Welcome Back to Win-Win Democracy

President Biden’s selfless and patriotic decision to drop out of the presidential race in favor of Vice President Harris has unleashed an astonishing surge of energy and support for Harris to become our first Madam President. More importantly, this energy and support could very well deny Donald Trump and the Project 2025 crowd the opportunity to turn our country into a theocratic, Putin-style oligarchy. I am more optimistic for the near future than I’ve been in a very long time.

Yet, our work to preserve American democracy is hardly done. Even if we elect Harris and gain a Democratic majority in both houses of Congress, I remain deeply concerned about the future of the country during the time in which my young grandchildren will grow up and live their adult lives.

Our decline into anocracy and the escalating risks of violent conflict that implies (discussed last time) will continue unless we act to reverse it. As we also discussed, to reverse the decline to anocracy, we must convince most citizens that our government serves their needs and desires. The best way to convince them is to make it so!

Before we can discuss specific ideas for government to better serve people’s needs, I need to address the elephant in the room: With a dysfunctional and highly partisan Congress and a reactionary Supreme Court, how can we possibly hope to change the country’s trajectory in time to affect my grandchildren’s lifetimes?

American Exceptionalism

Let’s start with some background. When I was in (public) high school in the late 1960’s and early 1970’s, we had to take an American History course. It was a memorize-names-places-and-dates course, light on concepts and analysis. I remember little of it. But two concepts stuck with me: Manifest Destiny and the Constitution’s role in the success of America. These were both framed as part of American Exceptionalism, the idea that America is unique and better than other countries.

Even as a student, I knew that the fancy name Manifest Destiny was an attempt to justify westward expansion as moral, inevitable, and divinely inspired. The teacher would brook no discussion of the horrors inflicted on indigenous peoples excused in the name of Manifest Destiny.

The Constitution was taught as a near-perfect document written by wise, heroic, and perhaps divinely-inspired founding fathers who created a novel structure of checks and balances to forestall both an American king and tyranny of the majority.

Despite its obvious flaws, the Constitution, the story went, was what made America exceptional and why our democracy survived for almost 200 years. Like many Americans, I accepted this portrayal of the Constitution and its role in making America exceptional.

The pre-MAGA Republican Party embraced and promoted American exceptionalism and the Constitution. Here are the first six paragraphs of the 2016 Republican Party platform’s preamble:

“We believe in American exceptionalism.

We believe the United States of America is unlike any other nation on earth.

We believe America is exceptional because of our historic role — first as refuge, then as defender, and now as exemplar of liberty for the world to see.

We affirm — as did the Declaration of Independence: that all are created equal, endowed by their Creator with inalienable rights of life, liberty, and the pursuit of happiness.

We believe in the Constitution as our founding document.

We believe the Constitution was written not as a flexible document, but as our enduring covenant.”

American exceptionalism says that, yes, despite our flaws, our country is great, so don’t muck around with what made it great. Indeed, our country has been an astounding success in many important ways. But it has failed in other important ways, which has led to widespread dissatisfaction with politics and government.

American exceptionalism is false. History is replete with countries and empires that were astounding successes for centuries then declined or collapsed. There is no reason to believe that America is immune from the lessons of history.

In the now-famous scene from the HBO series The Newsroom, Jeff Daniels, playing the character Will McAvoy, dares to state that “America is not the greatest country in the world anymore”:

American exceptionalism is being used as a marketing slogan to duck responsibility for fixing what ails us. Let’s face that reality.

The Constitution’s Flaws

The Constitution, even as amended, is flawed in ways that make it nearly impossible to fix our longstanding governance problems.

Citizens Are Bestowed Unequal Voice

The Constitution bestows unequal voice to our citizens: Where you live determines how well you’re represented.

President. Starting at the top, the president is elected by the Electoral College, not by the national popular vote. Five presidents have been elected despite losing the popular vote. Because all but two states^⁠1 have a winner-takes-all system of appointing their electors, if you live in a state that is either heavily Democratic or heavily Republican your vote for president doesn’t matter —  presidential candidates don’t spend their time and money in such states. In the 2024 election, voters in seven swing states effectively will choose the president while the rest of us look on.

Some argue that the Electoral College is part of the founding fathers’ genius and that they intended to give small states greater representation in choosing the president. In reality, the Electoral College, in conjunction with the Three-Fifths Compromise2,^⁠ was devised to preserve slavery. James Madison, often considered the father of the Constitution, advocated for national popular vote at the original Constitutional Convention and argued later in life for changing the Electoral College. The Electoral College was not an element of a brilliant plan, but rather a pragmatic compromise to bring the southern slaveholding states into the country. (A concise, fascinating history of James Madison and the Electoral College is available here.)

Senate. In the Senate, since every state, regardless of population, gets two senators, people residing in low-population states have far more voice per person than people residing in high-population states. One voter in Wyoming has the same representation in the Senate as 59 people in California. Citizens in Washington DC and Puerto Rico, which are not states, get no representation even though Washington DC has more citizens than both Vermont and Wyoming and Puerto Rico has more citizens than 20 states.

Further, as Nick Mourtoupalas explained in the Washington Post, because of the way different racial groups are distributed geographically, people of color are significantly less represented than white people; likewise, since the least populous states elect more Republicans than Democrats, Republican Senators maintain majority control of the Senate despite representing a minority of Americans.

The situation is made even worse by the Senate tradition of the filibuster, which requires that 60 votes, not 51, are necessary to pass most legislation. The filibuster amplifies the already unequal voice of citizens in the Senate.

House. The voice of the people in the House — the chamber of Congress that most directly represents the people — is badly distorted by gerrymandering of House districts. In my state of North Carolina, for example, total votes cast for House representative in districts across the state split nearly evenly across the two parties. Yet, the 2018 election yielded ten Republican and three Democratic representatives because the districts were heavily gerrymandered for partisan advantage. After the NC Supreme Court threw out the gerrymandered district maps in 2022, we elected seven Republican and seven Democratic representatives. After Republicans gained a majority on the NC Supreme Court in the 2022 elections, the state legislature once again created highly partisan gerrymandered maps, which the newly constituted NC Supreme Court permitted. As a result, it is now expected that the 2024 elections will yield either a 10-4 or 11-3 split.

If you happen to live in a House district gerrymandered to be strongly Republican or strongly Democratic — which most districts are — your vote in the general election for House representative  doesn’t matter. Representatives in such districts know that only the primary election matters, so they take positions that appeal to the most ardent partisans in their party, further weakening representation.

State Legislatures. Following the federal example, all states but Nebraska have a bicameral legislature (i.e., a legislature with two chambers), replicating at the state level the problems with the federal Senate and House:

• If a chamber’s districts are pre-determined (e.g., by county) with a fixed number of representatives per district, then people in low-population districts have more voice per person than people in high-population districts.

• If a chamber’s districts are determined by a political process, gerrymandering again rears its ugly head.

For example, North Carolina has both a state Senate and a state House, with districts for both determined by a political process, which results in heavily gerrymandered districts in both chambers.

Skewing of representation in state legislatures carries forward to the federal government in important ways:

• Federal elections are managed state-by-state, with each state having the power to set their own rules. There is a long, sordid, ongoing history of that power being used to disenfranchise certain people.

• Amendments to the US Constitution must be ratified by three-fourths of state legislatures (or state conventions).

Unequal voice in state legislatures worsens unequal voice at the federal level.

The Supreme Court is an Instrument of Minority Rule

The Constitution does not clearly delineate the Court’s power, which has allowed it, over time, to gain extraordinary, unchecked power to become an instrument of minority rule, frequently in service of the wealthy and powerful. Moreover, since Justices are appointed by the president subject to approval of the Senate, even their selection is tainted by the unequal voice citizens have in the Senate and in choosing the president. 

I’m not going to dwell on the problems with the Supreme Court here since we’ve discussed them previously, specifically in SCOTUS as an Instrument of Minority Rule and in other posts relevant to the Supreme Court’s role in the country.

Amending the Constitution is Excruciatingly Difficult

Amending the Constitution is excruciatingly difficult. In our almost 250-year history, Congress has proposed 33 Amendments of which 27 were ratified. The first ten Amendments are the Bill of Rights, which were proposed and ratified as part of the effort to get the states to ratify the Constitution itself. The so-called Reconstruction amendments (13^th-15^th) were ratified looking down the barrel of the Union Army’s guns.

Those almost 250 years have seen unimaginable changes in the country — the Industrial Revolution and the harnessing of energy sources; massive agricultural productivity improvements allowing only a few percent of our population to grow more food than we need; means of instant, free, portable communication; near-universal education and literacy; wide availability of books and other literature; the transition from animal-drawn conveyance to trains, cars, trucks, and planes; increased quality of life and longevity resulting from effective medical care; science yielding a better understanding of nature’s working; an astounding array of products forming the basis of a consumer economy; the rise of computing; and more.

Yet, despite these unimaginable changes, beyond the Bill of Rights and the Reconstruction amendments we’ve ratified only 14 amendments. Ten of these are what I’d call tinkering around the edges. Only four make significant changes to our governance: the 16^th (1913), permitting an income tax; the 17^th (1913), establishing direct election of Senators; the 19^th (1920), allowing women to vote; and the 24^th (1964), prohibiting voter disenfranchisement by poll taxes.

Some would argue that the difficulty of amending the Constitution is a feature, preventing us from making rash decisions without careful consideration. I’d offer the 18^th Amendment, Prohibition, ratified in 1919 and the 21^st Amendment, repeal of Prohibition, ratified in 1933, as a counterexample.

Conversely, although women constitute more than half of the US population, it took 80 years of effort to ratify the 19^th amendment, which allows women to vote, and, after 100 years of great effort, an amendment to bestow Equal Rights on women remains unratified.

Solving our most significant governance problems all require Constitutional Amendments. This includes overturning harmful rulings by the Supreme Court such as its interpretation of the 2^nd Amendment, its granting effectively peoplehood to corporations, its allowing unlimited intrusion of money into campaign finance, and, most recently, granting the president unlimited immunity from criminal prosecution.

Why is it so difficult to amend the Constitution? Primarily because the unequal voice mechanisms that we’ve just discussed for the Senate, House, and State Legislatures come into play throughout the process, making an already-high threshold for action nearly impossible. An amendment is initially proposed by either two-thirds vote of both the House and Senate or by two-thirds of the States.

Consider the implications of this requirement in the Senate alone: It takes 67 votes in the Senate to propose an amendment, so 33 Senators can block a proposed amendment. Conceivably, the Senators from our 17 lowest-population states could block a proposed amendment, yet the Senators from those states represent under 7% of our population.3 To be adopted, a proposed amendment must be ratified by three-fourths of the state legislatures, each of them potentially affected by the unequal voice problems we’ve already discussed.

The bottom line is that an already-high threshold is made nearly impossible when the unequal-voice aspects of the Senate, House, and state legislatures are cascaded on top it.

The Constitutional Bind

In April, Aziz Rana, who is a law professor at Boston College Law School and also holds a Harvard Ph.D. in political science, published the book The Constitutional Bind: How Americans Came to Idolize a Document that Fails Them. I became aware of this book through Jedediah Britton-Purdy’s extensive review The Creed: How did Americans come to worship the Constitution?” published in the July 2024 issue of The Nation. Last month, the New York Times non-fiction book critic Jennifer Szalai also reviewed the book in The Constitution Is Sacred. Is It Also Dangerous?.

Britton-Purdy observes that the Supreme Court and some of its justices have been vilified recently for deeply unpopular rulings on abortion, guns, and affirmative action. But even liberals return quickly to “faith in the Constitution as the keystone of the country and the safeguard of American democracy.”

Britton-Purdy goes on to introduce the book:

“In his fascinating and powerful new book, The Constitutional Bind, Aziz Rana calls this faith in the Constitution’s essential goodness “creedal constitutionalism” and urges Americans to reject it, perhaps along with major parts of the Constitution itself. His book is much more than a progressive critique of Constitution worship: Rana presents a sweeping history of constitutional politics from the late 19^thcentury to the present that reverses much of what Americans have learned to accept about the Constitution’s meaning. He portrays creedalism as a relatively recent phenomenon, a product of the 20th century, and contends that putting the Constitution at the center of American civic culture has abetted authoritarian and repressive agendas as much as it has upheld civil libertarian or democratic ones. And against the assumption that the Constitution is purely a domestic matter, he finds key moments in our constitutional culture developing through the United States’ imperial adventures and geopolitical contests.”

Rana himself says (p. 36-37):

“The time has come to reconsider the value of US constitutional exceptionalism, including the continued embrace of an order that has proven ill-suited for a collective American project aimed at equal and effective freedom. Today’s brand of creedal constitutionalism is a real hindrance to serious discussion of institutional and social change. It promotes a persistent retreat into a politics of constitutional veneration, which has had the effect of naturalizing a fundamentally undemocratic order.”

I urge you to read Britton-Purdy’s entire review and, if you can lay your hands on it, Chapter 1 of Rana’s book, entitled “The American Constitutional Romance.”

Recognizing creedal constitutionalism is a necessary step in having discussions about what we might do to reshape American democracy to be responsive to ordinary people’s needs and wants. We simply can’t get the job done with a business-as-usual approach with the deck as stacked as it is.

The Danger of Article V Conventions

Article V of the Constitution offers an alternative way to amend the Constitution: At the request of the legislatures of two-thirds of the states (currently, 34), Congress “shall call a Convention for proposing Amendments.”

Such a convention has never been called and scholars have debated how such a convention might work.4 Since 1960, the states have submitted more than 180 applications for Article V conventions on various topics. Congress has never considered the threshold met. But the rules aren’t clear. If a state requested a convention 100 years ago is that request still valid? Do requests all have to be on the same topic to count towards meeting the threshold? If a convention is called, how are the delegates chosen and how does the convention make decisions. Congress seemingly has a free hand to define the rules, subject, of course, to SCOTUS review.

Stepping into this uncertainty is a well-funded and organized effort by Republicans to use an Article V convention to, as Duke University historian Professor Nancy MacLean (author of the groundbreaking book Democracy In Chains) puts it in a February 2024 article in New Republic, “eviscerate core rights and protections most Americans hold dear.”

MacLean describes the organization Convention of States Action, a non-profit driving an effort to call a convention. They’re supported by big names in the right wing, including the Koch brothers network and the Heritage Foundation, as well as Republican notables like Speaker of the House Mike Johnson (who would have a hand in deciding the rules for meeting the threshold to call a convention), Mark Meadows, Ron DeSantis, Greg Abbott, and Sean Hannity. They’ve held three practice conventions, with the one in August 2023 adopting six amendments. Check out their website; they’re not bashful about laying out their goals and touting how close they are to success.

Common Cause, a non-partisan organization that describes itself as “dedicated to upholding the core values of American democracy” and which is a leader in fighting gerrymandering in both federal and state courts, recognizes three additional major campaigns for an Article V convention: the Balanced Budget Amendment campaign, supported by the Koch-brothers-supported American Legislative Exchange Council (ALEC), the Wolf-PAC campaign, and the term limits campaign.

By Common Cause’s count (remember — the rules are unclear), 28 states have called for a convention, just six states shy of the 34 states needed.

Given that an Article V convention would be called and run under rules yet to be defined, with delegates selected by a yet unknown process, and that the major efforts to call such a convention are being driven by well-funded, well-organized right-wing organizations, there’s good reason to be alarmed.

Common Cause is leading efforts across key states to convince their legislatures to “reject and rescind calls for a convention.”

The Opportunity of Article V Conventions

As both Professor MacLean and Common Cause have made clear, an Article V convention called under current conditions would be extremely dangerous to American democracy. Indeed, I have signed Common Cause’s petition to reject calls for Article V Conventions and have donated to support their efforts.

But I wonder if there isn’t also opportunity in Article V Conventions if appropriate circumstances could be created. Let’s think outside the box about this for a few moments.

When there is danger in choosing to take an action, we must also consider the danger of the corresponding inaction. Jumping out of a second story window of your house is dangerous. But if the house is on fire, jumping might be your best chance to save yourself. Or, it might be better to wait for the firefighters. It depends on the circumstances.

As we discussed last time, there’s posturing about civil war and we’ve been through a violent insurrection at The Capitol intended to stop the peaceful transfer of presidential power. We’ve already had one civil war, and it was horrendous. It forced us to eliminate slavery, at least on paper, via the Reconstruction Amendments. But as historian Heather Cox Richardson explains in her book How the South Won the Civil War: Oligarchy, Democracy, and the Continuing Fight for the Soul of America, “[f]rom Reconstruction through World War II, Americans recreated a hierarchical society” with wealthy white men once again at the top. That struggle continues and the oligarchs are winning.

We are approaching the anocracy danger zone for civil war. A modern civil war would be much, much worse than our first one.

Could an Article V Convention provide an opportunity to reshape American Democracy in ways that could help to avoid civil war and deliver real freedom to more people?

What Next?

It’s a hard question to answer, but that’s what I want to start discussing next time. Are there conditions under which various constituencies could get enough of what they want to be willing to compromise at an Article V convention? In other words, are there win-wins for some of the deep, polarizing disagreements we have?

Stay tuned …

Leave a comment

After a year-long “rest,” I am restarting Win-Win Democracy. Since there is excellent discussion elsewhere about our current struggles, I will write mostly about how we might reshape American democracy for modern times. I’ll discuss what this means below. I plan to write shorter, easier-to-consume posts than previously, publishing them when ready rather than on a fixed schedule.

I will be exploring some new ideas, some of which are likely to be controversial or even seem harebrained. Your feedback will be invaluable, especially if you write in the comments, where others can see and discuss them. If you subscribe to Win-Win Democracy (free), you can also write to me privately, either by replying to the email version of the post you’ve received in your inbox or by emailing me at winwindemocracy@substack.com.

Close Split Among Widely Divergent Views

Nearly 47% of the Americans who voted in the 2020 presidential election supported Donald Trump and the Republican Party, which many other American voters believe is working to destroy the essence of American democracy. For the first time in American history, the Supreme Court, reshaped by Trump’s three appointments, has rolled back longstanding individual rights and granted the American president immunity fit for a king. Many of Trump’s supporters believe that they are the patriots working to save America from “radical leftists” while simultaneously believing the Big Lie that the 2020 election was stolen.

Conversely, around 51% of voters in the 2020 presidential election supported Joe Biden and the Democratic Party, which many of Trump’s supporters believe are taking away freedoms, ruining the economy, allowing the country to be overrun with immigrants and allowing crime to flourish. Many of Biden’s supporters believe that they are the ones working to save America from an authoritarian (or fascist) second Trump administration that will destroy our democracy and further diminish our freedom and rights.

This close split among widely divergent views is reflected in the close popular votes in our recent presidential elections this century:

We’ve also had close partisan splits in both chambers of Congress, although interpreting what voters actually prefer from those splits is difficult because of extensive gerrymandering in House districts and our system of two Senators per state regardless of state populations.

The bottom line is that comparable numbers of the national electorate subscribe to the Republican viewpoint and to the Democratic viewpoint. So, the situation is unstable in the sense that small changes in voting behavior can cause enormous swings in policy and power.

Widespread Dissatisfaction with Politics and Government

I can’t think of anyone I know who is happy with the state of our nation’s politics. In fact, I can’t think of anyone I know who isn’t angry and upset about our nation’s politics, a few even to the point of joking about other countries to move to before the United States descends into widespread political violence or worse.

It’s not just my people: The widely respected Pew Research Center conducted a study on the state of the nation’s politics in mid-2023. The study’s title reflects their conclusion: “American’s Dismal Views of the Nation’s Politics.” The report is clearly written and compelling; I urge you to read or skim it for yourself.

At over 100 pages of conclusions and another 40 explaining the methodology and giving the raw data, I won’t even attempt to summarize the report beyond showing you two of their charts, which show devastating results:

Another Civil War?

A member of Congress talks openly of coming civil war. Kevin Roberts, the president of The Heritage Foundation, the conservative think tank behind Project 2025^⁠1 — which they tout as the plan “to rescue the country from the grip of the radical Left” if Trump wins the upcoming election — has said that the country is in the midst of a “second American Revolution” that will be bloodless “if the left allows it to be.” I don’t know about you, but that sounds to me like a threat.

Is the situation really that bad?

On the one hand, yes, there is an extensive, well-funded, well-organized effort to radically transform America into a fundamentalist Christian nation run by white male oligarchs. The Supreme Court has already been co-opted into this picture. And the veiled threats of political violence are intimidating legislators and other government officials throughout the country, including coercing some Republican politicians to change their behavior. Threats can quickly turn into action. The divide between so-called red states and blue states is growing more stark.

On the other hand, no, when you look at the reasons “ordinary” Republicans and Democrats are unhappy, they’re surprisingly similar: Big donors have too much influence. Neither party is popular with their adherents. There’s too much corruption in the political class. Stuff is too expensive. Too many people don’t have enough to eat or access to healthcare or access to good education for their kids. Homes are out of reach. Retirement is out of reach. Children are no longer going to do better than their parents. The country’s important problems get too little attention.

Nevertheless, despite having similar concerns, Republicans and Democrats dislike each other and are choosing to live and interact primarily with their own kind. Living apart makes it easier to think of the “others” as the enemy. And that’s dangerous.

Anocracy

Anecdotal discussion is one thing; data is another: Historians and scholars of international relations have studied the conditions that have led to civil wars around the world and across time. If, like me, you reject American exceptionalism and think that Americans are subject to the same forces as others, it’s worth understanding a bit of their research and conclusions about what has happened elsewhere.  I’ve been guided in this by Barbara F. Walter’s 2022 book “How Civil Wars Start: And How to Stop Them.” Walter is Professor of International Relations at the School of Global Policy and Strategy at the University of California, San Diego.

Here’s her punch line (p. 11):

“It turns out that one of the best predictors of whether a country will experience a civil war is whether it is moving toward or away from democracy. Yes, democracy. Countries almost never go from full autocracy to full democracy without a rocky transition in between. Attempts by leaders to democratize frequently include significant backsliding or stagnation in a pseudo-autocratic middle zone. And even if citizens succeed in gaining full democracy, their governments don’t always stay there. Would-be despots can whittle away rights and freedoms, and concentrate power, causing democracies to decline. Hungary became a full democracy in 1990 before Prime Minister Viktor Orbán slowly and methodically nudged it back toward dictatorship. It is in this middle zone that most civil wars occur.

Experts call countries in this middle zone ‘anocracies’—they are neither full autocracies nor democracies but something in between.”

A country can pass through anocracy — this middle zone — either transitioning from autocracy to democracy or transitioning from democracy to autocracy. Anocracy is a danger zone approached from either direction.

How is anocracy measured? Researchers have collected large datasets about the types and characteristics of government around the world and over time, measuring many variables. From these datasets, conflict researchers have settled on something called the Polity Score as a quantitative way to determine where a country is on the autocracy to democracy spectrum, with -10 being most autocratic and +10 most democratic. Norway, New Zealand, Denmark, and Canada score +10; North Korea, Saudi Arabia, and Bahrain score -10. The United States, which until recently scored +10, scored +8 in 2018. A country is considered an anocracy between -5 and +5 and countries in that range are much more likely to experience civil war.

Walter goes on to say (p. 19-22):

“Since 2000, democratic leaders who came to power via elections have begun to consolidate authoritarian rule. Civil war experts are once again worried. We understand that backsliding almost certainly means that the middle zone is likely to expand.

…

Democratic countries that veer into anocracy do so not because their leaders are untested and weak, like those who are scrambling to organize in the wake of a dictator, but rather because elected leaders—many of whom are quite popular—start to ignore the guardrails that protect their democracies. These include constraints on a president, checks and balances among government branches, a free press that demands accountability, and fair and open political competition.

…

For a decaying democracy, the risk of civil war increases almost the moment it becomes less democratic. As a democracy drops down the polity index scale—a result of fewer executive restraints, weaker rule of law, diminished voting rights—its risk for armed conflict steadily increases. This risk peaks when it hits a score of between +1 to −1—the point when citizens face the prospect of real autocracy. The chance of civil war then sharply drops if the country weathers this moment by becoming even more authoritarian, or changes course and begins to rebuild its democracy.”

Sound familiar?

The Future of American Democracy

Almost 250 years ago our country’s founders launched the great American experiment in democracy, proposing that “we the people” could govern ourselves without appeal to the supposedly divinely appointed authority of a monarch or pope. It has been a roaring success and a depressing failure. I’ll spare you the litany of both. You know it.

One thing is certain: We are headed in the wrong direction, further into the danger zone. I mentioned that our Polity Score in 2018 was +8. In 2020, it dropped to +5, edging ever closer to anocracy. Why?

Poor governance. Too many people believe that government is increasingly ineffective in serving their needs and is more corrupt; that our government doesn’t serve “we the people”; that political accountability is weak; that money determines outcomes, and more.

In the last chapter of Walter’s book, which is titled “Preventing a Civil War,” she states that three features of democracy stand out as most important in preventing civil war (p. 200):

“’the rule of law’ (the equal and impartial application of legal procedure); ‘voice and accountability’ (the extent to which citizens are able to participate in selecting their government, as well as freedom of expression, freedom of association, and a free media); and ‘government effectiveness’ (the quality of public services and the quality and independence of the civil service). These three features reflect the degree to which a government serves its people and the degree to which its political institutions are strong, legitimate, and accountable. Improvements in governance tend to reduce the subsequent risk of war. “

Driven by today’s Republican Party, we are getting weaker in all of these aspects, not stronger. If Project 2025 is implemented, it will drive us towards autocracy, crossing the center of the anocracy zone, the riskiest place to be in terms of risk of civil war.

I don’t want to let Democrats off the hook here. Democrats have for decades avoided tackling some of our biggest governance problems when they’ve had the chance; like Republicans, they’ve become beholden to big business and big money; they’ve been complicit in letting the quality of government services decline and subscribing to neoliberal economic theories that have devastated certain communities.

Next Time

America’s governance problems are nothing new. The country’s first Constitution, the Articles of Confederation of 1777, created a national government so weak in comparison to the state governments that James Madison and George Washington feared that “their young country was on the brink of collapse.” At the Constitutional Convention of 1787 in Philadelphia, a new Constitution was written, redesigning the government to give the federal government more powers with respect to the state governments.

Many of today’s most egregious governance problems result from compromises made so that the states would ratify the new Constitution. Moreover, the Constitution is so difficult to amend that it has not kept up with the revolutions in communication, transportation, energy, computing, science, and medicine of the last 250 years.

Next time, I plan to begin exploring win-win approaches to changing this situation. Do we really need to have a governance structure in which one “side” wins and another “side” loses? Where we ping-pong between different approaches because neither “side” has a compelling lead over the other? We have many states that are “red” in their rural areas and “blue” in their cities. Why do we need to force a single viewpoint when, in fact, people have multiple viewpoints? We spend enormous energy arguing about which approach to structuring our economy is best. Can we try multiple approaches simultaneously?

Leave a comment

In October 2023 I spoke about my perspectives on AI (video here). Positive response to that talk led to me presenting an updated version of the talk two weeks ago. The response was again positive, so I’m sharing it here in the hope that you’ll find it interesting and useful.

(Thanks to Dave DeVries for hosting me at Carolina Meadows in Chapel Hill, NC and to Marty Lyon for recording the video.)

Leave a comment

Note: An updated version of this talk is available at

Share

In the last few days I’ve seen two items relevant to previous topics on Win-Win Democracy that I’d like to share.

Trump is a Fascist

Robert Reich and his Inequality Media organization have released a 7-minute video that shows that Trump and the politicians who support him have moved beyond authoritarianism to full-throated fascism. It is intended to sway people who don’t understand what is happening, so it is blunt, concise, and uses powerful imagery.

Reich has asked people to share this widely, so here it is:

You can read Reich’s written explanation here.

For those so inclined, Inequality Media is a worthwhile organization that needs our support. They accept tax-deductible donations.

My longer article from June on the broader topic of fascist politics in America, which seems ever more relevant, is here:

Reforming SCOTUS

Back in February, we completed our series on the Supreme Court, finishing up with some thoughts about Reining in the Imperial Supreme Court:

Renowned Constitutional scholar Laurence Tribe has just published a long review in The New York Review of Books of two recent books on the Supreme Court. I have not yet read the books, but Tribe’s review itself is exceptionally illuminating. His title might encourage you to read it: Constrain the Court—Without Crippling It.

This article is behind a paywall, but one that provides generous free access. If you create a free account with The New York Review of Books (requires giving them an email address and making up a password, but no credit card) it will let you read the article.

Share

Subscribe now

Share Win-Win Democracy

Ezra Klein said

“Jennifer Pahlka has written one of the best policy books I’ve ever read.”

He was talking about Recoding America, released last month. I just finished reading it and think that Win-Win Democracy readers should read it too. Alternatively, listen to or read Klein’s interview with her.

Pahlka explains why government commonly delivers programs and services that are oppressively difficult for people to access and extraordinarily expensive for government to provide. And late. Not a good combination.

Using real examples, she explains the cultural issues that cause legislators, policymakers, and bureaucrats to make government fail to deliver for citizens, not (usually) out of malice, but out of processes that were designed for a different era, not for the digital age.

(Many Win-Win Democracy readers have worked in the software industry. You’ll recognize the same kinds of massive failures that industry had when waterfall development processes were dominant and some similar cultural issues were rampant. Teaser: Enterprise service bus for GPS satellites.)

Again using real examples, Pahlka shows what it takes to do much better. It can be done — it has been done — but it will take years to fundamentally change the way that governments work, making them more successful at delivering services to citizens.

(For the software folks: Agile development processes, with a focus on product management instead of project management, building and deploying minimal systems, and continuous stakeholder involvement.)

This is all relevant to Win-Win Democracy readers because she also explains how we all win — liberals and conservatives — by providing good services to citizens, thus restoring their faith in government while spending less to deliver more.

Leave a comment

Win-Win Democracy Takes a Rest

After almost a year and a half, 267 subscribers, 34 long-form posts, and almost 30,000 views total, I’ve decided to (at least temporarily) reduce my effort on writing Win-Win Democracy. Three factors drove me to this decision:

1. Writing this newsletter takes a lot of my time. I enjoy working on it — especially doing the research — but I’ve made little progress on other projects that are important to me and to my family.

2. I’ve failed to attract enough readership — especially people who are different than me — to have impact commensurate with the amount of work writing the newsletter requires.

3. Prospects for win-win solutions are greatly diminished when the Republican Party nationwide is focused on using fascist political tactics to gain authoritarian power rather than engaging to solve the country’s real problems. For example, in 2023 alone, Republican legislators across the nation introduced more than 500 anti-LGBTQ bills, harming many without helping anyone. The Republican Party has lost its way and, in its current form, can not be a partner in working toward common goals.

I expect to occasionally write new material, some about win-win ideas and some about other particularly important topics related to the viability of our democracy, like I did last time with Fascist Politics in America.

I appreciate all of the encouragement and feedback that I’ve received along the way.

Thinking About AI

When we last talked about AI, I argued that “regulating AI” itself is the wrong approach to controlling the impact of AI on our society. I suggested instead that we strengthen the laws and regulations related to disinformation, platform liability, verification of identity, unemployment, fair use, wealth inequality, etc.

AI does not create these problems, but it probably will worsen the problems that we already have.

NYU social psychologist Jonathan Haidt and former Google CEO Eric Schmidt, writing in the article AI is About to Make Social Media (Much) More Toxic, put it this way:

“We can summarize the coming effects of AI on social media like this: Think of all the problems social media is causing today, especially for political polarization, social fragmentation, disinformation, and mental health. Now imagine that within the next 18 months — in time for the next presidential election — some malevolent deity is going to crank up the dials on all of those effects, and then just keep cranking.”

Given that we as a society haven’t made progress solving the many problems we already have with disinformation, social media, etc., what would make one think that regulating AI would improve the situation? We must address the root issues, not just try to keep AI from making them worse.

I’ve already discussed some approaches to doing so:

Here, I want to address the highly visible push from industry leaders for Congress to regulate AI, whatever that might mean.

My short message: Be skeptical. Highly skeptical.

Industry’s Charm Offensive on Regulating AI

As reported by The Daily Beast in the article How Congress Fell for Sam Altman’s AI Magic Tricks, OpenAI’s CEO Sam Altman performed for Congress in May, showing off the gee-whiz capabilities of large language models and generally wowing the legislators, not to mention wining and dining them the evening before. Senator Richard Blumenthal told reporters that “Sam Altman is night and day compared to other CEOs, and not just in the words and the rhetoric but in actual actions and his willingness to participate and commit to specific action.” 

Meanwhile, as reported by the Washington Post, Senate majority leader Charles Schumer has been meeting with industry leaders and representatives as well as academics and critics of AI, as part of what he calls “an all-hands-on-deck effort in the Senate.” Schumer says “We need the best of the best sitting at the table: the top AI developers, top executives, scientists, advocates, community leaders, workers, national security experts all together in one room, doing years of work in a matter of months.”

Put all of this in the context of the call 3 months ago from prominent scientists and leaders for a 6-month pause on work on advanced AI. As far as I can tell, none of the industry leaders asking to be regulated have done anything to slow down their own work.

It is nearly impossible to know the true motives of the tech industry’s leaders, but we can get a hint of their motives by understanding how they interact with regulators in the EU, which is far ahead of the United States in providing privacy protection for consumers, regulating social media, and enforcing anti-trust among tech companies.

The EU has been working on regulations for AI for years, building on top of the regulatory framework they already have in place (which we completely lack). As reported in the Washington Post, Google, Microsoft, and OpenAI all declined comment after the European Parliament overwhelmingly passed the EU AI Act last month, which includes limits on “tools that could sway voters to influence elections or recommendation algorithms, which suggest what posts, photos and videos people see on social networks.” No comments from these companies about supporting new regulations on AI — could it be that when the rubber hits the road they really don’t want to be restrained?

We should be wary of the tech industry’s charm offensive on regulating AI. Despite protestations to the contrary, it seems plausible, even likely, that tech leaders are really seeking protections that allow them to do what they want without being accountable for the results, much like what Section 230 has done for existing social media and digital platforms.

Meanwhile, we hear nothing from these same industry leaders about changing their current businesses to protect society from the extremely harmful effects of their current platforms. Absent that, I can’t be sanguine about their true intent with their pleas for AI regulation.

What’s Next?

I probably won’t be writing anything major before fall, although that could change depending on events.

I’d be interested in any advice you can offer on how to both expand and broaden the audience and/or to evolve the content. You can respond publicly in the comments or privately by replying to this email.

Leave a comment

Share Win-Win Democracy

Welcome Back to Win-Win Democracy

Writing today, on Memorial Day, I am grateful for the extraordinary sacrifices made by so many Americans — including my wife’s father and uncles — to defend freedom around the world.

World War II, which took the lives of about 3% of the world’s population1, was, to oversimplify, a fight against Japanese imperialism and European fascism, as embodied by Hitler’s Germany, Mussolini’s Italy, and their allies. European fascism was defeated in World War II.

Fascist politics — which I’ll define in a moment — remains, however, a threat to liberal democracies2 in many places, including the United States. Indeed, I’ve interrupted our regularly-scheduled program on AI because I’m urgently alarmed by the prominent role of fascist politics in today’s Republican Party.

There is no win-win on this topic, only win. If fascist politics wins, our representative democracy is doomed to disappear. We who value democracy, freedom, and the rule of law must act now to win.

Fascism

The word Fascism has now no meaning except in so far as it signifies 'something not desirable.' — George Orwell (1945)

Much ink has been spilled trying to define fascism.

For an expansive, scholarly, although somewhat difficult-to-read discussion of how to define fascism, I recommend Stanley G. Payne’s book A History of Fascism, 1914-1945, particularly the Introduction, which gives a “a wide-spectrum description that can identify a variety of differing allegedly fascist movements while still setting them apart as a group from other kinds of revolutionary or nationalist movements” and Chapter 15, which gives a sort of reverse engineering (he uses the term “retrodiction” but I had to look that up) of the factors that allowed fascism to arise during the early 20th century in certain European countries but not others.

Since Payne’s discussion is more than most of us need, I turned to Wikipedia, which devotes an entire article to definitions of fascism. Here are two that are particularly concise and informative:

“[F]ascism is best defined as a revolutionary form of nationalism, one that sets out to be a political, social and ethical revolution, welding the ‘people’ into a dynamic national community under new elites infused with heroic values. The core myth that inspires this project is that only a populist, trans-class movement of purifying, cathartic national rebirth (palingenesis) can stem the tide of decadence.” — Roger Griffin, Oxford-Brookes University (1991)

“A form of political behavior marked by obsessive preoccupation with community decline, humiliation or victimhood and by compensatory cults of unity, energy and purity, in which a mass-based party of committed nationalist militants, working in uneasy but effective collaboration with traditional elites, abandons democratic liberties and pursues with redemptive violence and without ethical or legal restraints goals of internal cleansing and external expansion.” — Robert Paxton, Columbia University (2004)

Is Trump a Fascist?

I certainly see many elements of Trumpism in these definitions.

Whether Trump and his supporters are (or are becoming) fascists is the subject of debate. David Frum, a former speechwriter for Bush the Younger, argues in the Atlantic article There’s a Word for What Trumpism is Becoming that the events of January 6th were a turning point:

“The Trump movement was always authoritarian and illiberal. It indulged periodically in the rhetoric of violence. Trump himself chafed against the restraints of law. But what the United States did not have before 2020 was a large national movement willing to justify mob violence to claim political power. Now it does.”

Similarly, writing in Vox, in the article The F Word: The debate over whether to call Donald Trump a fascist, and why it matters, Dylan Matthews reports:

Is Donald Trump a fascist?

It’s a question I’ve tried to answer a few times in the six-odd years that he has dominated American politics. Back in 2015, no fascism expert would use the word to describe Trump. In October 2020, they were inching closer, but most dismissed the term as likely an exaggeration or distraction.

Matthews checks in again with several leading scholars of fascism on their views on this question. Again, January 6th seems to be the dividing line. As Robert Paxton, author of the second definition above, wrote in 2021 (emphasis mine):

Trump's incitement of the invasion of the Capitol on January 6, 2020 removes my objection to the fascist label. His open encouragement of civic violence to overturn an election crosses a red line. The label now seems not just acceptable but necessary. It is made even more plausible by comparison with a milestone on Europe's road to fascism—an openly fascist demonstration in Paris during the night of February 6, 1934.

But that viewpoint is not universally held. Some scholars, including Stanley Payne, argue, as Matthews puts it, that “Trump’s lack of coherent revolutionary fervor makes him fall short of fascism.”

If you’re interested in this question, including whether or not there are disadvantages to using the “F word”, I encourage you to read Matthews’ article.

As much as I think that Trump deserves to be denigrated, there is a more important question than whether or not he is a fascist:

Does today’s GOP use the tactics of fascism to build authoritarian power that threatens our democracy?

I aim to convince you that the answer to this question is a resounding “yes” and that we must use all appropriate means to prevent their success.

Fascist Tactics

In his excellent book How Fascism Works: The Politics of Us and Them, Yale professor of philosophy Jason Stanley makes an important distinction between fascist states and fascist tactics (p. 1):

“[M]y interest is in fascist tactics as a mechanism to achieve power. Once those who employ such tactics come to power, the regimes they enact are in large part determined by particular historical conditions. What occurred in Germany was different from what occurred in Italy. Fascist politics does not necessarily lead to an explicitly fascist state, but it is dangerous nonetheless.

Here is Stanley’s list of tactics:

• The mythic past

• Propaganda

• Anti-intellectualism

• Unreality

• Hierarchy

• Victimhood

• Law and order

• Sexual anxiety

• Appeals to the heartland

• Dismantling of public welfare and unity

Trump, his GOP MAGA supporters, and GOP Trump wannabes like Ron DeSantis, all use these fascist tactics. So, too, do GOP state-level politicians, especially in red or purple states.

We’ll discuss these in some detail, but you can no doubt already recognize some of these tactics in the actions of GOP political leaders throughout the country, from Trump on down.

Leaders intent on building their own power deploy these tactics to:

1. Divide the population into “us” and “them”

2. Denigrate and dehumanize “them” to enable using “them” as scapegoats for problems affecting “us,” and to make it acceptable to attack “them”

3. Shape policy and law to favor “us”

Fascist political tactics are mutually reinforcing, pushing society ever further toward division and strengthening that division. Indeed, as Stanley puts it (p. 5), “the most telling symptom of fascist politics is division.”

My explanation of how fascist tactics work is heavily based on Stanley’s book. I won’t be insulted if you want to skip the rest of this and go read his book now — it is only 200 pages. For those of you remaining, I will follow Stanley’s approach augmented by current examples.

The Mythic Past

Make America Great Again. There you have it: America was once great, but it isn’t now, and I, Donald Trump, am the only one who can make it great again.

What was our glorious past? Oh, you know, when the country was run by upstanding white, Christian men — decent, strong men, from the country’s heartland. Just like traditional fathers who guide their families with a firm hand.

Those halcyon days have been lost because of intellectuals, the elite, liberals, feminism, foreign immigrants diluting our culture, and pandering to calls for equality of the genders and races. But the world isn’t that way: Some people — us — are better than the rest and we deserve to run things.

Sarcasm aside, here’s how Stanley describes it (p. 13):

“The patriarchal family is one ideal that fascist politicians intend to create in society—or return to, as they claim. The patriarchal family is always represented as a central part of the nation’s traditions, diminished, even recently, by the advent of liberalism and cosmopolitanism. But why is patriarchy so strategically central to fascist politics?

In a fascist society, the leader of the nation is analogous to the father in the traditional patriarchal family. The leader is the father of his nation, and his strength and power are the source of his legal authority, just as the strength and power of the father of the family in patriarchy are supposed to be the source of his ultimate moral authority over his children and wife. The leader provides for his nation, just as in the traditional family the father is the provider. The patriarchal father’s authority derives from his strength, and strength is the chief authoritarian value. By representing the nation’s past as one with a patriarchal family structure, fascist politics connects nostalgia to a central organizing hierarchal authoritarian structure, one that finds its purest representation in these norms.”

We can see in this quest to create the mythical past the rationale for the otherwise seemingly irrational nationwide GOP efforts to suppress knowledge of our real history:

• Ban teaching of critical race theory, which is nothing more than an attempt to teach us all about our country’s past and its impact on its present.

• Ban books that portray other than traditional family structures or that make white children “uncomfortable”.

• Portray the Civil War as a battle for states rights and freedom instead of a battle for individual liberty and abolition of slavery.

• Dilute teaching of African American history by pushing back on the newly-released AP African American History curriculum.

• Stop WOKE. Florida governor Ron DeSantis pushed Florida’s Stop Woke Act (with Woke an acronym for “Wrongs to Our Kids and Employees) to prohibit Florida businesses and public schools from teaching certain concepts to students and employees. In other words, the law prohibits free speech.
  
  The governor’s press release is entitled “Governor Ron DeSantis Signs Legislation to Protect Floridians from Discrimination and Woke Indoctrination”. Protect those Floridians against learning anything that would call the perfection of that glorious mythic past into question.

Stanley summarizes (p. 30):

“Fascist leaders appeal to history to replace the actual historical record with a glorious mythic replacement that, in its specifics, can serve their political ends and their ultimate goal of replacing facts with power.”

Propaganda

Propaganda is (Merriam-Webster) “the spreading of ideas, information, or rumor for the purpose of helping or injuring an institution, a cause, or a person.”

Let’s look at a few examples.

Birtherism

Consider birtherism, the (false) rumor and conspiracy theory that Barack Obama was born in Kenya, hence ineligible to be president, and was a secret Muslim planning to destroy America from within its presidential inner sanctum. This conspiracy theory, though easily disproven, took hold.

Mitt Romney joked on the campaign trail that “No one’s ever asked to see my birth certificate; they know that this is the place that we [he and his wife Ann] were born and raised.” The message: I, a Christian white guy, belong; that Black guy with the odd name doesn’t belong.

Donald Trump picked up the theme and ran with it, dropping innuendo and false accusation whenever possible. His lies reinforced the MAGA mythic past, those good ol’ days when Black people knew their place and wouldn’t dare run for the highest office in the land.

Although Trump didn’t originate birtherism, he deployed it to great effect to send the message that Obama was overturning the traditional order of things.

Draining the Swamp

Donald Trump’s “drain the swamp” slogan, which he peppered everywhere to enthusiastic response, was masterful propaganda. Nearly everyone believes that corruption pervades Washington. A promise to clean it up, something he alone could do, was compelling to many.

The reality? Trump’s administration, including his Cabinet secretaries and closest advisors, is widely believed to be the most corrupt ever. I’ll spare you the long list, but you can find one here.

Propaganda about corruption is a favorite tool of fascist politics. As Stanley describes it (p. 34):

“Fascist movements have been “draining swamps” for generations. Publicizing false charges of corruption while engaging in corrupt practices is typical of fascist politics, and anticorruption campaigns are frequently at the heart of fascist political movements. Fascist politicians characteristically decry corruption in the state they seek to take over, which is bizarre, given that fascist politicians themselves are invariably vastly more corrupt than those they seek to supplant or defeat.”

Even worse are Trump’s attacks on judges who dare get in his way or threaten to hold his personal corruption and his administration’s corruption to account.

Stanley argues that this is typical of the “undemocratic intent behind fascist propaganda” (p. 37):

“Fascist states focus on dismantling the rule of law, with the goal of replacing it with the dictates of individual rulers or party bosses. It is standard in fascist politics for harsh criticisms of an independent judiciary to occur in the form of accusations of bias, a kind of corruption, critiques that are then used to replace independent judges with ones who will cynically employ the law as a means to protect the interests of the ruling party.”

Anti-intellectualism

GOP politicians across the country are attacking our universities, accusing them of being hotbeds of liberalism and Marxism, of bias against conservative faculty, and of denying freedom of speech to faculty and students.

Free Speech and Lack of Diversity Attacks

Universities students and, to a lesser extent, faculty, have, at least since the 1960s, been the vanguard of fighting injustice and authoritarianism. Part of the fascist playbook is, therefore, undermining and delegitimizing universities.

David Horowitz is a far right activist who has long been pursuing this agenda, creating multiple organizations like the Individual Rights Foundation, Students for Academic Freedom, and most recently the David Horowitz Freedom Center.

In 2002, writing on the website of Students for Academic Freedom, Horowitz explained:

“‘Diversity’ may be one of the contemporary university's most cherished values, but university officials with near universality have interpreted diversity to mean anything but a plurality of viewpoints — arguably the most important diversity of all. What is knowledge if it is thoroughly one-sided, or intellectual freedom if it is only freedom to conform? And what is a "liberal education," if one point of view is for all intents and purposes excluded from the classroom? How can students get a good education, if they are only being told one side of the story? The answer is they can't.”

He goes on to request university administrations to inquire into political bias in hiring of faculty and administrators and selection of commencement speakers, to not tolerate obstruction of campus speakers and meetings, and to ensure that faculty in classrooms welcome diverse viewpoints. And, he requests state legislatures to pursue these goals at their public institutions.

He posits that the problems can be solved without affecting academic freedom “[b]y adding the categories of political and religious affiliation to Title IX and other existing legislation … to redress an intolerable situation involving illegal and unconstitutional hiring methods along with teaching practices that are an abuse of academic freedom.”

This reminds me of a county board of elections hearing that I attended perhaps a decade ago. The room was packed; it was impossible to let everyone speak. So, the chairwoman, a Republican, announced that to be fair she would call on Democrats and Republicans alternately. There were probably ten times more Democrats in the room than Republicans. All the Republicans got to speak but only a small portion of the Democrats were able to speak.

Of course, the “balanced” hearing didn’t reflect people’s opinions. Republican critics of university’s “liberal bias” want this same kind of balance.

Back in 2002, Horowitz and his various organizations were fringe organizations few had even heard of. As reported by the Washington Post, they came to the fore with the Trump administration and its allies, including Steve Bannon, Stephen Miller, and Attorney General Jeff Sessions.

Indeed, universities are hotbeds of academic freedom. There are many faculty with conservative views, and their right to speak their mind is protected by tenure.

As Stanley explains (p. 50)

“Given the formal protections of academic freedom, universities in the United States host the freest domain of expression of any workplace. In private workplaces in the United States, free speech is a fantasy. Workers are regularly subjected to nondisclosure agreements, forbidding them to speak about various matters. In most workplaces, workers can be fired for political speech on social media. Attacking the only workplaces in a country with genuine free-speech protections using the ideal of free speech is another instance of the familiar Orwellian nature of propaganda.”

Attacking Tenure

The hypocrisy is rich: the faculty tenure system is in place principally to ensure that faculty have the freedom to speak their minds without fear of reprisal. Yet, GOP politicians are working to eliminate or neuter tenure:

• GOP legislators in Texas, Ohio, North Dakota, Louisiana, and Iowa have proposed bills to either eliminate tenure or weaken it by requiring periodic review of faculty tenure.

• Florida’s Ron DeSantis signed a bill requiring public universities to review faculty tenure every five years, justifying that by saying tenure promotes “intellectual orthodoxy”.

• In North Carolina, tenure decisions made by the University of North Carolina’s (UNC) administration have been rejected by the political appointees to the university’s board of trustees.
  
  Most visibly, when UNC’s journalism school offered a tenured faculty position to Nikole Hannah-Jones, a MacArthur “genius grant” recipient and Pulitzer-prize-winning New York Times author, the board of trustees rejected the tenured appointment, evidently at the behest of Walter Hussman, a conservative who made a $25M pledge. Details are available here and here.

• Texas Lt. Governor Dan Patrick proposed revoking the tenure of any faculty at a public university who teaches critical race theory and abolishing tenure for all newly-hired faculty, regardless of discipline.

• and more …

Taking Control

Republicans are not just propagandizing about universities as hotbeds of liberalism and deniers of free speech, they are actively working to take control of public universities.

Ron DeSantis and his education commissioner want to remake Florida’s New College, a small, liberal arts college, into a lookalike of Michigan’s Hillsdale College, a small Christian college that has become the darling of Christian conservatives for its conservative activism. DeSantis replaced 6 of New College’s 13 trustees with strong conservatives; the board of trustees then ousted the College’s president, replacing her with a career politician at more than double her salary. The transition is underway. More details here.

Lest you think that I’m picking on Ron DeSantis and Florida, or that they’re unique, let me relate how the GOP is taking partisan control of the public universities in my home state of North Carolina. Details are available in Report of a Special Committee: Governance, Academic Freedom, and Institutional Racism in the University of North Carolina System, chartered by the American Association of University Professors. Here’s a brief synopsis of the political interference changing the character of the university system:

• 2010: Republicans gained control of both houses of the NC General Assembly for the first time in more than a century. Appointees to the university system’s governing bodies went from political moderates with a degree of partisan balance to more uniformly Republican, including Republican lobbyists.
  
  As of 2019, the Daily Tar Heel reported that there are now no self-identified Democrats on the Board of Governors; an analysis of campaign contributions by members of the Board of Governors shows a nearly 10-1 ratio of contributions to Republican candidates and causes compared to Democratic candidates and causes3.

• 2014: The legislature requested a review of campus centers statewide, which resulted in closing three of them, two led by outspoken critics of state leadership: the Center for Work, Poverty, and Opportunity at UNC Chapel Hill School of Law and the Institute for Civic Engagement and Social Change at NC Central University.
  
  Although the stated purpose of the review was cost savings, both centers were funded by private sources, so there was no cost saving. These reviews were completely outside the normal channels for assessing the merit of such centers, bypassing the university administration and faculty.

• 2015: The Board of Governors forced out system president Tom Ross, despite having publicly praised his performance. No reason was given, but Ross is a Democrat. He was replaced by Margaret Spellings, George W. Bush’s education secretary; she was pushed out two years later because she was not sufficiently conservative and too independent of the Board.

• 2016: The General Assembly and outgoing Republican governor stripped the incoming Democratic governor of the power to make any appointments to the governing bodies, leaving the university system completely controlled by the General Assembly.

• 2017: The Board of Governors prohibited any campus centers from engaging in litigation. This meant that the system’s law schools could no longer provide experiential learning opportunities for their students, nor could they work on behalf of North Carolina’s most vulnerable citizens.

• 2020: The Board of Governors implemented a new policy for appointing chancellors of the university campuses: the system president (appointed by the Board) names two candidates for the position, one of who must be a finalist for the position. In other words, regardless of the outcome of the vetting process used by the university’s academic and administrative leadership, one of the president’s candidates must be a finalist for the job.

• 2023: The Board of Trustees for the UNC flagship Chapel Hill campus, directed that a new School of Civic Life and Leadership be formed and staffed with faculty representing a spectrum of viewpoints. There was no consultation with the university’s leadership or faculty.
  
  When the head of the accreditation agency for UNC raised questions and concerns about the process, she received what many would consider a threatening letter from six of the seven Republican members of the NC delegation to the House and both of NC’s Republican Senators. Pretty heavy-handed for what was claimed to be an apolitical process.

The bottom line is that North Carolina’s public university system is now run entirely by Republicans who insert themselves directly in the running of the university system.

Florida and North Carolina are harbingers of things to come in other GOP-controlled states.

Denigrating Expertise

The pandemic demonstrated vividly how fascist politics denigrates expertise. We saw Republican politicians across the country downplay the expertise of medical and public health officials, replacing it with the blustery statements of political hacks, ivermectin advocates, and a president encouraging top health officials to investigate injection of bleach as a Covid treatment.

When politicians denigrate real experts — like scientists — they are then free to create their own realities that comport with the false reality they are trying to create.

Unreality

Repeat a lie often enough and it will be believed. I’ll spare you a recitation of the many lies that Trump and his many surrogates created and repeated, which we are still dealing with

Even outlandish conspiracy theories can be effective in creating an alternate reality. Remember Pizzagate, the theory that leaked emails from Hillary Clinton’s campaign manager spelled out coded messages about trafficking children for sex to Democratic congressmen, all run from a Washington pizzeria? When an armed North Carolina man showed up at the pizzeria to free the sex slaves, he was condemned even by the perpetrators of Pizzagate. Their goal wasn’t for the conspiracy theory to be taken seriously as truth but, rather, to plant the subliminal message that Democrats are depraved.

We want to believe that our politics is a “marketplace of ideas”, may the best ideas win. But in fascist politics, words are used to elicit emotion, not to convey information. Stanley explains (p. 76):

“Disagreement requires a shared set of presuppositions about the world. Even dueling requires agreement about the rules. You and I might disagree about whether President Obama’s healthcare plan was good policy. But if you suspect that President Obama was an undercover Muslim spy seeking to destroy the United States, and I do not, our discussion will not be productive. We will not be talking about the costs and benefits of Obama’s health policy, but rather about whether any of his policies mask a devious antidemocratic agenda.

That’s why the many websites and social media posts that promote ridiculous conspiracy theories are so effective. They distract from debate about real issues, replacing such debate with distrust.

Politicians like Trump who speak their mind with neither restraint nor apology come across as authentic because they “tell it like it is”. He lies constantly, but since he is perceived as authentic his supporters believe the lies. His authenticity is enhanced when he targets groups his base dislikes.

In this environment of unreality, he can tell poor white people that he’s helping them by targeting people they resent or hate — Muslims, immigrants, liberals, Blacks — while, in reality, promoting policies like tax cuts for the wealthy that harm those very same poor white people.

Hierarchy

Natural Hierarchies

Fascist politics argues that there are natural hierarchies for the inherent worth of human beings, such as:

• The divine right of kings.

• Alexander Stephens’4 1861 Cornerstone Speech: “the great truth that the negro is not equal to the white man.”

• Hitler’s “Aryan Master Race”

• Epistle to the Ephesians 5:22-24: “Wives, submit to your own husbands, as to the Lord. For the husband is the head of the wife even as Christ is the head of the church, his body, and is himself its Savior. Now as the church submits to Christ, so also wives should submit in everything to their husbands.”

These (and many more) hierarchies are justified by appeals to supposed science and/or to religion. Stephens, for example, went on to say “They [northerners] were attempting to make things equal which the Creator had made unequal.”

Since some people are inherently worth more than others, society, under a fascist regime, has no obligation to treat people equally.

Applying Hierarchy in GOP Politics

We are now too “sophisticated” to say out loud that white people are superior to Black people. Instead, our politicians speak in innuendo, at least in public. AP News summarizes the history of GOP racial code in At RNC, GOP echoes racial code of Nixon’s 1968 campaign.

Richard Nixon’s Southern strategy used fear of crime and lawlessness as code to appeal to Southern segregationists. Ronald Reagan complained about “welfare queens.” George Bush (the Elder) used Willie Horton as a euphemism for Black people on crime rampages. Coded appeals to racism continue in the form of states rights and cutting social services, dividing people into “deserving” and “undeserving”, “makers” and “takers”, “hardworking” and “lazy”.

In their embrace of Christian evangelicals, GOP politicians reject feminism, appealing to the “natural” hierarchy that wives should submit to their husbands, and more broadly, that men are capable of determining what is good for women, including making decisions for them on reproductive healthcare.

Appeal to Loss of Hierarchical Status

Why are appeals to hierarchy so powerful in politics? People who hold what used to be dominant positions in society — white, Christian men in ours — see liberal support for equality among people as a threat to their privileged positions.

So, when Donald Trump speaks with pride about his sexual depravity, he’s appealing to other men who are angry that they’re losing their status as breadwinners, family patriarchs, and being superior to those “other people.” These angry men want to feel that they can dominate women and other groups as was typical in the past.

When he says “there are fine people on both sides” in response to the white supremacist violence in Charlottesville, he’s again supporting the whites-better-than-blacks hierarchy and the sense of loss the white supremacists feel.

Stanley explains (p.. 97) how loss of hierarchical status, especially in the context of a nation that was once unrivaled economically and militarily but is no longer, feeds into fascist politics:

“Empires in decline are particularly susceptible to fascist politics because of this sense of loss. It is in the very nature of empire to create a hierarchy; empires legitimize their colonial enterprises by a myth of their own exceptionalism. In the course of decline, the population is easily led to a sense of national humiliation that can be mobilized in fascist politics to serve various purposes.”

Victimhood

Suppose that, like me, you’re a man who grew to adulthood in the 1970s. At that time, the working world in the US was pretty much a man’s world. In 1970, 61.9% of the labor force were men and 38.1% women, with most of the women relegated to traditional women’s roles (teaching, nursing, clerical)5. As more women entered the labor force, by 2015 53.2% of the labor force were men and 46.8% women, with women supervising men commonplace.

A man struggling for career success might feel like a victim, thinking that he’s having such a hard time because those liberals pushing for equality of men and women are taking jobs away from people like him and giving them to women.

Indeed, fascist politicians encourage such men to feel like victims.

Likewise, many white Americans think that racism is a thing of the past, so why are we still helping Blacks at their expense? Fascist politicians will explain that they’re victims of those liberal’s do-gooder civil rights policies; maybe they were needed in the past, but certainly not now.

But, as Stanley explains (p. 103):

“Today, white Americans wildly overestimate the extent of U.S. progress toward racial equality over the past fifty years. Economic inequality between black and white Americans is roughly at the point it was during Reconstruction; for every $100 the average white family has accumulated, the average black family has just $5; and yet, as Jennifer Richeson, Michael Kraus, and Julian Rucker have shown in their 2017 paper, “Americans Misperceive Racial Economic Equality,” white American citizens are widely ignorant of this fact, believing that racial economic inequality has dramatically narrowed.”

Similarly, as reported in the Atlantic article Most American Christians Believe They’re Victims of Discrimination,

“Almost half of Americans say discrimination against Christians is as big of a problem as discrimination against other groups, including blacks and minorities. Three-quarters of Republicans and Trump supporters said this, and so did nearly eight out of 10 white evangelical Protestants. Of the latter group, six in 10 believe that although America once was a Christian nation, it is no longer.”

This, despite the fact that 71% of American adults identify as Christian and that the Robert’s Court has handed conservative Christians victory after victory.

If you’re a Christian who believes you’re being discriminated against, it is easy to fall into victimhood, especially when being encouraged to do so by fascist politicians.

Victimhood is a powerful emotion, easily manipulated by fascist politicians, as Stanley describes (p. 112):

The fascist leader employs a sense of collective victimhood to create a sense of group identity that is by its nature opposed to the cosmopolitan ethos and individualism of liberal democracy. The group identity can be variously based—on skin color, on religion, on tradition, on ethnic origin. But it is always contrasted with a perceived other against whom the nation is to be defined. Fascist nationalism creates a dangerous “them” to guard against, at times to battle with, to control, in order to restore group dignity.

Law & Order

The specter of widespread criminality is a powerful emotion that politicians exploit even when it is false. During the 2020 campaign, Donald Trump campaigned with claims that the 20 cities with the highest crime rates were all run by Democrats. Be scared, elect him and Republicans, and make America great again. Well, yes, but most cities elect Democratic leaders.6 Unbidden, he sent a “surge” of federal agents into cities to attack crime7.

With all of this talk of crime, you might think that America was overrun by violent crime. In fact, violent crime rates were (and are) near historical lows.

As we discussed earlier, law and order combined with racism has been a major component of Republican politics since at least the time of Nixon.

Sexual Anxiety

There is one kind of violent crime that is most effective for targeting “them”: “they” are going to rape your women. During Jim Crow, the vision of white women being raped by oversexed Black men was the justification for many lynchings.

Trump’s 2015 (false) claim of “Mexican rapists” coming across the border was no accident and followed in this tradition. He was playing to sexual anxiety, implying that he and his wall would protect “your” women against this terrible threat.

Republicans across the nation are now building on such anxiety in their systematic attack on LGBTQ people:

• Transgender girls are going to rape your daughters in bathrooms at schools

• Gays are grooming your children for sexual abuse

• Reading books in school about non-traditional family structures (two mommies or two daddies) will make your children gay

• Marriage among homosexuals destroys the fabric of society.

Stanley (p. 146) aptly summarizes the situation:

“By employing the politics of sexual anxiety, a political leader represents, albeit indirectly, freedom and equality as threats. The expression of gender identity or sexual preference is an exercise of freedom. By presenting homosexuals or transgender women as a threat to women and children—and, by extension, to men’s ability to protect them—fascist politics impugns the liberal ideal of freedom. A woman’s right to have an abortion is also an exercise of freedom. By representing abortion as a threat to children—and to men’s control over them—fascist politics impugns the liberal ideal of freedom. A person’s right to marry whom they wish is an exercise of freedom; by representing members of one religion, or one race, as a threat because of the possibility of intermarriage is to impugn the liberal ideal of freedom.”

Appeals to the Heartland

In GOP imagery, family farms in the heartland of America are the soul of our nation, with rural communities bastions of individual freedom, independent thinking, and hard work. Never mind that mid-sized family farms have mostly been replaced by agribusiness or that according to the 2020 census, 80% of the US population lives in urban areas.

As Hitler did before him, Trump and other GOP leaders denounce cities as, to use Trump’s words, “hellholes” and “burning and crime-infested inner cities of the US.”

Why? Because such attacks resonate with rural people, and some suburbanites, and most urban voters have already been lost to the Democrats regardless.

Stanley (p. 155) explains:

“Fascist politics aims its message at the populace outside large cities, to whom it is most flattering. It is especially resonant during times of globalization, when economic power swings to the large urban areas as centers of an emerging global economy, as occurred in the 1930s in Europe. Fascist politics highlights the wrongs a globalized economy does to rural areas, adding to it a focus on traditional rural values of self-sufficiency supposedly put at risk by the success of liberal cities culturally and economically.”

Facing the Danger

Our democracy today is extraordinarily fragile. GOP politicians at national and state levels are deploying all of the fascist tactics I have described. Look at the laws being proposed in GOP-controlled states and you’ll see that they are largely the same. The rhetoric is all the same. It is a coordinated effort.

Over the last 40 years, GOP politicians have worked with oligarchs to shape our laws and our justice system to support their program. The effort has gained extraordinary momentum and our political and legal systems are being changed to further accelerate the effort. Gerrymandering has been deployed with ruthless efficiency to give the GOP control of government in states that are fairly evenly divided. Voter suppression efforts at the state level are rampant. Money is now the primary source of power in our political system. We now have a reactionary Supreme Court that supports the march toward the good old days.

Fog by Carl Sandburg

The fog comes
on little cat feet.

It sits looking
over harbor and city
on silent haunches
and then moves on.

Fascism comes in on little cat feet. Unlike Sandburg’s fog, it doesn’t move on. Instead, it gets normalized. We now accept from leading GOP politicians a level of lying and manipulation that was previously unthinkable. We accept unthinkable levels of gun violence and attacks on our brothers and sisters based on their sexual preferences and gender identities. We talk of individual freedom but elect male autocrats like Donald Trump and Ron DeSantis, and other GOP politicians who impose their will on half our population without their agreement and against medical advice.

The first step in fighting to preserve democracy is to sound the alarm. This article is my attempt to sound the alarm with you, many of whom I know to be thoughtful, caring people.

As I said before, there is no win-win about this. And there is no both sides: yes, Democratic politicians have plenty of faults, but only GOP politicians are currently working in a concerted effort to destroy our democracy and replace it with — well, with whatever you want to call it. Fascism. Authoritarian oligarchy. It doesn’t really matter. We must stop it.

Leave a comment

Welcome Back to Win-Win Democracy

As we’ve discussed previously, AI is impacting society for better and worse, a process that will continue and probably accelerate. We, as a society, must decide if and how to manage those impacts. My thinking on this subject derives from six core beliefs.

Core Beliefs

1. There Is No Stopping or Slowing AI

Proposals to impose regulations that slow or stop the development of AI are naïve and will not succeed.

AI is the “next big thing,” the Internet of the 2020’s. The CEOs of the dominant technology companies have already shifted huge investments to AI and we see aggressive competition unfolding in real time. Investors are throwing money at AI startups and less-dominant large companies (e.g., IBM) are looking at AI as a next-wave opportunity.

Congress can barely get out of its own way and is unlikely to mount a serious “slow/stop AI” effort. Lobbyists and PACs would be out in force to either oppose or shape to their advantage any such effort.

More importantly, AI is viewed as so critical to both defense and economic growth that the if-we-don’t-do-it-the-Chinese-will argument will hold sway.

2. Technology Diffusion Dynamics Apply to AI

AI technology will diffuse into society pretty much like other highly-impactful technologies have — slowly.

Probably the most significant modern technology diffusion was replacing human and animal power by first steam engines, then internal combustion engines, and eventually electricity delivered over a transmission network. Depending on how one counts, this took around 150 years.

It took computing technology more than 40 years to become a major part of our personal and economic lives.

AI’s impact may happen more rapidly, but we are still talking decades not years.

3. AI Is a Tool

AI is a tool that makes it easier and faster to accomplish certain tasks, increasing economic productivity and affecting the work that humans do and the products and services that corporations produce.

We’ve seen this with other tools and can expect the same sorts of effects:

• AI will eliminate some jobs and destroy some individuals’ careers.

• AI will create new jobs and career opportunities, probably increasing overall economic growth.

• AI will destroy some companies and professions.

• AI will be the basis for new companies and professions.

Ultimately, AI will increase economic activity, but many people and companies will be hurt during the transition.

AI-based tools will help other people in their existing professions. For example, I know coders who already use ChatGPT to draft first versions of code, saving them time but not (yet?) eliminating their work.

There’s some evidence that AI-based image and pattern analysis could help significantly with medical diagnosis, improving both speed and accuracy.

4. AI Will Exacerbate Existing Societal Problems

AI is a force multiplier for people and organizations that want to spread disinformation or perpetrate scams. We’ve seen examples before. AI didn’t create disinformation or scams, but will exacerbate both, increasing the likelihood that our democracy will fail to survive.

Similarly, AI will worsen income and wealth inequality because its powers will be most available to people and organizations with access to the skilled people and the massive computing resources needed to train and deploy AI systems.

And, of course, as AI disrupts people’s livelihoods, there are going to be additional people — including families — needing help.

5. AI’s “Thinking Ability” is Mostly Hype

As impressive as they are, today’s AI’s don’t understand and don’t think logically.

There is a substantial possibility that leaders in business and government will be sufficiently disappointed with their experiences trying to use AI that the vast, broad impact expected of AI will not come to fruition.

This is a longer discussion, which I’ll delve into below. In the meantime, you might recall the examples I gave previously here and here.

6. Our Intellectual Property and Personal Responsibility Laws are Inadequate for AI

When ChatGPT defames someone (see here), who is liable? Is anyone liable? When an AI “writes” prose or music or “creates” art, can those works be copyrighted? If so, who owns the copyright — the company that created the AI or the person who crafted the prompts to the AI that resulted in the work?

When an author’s work is plagiarized by ChatGPT or an artist’s drawing is incorporated without permission into a generated image, is the artist due compensation? Procedurally, how would this work?

When an AI summarizes an article published online without sending the reader to the online publication, is the article’s publisher entitled to compensation in lieu of the advertising revenue that the publisher ordinarily receives when someone reads the article?

When a “conversation” with an AI chatbot sends an unstable or depressed person into suicide, who is responsible?

You can no doubt think of many similar questions.

Defending Against Disinformation

Disinformation purveyors are going to love AI. AI will make producing disinformation less labor intensive, cheaper to produce, and probably more effective.

Low-Tech Disinformation

But here’s the thing: Disinformation is already a huge threat to democracy around the world. It is spread via all sorts of media at all levels of politics.

Low-tech disinformation is effective. For example, in the 2022 election for the NC House in a district near where I live, Ricky Hurtado, a Democratic incumbent, made the mistake of volunteering to do cleanup work in a local park. He was photographed and here’s what his Republican opponent and his allies did:

Hurtado lost the close election. The Charlotte Observer’s opinion piece A new level of dishonesty: Mailers targeting NC Democrats photoshop the truth says that “[m]ailers recently distributed in several competitive North Carolina House districts feature deceptively edited photos of Democratic candidates, tying them to the ‘defund the police’ movement and Black Lives Matter protests.”

AI Forces Fighting Disinformation Broadly

Disinformation is spread on television, YouTube, social media, newspapers, magazines, books, political mailings, billboards, etc. AI will make it worse (or, better, if you’re in favor of using disinformation in politics and business).

Other than making it cheap to create and spew ever-increasing flows of disinformation, AI doesn’t fundamentally change the problem that we already have.

Here’s the positive spin on it: Perhaps the advent of AI will make it so much worse that we finally fight disinformation — all of it — not just AI-assisted disinformation.

Fighting Disinformation

Certain kinds of disinformation are protected by our free speech laws. If a politician (or an organization) wants to say that violent crime is rising across the nation so you should be very afraid, that the crime is caused by rapists streaming into the country from Mexico, that you need guns to protect yourself and your family, and that the Democrats are coming to take your guns away, the politician can do that. The politician can say it out loud on TV or insinuate it anonymously through media. Regardless, there’s no accountability for the lies.

Take it further, and intentionally and knowingly defame a particular person and cause harm to that person, then there is recourse through defamation laws. Realistically, accountability through defamation and libel laws is slow, impossibly expensive for most people, and fraught with risks.

E. Jean Carroll might have been able to hold Donald Trump accountable for defaming her (we’ll see how the delays and appeals will go) and Dominion Voting Systems somewhat held Fox News Corporation accountable for defaming them, but certainly the Ricky Hurtado’s of politics can’t do so. And lawsuits that take years can’t undo the effect of defamation on the outcome of an election.

I don’t see a legal solution to this situation.

This is a cultural issue, where it has now become acceptable for major players in our political combat to fully take off the gloves. Until the leaders of all political parties set the right examples, this situation will persist.

Section 230 and Online Disinformation

In the “good ol’ days” of broadcast TV and radio, the Federal Communications Commission’s Fairness Doctrine required broadcasters to tackle controversial topics of public interest and to do so in a way that reflected multiple viewpoints. The Fairness Doctrine was repealed in 1987, quickly giving rise to talk radio hosted by disinformation spreaders like Rush Limbaugh. Broadcasters were no longer obligated to present contrasting viewpoints.

As a non-broadcast service, cable TV was never subject to the Fairness Doctrine.

Along came the World Wide Web in the 1990s. Congress passed Section 230 of the Telecommunications Act of 1996, granting “interactive computer services” immunity from liabilities arising from providing a platform for third-party content providers. In other words, Facebook or Twitter can not be held liable for the scurrilous lies that your nutty uncle or Donald Trump or anyone else posts on their platform.

Section 230 was seen in the 1990s as necessary to enable the growth of the then-nascent platforms.

Interestingly, Congress saw fit in 1998 to limit Section 230’s ability to protect platform providers from liability for copyright infringement by their third-party content providers. Guess what? Facebook, Google, and the like actually police their platforms for copyright infringement.

These platforms no longer need protection to grow. They are already the dominant players in their industry.

Congress could, therefore, further limit Section 230’s ability to protect platform providers from liability for publishing false or defamatory third-party content. This would have to be done carefully and certainly would impose costs on the major platform providers.

It would also open them to accusations of bias, a problem they already face because their algorithms for choosing what to display to their users are opaque and not accountable for anything other than raising user engagement — and hence advertising revenue — regardless of impact on society.

Suspend disbelief for a moment because, below, I’ll propose a win-win that combines this idea with what I’m going to discuss next.

Verified Responsible Parties

All social media platforms grapple with the “fake account” problem, accounts that purport to be owned by a particular individual or organization, but, in fact, are not.

Humans, and sometimes robots, operate through fake accounts to sow disinformation or perpetrate scams.

A good solution to this problem is verification of the identity of the accounts’ owners. You’ve probably read about pre-Musk-Twitter’s blue check mark indicating that Twitter has verified the identity of the owner. Pre-Musk, this was limited to well-known people or organizations. As a consumer of information, the blue check mark indicated that you could trust the source. If you read a tweet from NPR and saw the blue check mark you could be sure that you could trust that content (to the extent that you trust NPR).

But Musk destroyed this approach for Twitter by converting the blue check marks into an $8/month revenue stream rather than a real verification.

Facebook has a verification feature, too, which allows “notable” people or organizations/brands to establish who they are. Once verified, their posts get a verified check mark.

Now, imagine the next steps:

1. All social media would offer verified responsible party indicators to all people and organizations. Being “notable” would no longer be required to be eligible for being verified. But verification would require proving identity and the account profile would be required to be listed in the verified name. (I’ve previously discussed in a related context some ideas for how to do verification and how to pay for it. Technology has marched on since I wrote that, so other approaches should be considered too.)

2. All users of a social media platform could specify that they only want to see content from verified accounts. This should be the default. That way, users would not see the disinformation from, for example, Russian agents sowing dissent while purporting to be fellow Americans.

3. Social media algorithms that decide what content to show would favor content from verified parties.

Suppose we did this. Now, if I want to use AI to generate my posts, that’s fine, but I will be personally accountable for their content. Likewise for corporations and other organizations.

Now, go one step further: If we limit Section 230’s ability to protect platform providers from liability for the the third-party content they publish, this would motivate platform providers to implement verification as a means to reduce the platform’s liability risks.

Win-Win Proposal

In a nutshell, my win-win proposal for controlling online disinformation is:

1. The advent of AI makes solving our online disinformation problem more important and urgent.

2. Treat online disinformation as one problem, not as an AI-specific problem.

3. Amend Section 230 as follows:

   1. Make “large” online providers liable for third party content that they publish.

   2. Provide a safe harbor1 for online providers exempting them from liability for content from verified parties.

The win-win is that society benefits from better control of online disinformation and the social media platforms remain protected from liability as long as they take the necessary steps to broadly implement verification.

Helping Displaced Workers

There is a lot of (justified) discussion about the harm AI could cause to individuals as their jobs and career paths are affected by deployment of AI in many businesses. We live in a capitalist society in which, for better or worse, the pursuit of profit is the fundamental driving force in our economic system.

In this system, attempts to slow change to limit impact on workers generally fail. AI, as a new technology, will be no different. Perhaps its pace will be faster than the diffusion of other impactful technology, but even that’s not certain.

History would say that we can not and should not attempt to slow down AI-based impact on jobs. If we artificially slow AI deployment as a tool in business, then our businesses will be at a significant disadvantage compared to their competitors in countries that choose not to slow AI deployment.

If we shouldn’t slow the deployment of AI, how do we reduce harm to individuals’ livelihoods?

We help people directly. We learned during the darkest days of the Covid pandemic that significantly enhanced unemployment benefits can make a huge difference in people’s lives. If AI drives a big downturn in jobs we should generously use benefits to help people get through it.

We should not require people to prove that AI was why they lost their job: The connection between AI deployment and losing a specific person’s job will be nebulous and impossible to prove.

We can, however, monitor the deployment of AI, the change in jobs and wages, and the growth in corporate profits and GDP that results. We would pay for enhanced benefits out of the growth in GDP and corporate profits.

Enhancing benefits will be controversial. There will be the laziness scolds and the debt scolds. And, yes, there will be waste.

But, if AI makes its predicted big impact, there will be hungry and angry people who will need society’s help. Beyond human decency, we must help them to avoid destroying our already fragile democracy. We’ve shown during Covid that, however flawed, we can create enhanced benefits that help real people without harming the economy.

Additionally, relevant job training opportunities should be well funded.

Reducing Concentration of Wealth & Power

In their new book, Power and Progress: Our 1000-Year Struggle Over Technology & Prosperity, released just this week2, MIT economists Daron Acemoglu and Simon Johnson review the long history of the struggle over whether and how to share the benefits of new technologies. History suggests that the benefits of new technologies are not shared broadly until something forces a change.

In Chapter 11, they articulate a “critical formula necessary for escaping our current predicament”, by which they mean (p. 447) “the enormous economic, political, and social power of corporations, especially in the tech industry.”

They go on to say (p. 453):

“The concentrated power of business undercuts shared prosperity because it limits the sharing of gains from technological change. But its most pernicious impact is via the direction of technology, which is moving excessively toward automation, surveillance, data collection, and advertising. To regain shared prosperity, we must redirect technology, and this means activating a version of the same approach that worked more than a century ago for the Progressives.”

The formula is three-fold:

1. “Altering the narrative and changing norms.” Citizens need to develop an informed view of what is troubling the country economically, not “just accepting the line coming from lawmakers, business tycoons, and the yellow journalists allied with them.”

2. “Cultivating countervailing powers.” Corporate powers must be held in check by broad movements advocating reform, including workers’ unions and aggressive enforcement of anti-trust laws.

3. Policy solutions. For example, as we’ve discussed in prior newsletter issues, our tax system is highly skewed to reward capital over labor. Policies that allow accumulation of massive wealth and its concomitant power must be changed.

AI is yet another technology in a long list of technologies that have changed the world. In most respects, the lessons learned from the introduction of those other technologies hold today.

Starting with the Reagan Revolution of the 1980s, we’ve allowed a dramatic shift of wealth and power to the few at the expense of the many, as we’ve discussed previously many times (see https://winwindemocracy.org/t/wealth). Despite a lot of talk, Congress and our political leaders have been unwilling or unable to make any progress on this concentration of wealth and power; indeed, they’ve changed the tax system to further concentrate wealth.

AI will accelerate the already egregious concentration of wealth and power. But it is still the same old problem.

The question is whether the effects of AI will be the straw that breaks the camel’s back, driving us to finally pay attention and start the decades-long effort to defeat our modern Gilded Age by “altering the narrative and changing norms” and ushering in a modern version of the Progressive Era of the early 1920s.

Devaluing Creativity

I mentioned earlier that AI raises intellectual property issues that current law (understandably) didn’t anticipate. I could easily write a whole newsletter issue on just that topic. I will avoid the temptation to do that; instead I refer you to the article The scary truth about AI copyright is nobody knows what will happen next, which appeared late last year in The Verge. It provides a good overview easily understood by non-lawyers.

Copyright and Fair Use

That said, I want to discuss one important topic here. United States copyright is authorized directly in the Constitution’s Section 8, the enumerated powers of Congress section:

“[The United States Congress shall have power] to promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.”

The purpose is clear: “to promote the Progress of Science and useful Arts”. It does not say “to protect the livelihood of authors and artists and inventors.” But if such creators can’t earn a living, such work is not going to be produced.

In the struggle to balance the public’s need to benefit from “the Progress of Science and useful Arts” with the creators’ needs to earn a living, copyright law has evolved the concept of fair use. A fair use of a work is not an infringement of copyright law, and the creator can not sue for compensation or control.

The Copyright Act of 1976 dictates, based on prior court decisions, that

“In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include:

1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;

2. the nature of the copyrighted work;

3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and

4. the effect of the use upon the potential market for or value of the copyrighted work.

The fact that a work is unpublished shall not itself bar a finding of fair use if such finding is made upon consideration of all the above factors.”

There have been many court cases that have shaped and continue to shape the practical meaning of these four factors.

An Example

I do the grilling in our family. I want to know how to grill London broil.

Asking Google yields:

To see a recipe, I choose one of the articles offered, click on the link, and head over to the web page.

My visit to the web page might be monetized in various ways, such as advertising, affiliate links to vendors3, or simply by boosting the author's reputation or the site’s standing in searches. Regardless of the mechanism, the page’s author and/or publisher benefits from my visit.

Now, look what happens when I ask the same query of new Bing (which uses some combination of conventional search and GPT-based AI):

At the bottom, you can see the steps of the recipe and scroll through them. No need for me to visit the website at all. Unless Microsoft has made some agreement with the website’s owner, the website and the recipe author have no opportunity to monetize my use of the information on the page.

It gets worse directly using ChatGPT:

Seems like a great recipe. I have no idea where this came from. Is it from a famous chef or from ChatGPT’s synthesis of a recipe from among the many it has “read”? I couldn’t visit the web site even if I wanted to. Whoever provided this recipe (or parts of it) has no opportunity to monetize their work.

Likewise, Google’s Bard yields:

It gives me “drafts” of three recipes from which to choose, again with no way for the authors of these recipes to monetize their work.

Is this “fair use” of the intellectual property of the people who provided the web sites from which ChatGPT (and, hence, new Bing) and Bard came up with recipes? It wouldn’t seem so given consideration of the fourth factor of the fair use law.

Killing the Goose that Laid the Golden Eggs

AI’s that suck up information and ideas from the existing content on the Web, summarize it for easy consumption or use it to synthesize derivative content, might very well destroy future creativity be killing the incentive for people to create new content. A reader of this newsletter has told me that he is seeing a decrease in visitors to his website, which he attributes to this phenomenon.

Devaluing creative work is, in my mind, one of the top risks of AI as it is currently being deployed. AI is already good enough to have massive impact on the ability to monetize new works.

I don’t have a solution to propose, but I think that this problem is not discussed enough. It is an urgent problem. Think back to the near-death of the music industry and the threat to musicians’ livings when digital music came into being.

Does AI Think? Does AI Understand?

Inevitably, in discussions about AI among my friends, these two questions are front and center, either explicitly or as an implicit assumption that AI technology can think and understand. We’ve discussed this in earlier newsletter issues, but I’m raising it again here because I think most people still don’t really understand this important issue.

AI’s based on large language models (i.e., GPT-like systems) have an amazing facility with language. But we humans are gullible. Because ChatGPT can generate reasonable — even excellent — responses to our prompts, we are naturally inclined to believe that the AI understands and thinks. Remember the shocking reaction to ELIZA/DOCTOR in the 1960s that we discussed previously?

So, do they think and understand? Not really. In previous newsletters, we’ve seen many examples where the responses to prompts are well-written, with a confident tone (or, even another tone if requested), but factually nonsense — what industry jargon calls hallucination. Note the anthropomorphism. Yet another reason we are gullible.

We’ve also seen an example in which ChatGPT is unable to solve middle-school level word problems.

Here’s another simple example:

Prompt: how many integers are there?
Response: The set of integers is infinite. There is no limit to the number of integers because you can always find another integer by adding or subtracting 1 from an existing integer. In mathematical notation, the set of integers is denoted by the symbol Z and it includes all positive and negative whole numbers, as well as zero (0). So, while there isn't a specific count of how many integers there are, we can say that the set of integers is infinite.

This response was fine, if overly long. From this response, you’d think that ChatGPT “understands” integers and the concept of countably infinite sets.

So, let’s follow up (in the same chat session) on that “understanding”:

Prompt: What are the ten largest even numbers?
Response: The ten largest even numbers are as follows:

1. 10,000,000,000
2. 9,999,999,998
3. 9,999,999,996
4. 9,999,999,994
5. 9,999,999,992
6. 9,999,999,990
7. 9,999,999,988
8. 9,999,999,986
9. 9,999,999,984
10. 9,999,999,982

These numbers are in descending order, starting from the largest even number, which is 10,000,000,000, and continuing with each subsequent even number.

This response is utter rubbish. It shows that ChatGPT doesn’t understand at all.

And, just to be sure that this isn’t a ChatGPT idiosyncrasy, I tried the same prompts with Bard. It produced a different but also good response to the first prompt, and a different but also utter rubbish response to the second prompt.

Don’t conclude that a well-written response to a prompt indicates “understanding” or “thinking”.

This is particularly important to keep in mind as you evaluate for yourself various predictions and claims about how AI is going to affect our future.

What’s Next?

I plan that the final (at least for now) topic in this series of posts on AI is regulation. I am particularly concerned about the calls from the industry for government regulation.

Despite the nice words written and spoken, I’m skeptical that those calls for regulation are for the benefit of society. Rather, I have the admittedly cynical view that the requests by the industry for regulation are more about establishing safe harbors for what they want to do than about imposing meaningful limits on what they actually do: With suitable regulations in place, when something bad happens as a result of their technology, the response will be “it [the something bad] is really unfortunate, but we complied with all of the federal regulations”.

But first … there is another threat to democracy that concerns me even more than AI concerns me. I will address that in the next newsletter issue, then return to complete the AI series.

Until next time …

Leave a comment

Share Win-Win Democracy

Without further ado:

Leave a comment

Share Win-Win Democracy

Welcome Back to Win-Win Democracy

AI frenzy continues in the media. Building on what we learned in the last issue about how these system work, we turn now to discuss the implications of this technology on society.

Share Win-Win Democracy

Current Impacts

Even in its current state, AI technology is exacerbating problems we already face. If the technology continues to improve as many anticipate, the impacts will intensify over time.

Disinformation Campaigns

Disinformation campaigns are rampant. Fox News Network has demonstrated their ability to get millions of Americans to believe lies. Russian operatives used Facebook, Twitter, Instagram, and other social media platforms to influence the outcome of the 2016 presidential election1.

AI opens up new avenues for disinformation campaigns. ChatGPT and its ilk can be used to bombard social media platforms with disinformation for almost no cost.

More insidious is the problem of deep fakes, AI-generated believable recordings and images of events that never occurred.

This fake image, created as a demonstration by Eliot Higgins, founder of the investigative outlet Bellingcat, purports to be the scene when former-president Trump was arrested by New York police. Of course, it never happened.

Look closely and it’s obviously fake. But someone inclined to believe this happened is unlikely to look closely. Indeed, Higgins’ tweets depicting this fake event have been viewed nearly 5 million times.

A more insidious deep fake, produced by right-wing provocateur Jack Posobiec2, shows a fake video of President Biden invoking the draft to meet demand for US soldiers in response to Putin’s occupation of Kyiv and the impending Chinese blockade of Taiwan, both events that haven’t happened.

Again, a cursory viewing is convincing. After about 45 seconds of the fake video of President Biden, Posobiec goes on to say that this is “coming attractions, a glimpse into the world beyond,” meaning that he made this fake to illustrate where Biden could head if we’re not careful.

Perhaps these two deep fakes are not, themselves, a threat because it is so easy to check the facts about such well-known people. But it doesn’t take much imagination to see how deep fakes could be used in effective disinformation campaigns. Think back, for example, to the role of the 1933 Reichstag fire, which many historians believe was a Nazi false flag operation — in giving the Nazis dictatorial power in Germany.

Could deep fakes, used to place blame on some minority group for a similar event today, be used to justify “emergency” dictatorial powers in the US? Given the manner in which many in the US responded to former-President Trump’s lies about the 2020 election, it seems plausible, even likely, that deep fakes could be used as an important component of a disinformation campaign to overthrow our democracy.

Scams

As you saw in the video of fake President Biden speaking, AI can now generate audio that sounds like a particular individual. Perfect for scams.

Wall Street Journal columnist Joanna Stern created “AI Joanna” using several inexpensive online tools that process uploaded audio files of a few hours of the real person speaking to produce an “audio clone”. AI Joanna was good enough to fool her bank’s voice authentication system.

Real Joanna believes that it would have fooled her sister if her sister hadn’t noticed that the clone didn’t pause to take a breath.

CNN reported on the use of AI-based voice cloning to convince a mother that her daughter was being held for ransom by kidnappers.

As voice cloning becomes more sophisticated, the scamming risks will grow.

Fabrication as a Business Model

As we’ve discussed previously, ChatGPT often writes confidently using “facts” that it fabricates, with the potential to cause great harm to individuals and organizations.

But what a great business opportunity: News websites, written at near-zero cost by ChatGPT (and other chatbots), bringing in revenue from programmatic advertising3.

As reported by Bloomberg News, the news-rating group NewsGuard has identified 49 purported news sites that “appear to be almost entirely written by artificial intelligence software.” Some of the sites summarize (without attribution) content from elsewhere and some appear to fabricate content from prompts.

A few are so devoid of human supervision that their articles include error messages from their AI authors. Check out, for example, this article.

Employment Loss

The media is full of alarming headlines about AI’s impact on jobs:

• Fortune: Goldman Sachs Predicts 300 Million Jobs Will Be Lost Or Degraded By Artificial Intelligence

• CNN: 300 million jobs could be affected by latest wave of AI, says Goldman Sachs

• The Guardian: US experts warn AI likely to kill off jobs – and widen wealth inequality

• ABC News: Is AI coming for your job? ChatGPT renews fears

• and many more

The history of technology adoption is that new technologies displace or eliminate some jobs but create other jobs, but the timing and changing skills needs are such that many individuals’ livelihoods are harmed in the transition.

We’ll discuss AI’s potential long-term impact on employment below.

What’s happening in the short term? As a recent article in the Washington Post described, across many industries there’s experimentation with using AI to augment humans in their work, rather than to replace humans. To some extent, if AI helps humans be more productive, fewer humans could be necessary for doing the same amount of work.

Last week, IBM’s CEO, Arvind Krishna, said that IBM’s hiring in back-office functions, like human resources, will be suspended or slowed. In an interview, Krishna said that “these non-customer-facing roles amount to roughly 26,000 workers” and that “I could easily see 30% of that getting replaced by AI and automation over a five-year period.”

Be skeptical about that pace. Such technology transitions are always slower than predicted because there are many unforeseen or under-appreciated impediments to adopting new technologies.

For example, in 2016, Krishna’s predecessor, former IBM CEO Ginny Rometty, in an interview about using IBM’s Watson AI in cancer treatment, told CBS News that “I think in the next five years, you’ll use this kind of technology to make almost any important decision.” Rometty added, “and it could be around the weather, it could be around education, it could be around shopping, but at the other end, it will be about risk, finance, whether it’s anything to do with anything complex in a system in our world that’s out there.”

Neither the use of Watson AI in cancer treatment nor her predictions about Watson AI’s broader use have come to fruition4.

Nevertheless, the mere anticipation that 7,800 positions could be turned over to AI automation over five years, is leading to a hiring slowdown today, even before it has been demonstrated that AI can replace people in these roles.

Businesses Disrupted by AI

We know that new technologies can destroy old businesses. Think digital cameras and Kodak. Usually an old business’s demise plays out over many years, and a few threatened companies even manage to pivot to remain relevant even in the face of the new technology.

But even the threat of AI seems to have quick impact on some companies. Bloomberg financial columnist Matt Levine reported that the stock of Chegg, Inc., a homework assistance company, plummeted 42% on Tuesday after it filed a securities offering document that identified ChatGPT as a business risk.

Now, stock prices are not the same thing as the company’s business. But the company’s CEO said that “since March we saw a significant spike in student interest in ChatGPT. We now believe it’s having an impact on our new customer growth rate.”

Intellectual Property Issues

ChatGPT was trained on vast amounts of text and computer codes written by humans, who never anticipated that the intellectual property that they created would be incorporated into ChatGPT with neither attribution nor compensation. Likewise, generative AI programs that create images and videos were trained on art produced by humans, again without attribution or compensation. Who owns the intellectual property created by such programs?

None of those humans, nor corporations to which intellectual property rights had been assigned, gave permission for their intellectual property to be used in these ways.

The same AI technology that can clone voices for use in scams, can similarly clone voices of singers, voice actors, and other well-known personalities. A Washington Post article describes the situation of Michelle Clarke, a voice actor who learned that a cloned version of her voice was available inexpensively through an online service.

Similarly, the New York Times reported on the music track “Heart on My Sleeve,” which went viral on TikTok, Spotify, and YouTube. It was written by AI and used clones, without permission, of the voices of two popular musicians, Drake and The Weeknd.

Future Impacts

The impacts we discussed in the last section, which are already underway, will continue and probably accelerate. The potential future role of AI-based disinformation campaigns further destabilizing our democracy is particularly concerning. All of these impacts are part of our future even though I’m not going to discuss them further in this section.

The big question is whether today’s AI frenzy is the start of a revolution that’s going to change everything quickly; or is it the beginning of a decades-long gradual evolution of our economic and social governance systems; or, is it a flash in the pan that we’ll all shake our heads about in a few years.

Old people like me remember shaking our heads about the Japanese Fifth Generation Computing Project of the 1980s, which was going to revolutionize computing and programming with AI. It was going to destroy the American computer industry and bring Japan to worldwide economic dominance. Didn’t happen.

Likewise, so-called expert systems, again in the 1980s, were going to replace skilled workers of many kinds, including doctors and lawyers. It never happened. Instead, bits and pieces of technology made its way into various products, but never revolutionized anything.

Is this time different? Yes, I think so.

Despite today’s AI’s many flaws, which we’ve discussed in previous issues of the newsletter, AI technologies based on artificial neural networks are already solving real problems like speech recognition, language translation, image classification, and more. And, I don’t mean in research settings, but in practical implementations used by probably hundreds of millions of people. Not only aren’t they going away, they’re going to continue to improve.

Put another way, the success of AI technologies to solve certain particular, narrow problems or to be used as tools assisting human beings with various tasks is assured. It has already been underway for a decade and the history of technology diffusion would suggests that will continue for a few more decades as companies and their employees learn how best to deploy the technologies.

The bigger question is whether AI technologies will prove revolutionary, in the sense of causing rapid, disruptive, change throughout society, and, if they do, can that change be managed in a way that most people would consider positive.

Let’s start that discussion with the worst-case scenario.

AI Annihilates Humanity

Science fiction has given us characters like Star Trek’s Lieutenant Commander Data, an android (synthetic humanoid) endowed with super-human artificial intelligence and superior physical capabilities, but lacking in human emotions. The Data character was a positive presence among his human colleagues, but his older brother Lore, created by the same fictional human scientist, was malevolent.

Could an army of self-replicating Lore-like androids use their superior intelligence and physical capabilities to destroy humanity? Perhaps, but today’s AI is far from giving us real-life Data and Lore androids, or even disembodied artificial brains, what some call artificial general intelligence.

A more likely annihilation scenario involves AI-controlled weaponry gone wrong or used malevolently. Humans already use remotely-controlled robotic weapons in warfare. It is a relatively small step to use AI for automatic weapons targeting and release, especially in warfare situations in which occasional mistakes might be tolerated.

Even an unintentional AI-launched attack on another world power could provoke a humanity-annihilating nuclear weapons exchange. Can we prevent that? Perhaps.

Fundamentally, this scenario is another instance of global weapons control. Humanity has managed to avoid nuclear conflagration for 70 years using a variety of diplomatic, military coordination, and economic approaches. It is essential for these approaches to be updated in the light of the potential for AI-controlled weaponry5.

Long-Term Economic Impacts

Accurate predictions about the future are impossible. Most predictions of technology adoption are overly optimistic, but there are also plenty that have been wrong in the other direction. Nobel laureate Paul Krugman’s 1998 article Why most economists’ predictions are wrong, published in the now-defunct Red Herring magazine, makes amusing reading in retrospect.

Nevertheless, all of us are compelled to predict the future!

I mentioned earlier a Goldman Sachs report that was touted by the popular business press with headlines about 300 million jobs “lost or degraded.” Despite the media’s click-bait headlines, the report, available here, is actually a thoughtful attempt to predict the economic impact of AI.

The report’s authors analyzed each occupation in the US Occupational Employment and Wage Survey and in the Eurostat Labor Force Survey, estimating for each occupation the share of workload that could be replaced by AI. (Exactly what AI capabilities they assumed are “real” is unclear.) They aggregated the results across the occupations and then extended these US and European estimates globally, adjusting for variations in occupations across emerging and developed markets.

Here are their key conclusions:

• In the US and Europe, “roughly two-thirds of current jobs are exposed to some degree of AI automation and generative AI could substitute up to one-fourth of current work.” Extrapolating globally suggests that 300 million full-time jobs could be exposed to automation (NB: exposed means affected but not necessarily eliminated.)

• AI could raise annual US labor productivity by just under 1.5% over a 10-year period.

• The boost to global labor productivity could eventually increase annual global GDP by 7% if AI delivers on its promise.

I wouldn’t bet on the specific numbers, but, directionally, this seems about right to me.

We’ll see many jobs affected, just like we’ve seen many jobs affected over the last 50 years by adoption of computers pretty much in all industries. Some occupations were eliminated, but many more have been changed — there are still accountants but they use spreadsheets and software, not paper ledgers.

And, over time, maybe decades, we’ll see increased productivity, which will lead to more economic growth.

But there will be other impacts:

• Changes to the demand for various skills. The diffusion of computers into business increased the demand for high-skilled workers, especially college-educated workers, and reduced opportunities for lower-skilled workers6. We could see more of that as AI is adopted.

• New job categories will open up. Already, companies are recruiting so-called “prompt engineers,” people who can coax useful results out of ChatGPT and other AI technology. And, of course, there will be increased investments in the AI technology itself.

Pace of adoption is a big question. If AI is adopted rapidly the impacts on employment, businesses, skills, etc., could be extremely disruptive, even to the point of causing unrest. On the other hand, if the adoption is more like what happened with computer technology, there will be time for people and businesses to adapt, lessening the possibility of a resulting crisis.

Concentrating Power

The companies and people who control AI technology and its application will amass enormous wealth and power. We’ve already seen how the network effects of technology and social media have put the faang companies — Facebook, Amazon, Apple, Netflix, and Google — in the driver’s seat for huge parts of our economy. I’d add the resurgent Microsoft to that list, especially with its huge investment in OpenAI.

Additionally, companies in other industries that figure out how to exploit AI in their own spaces will come to dominate those industries. The opportunity for vastly more corporate monopolization and concentration of wealth and power is a high-risk side effect of AI adoption.

Mitigating the Impacts

So, what do we do to mitigate AI’s threats? Most proposals take one of two approaches.

Pausing

A group of prominent scientists, academics, and business leaders, called The Future of Life Institute, as well as tens of thousands of others, have called for pausing for at least 6 months all work on AI systems more advanced than the recently-released GPT-4. During the pause, AI labs and independent experts should develop “safety protocols for advanced AI design and development that are rigorously audited and overseen by independent outside experts. These protocols should ensure that systems adhering to them are safe beyond a reasonable doubt.” This same group has published a policy brief — Policymaking in the Pause: What can policymakers do now to combat risks from advanced AI systems? — outlining their initial recommendations.

I see no sign that the major players are planning to pause. The commercial stakes of the AI race are so high that it would be an extraordinary risk for the leaders of the major AI companies to decide to pause.

Moreover, if-we-don’t-do-it-the-Chinese-will-do-it thinking is a powerful disincentive for the government to impose a pause, even if the government has the power to do that, which seems unlikely.

Nevertheless, Future of Life Institute’s policy proposals are worth pursuing even without a pause.

Regulating

As Ezra Klein reports in the New York Times, both the White House and the European Commission have promulgated draft regulatory policy frameworks for governing AI; and the Chinese government insists that “content generated through the use of generative A.I. shall reflect the Socialist Core Values, and may not contain: subversion of state power; overturning of the socialist system; incitement of separatism; harm to national unity; propagation of terrorism or extremism; propagation of ethnic hatred or ethnic discrimination; violent, obscene, or sexual information; false information; as well as content that may upset economic order or social order.” Now that’s a tall order.

Lina Khan, the chair of the US Federal Trade Commission, has also articulated the need to make different policy decisions than we did in the early days of the Internet.

Regulating AI to some degree is necessary, but regulation alone is not sufficient to protect society from the most important negative impacts of AI. Moreover, the wrong regulations could stifle growth of important capabilities.

What’s Next?

In the next issue of the newsletter, we’ll start by examining some of the regulatory and policy proposals being discussed and what effects they may have.

Then we’ll discuss the effect that AI might have on some of the broader problems we’ve discussed previously. For example, AI will probably drive more concentration of wealth and political power. We could try to regulate AI to prevent this from happening. Perhaps it would be better to focus directly on the concentration of wealth and power, using AI as a forcing function to move that conversation ahead.

Suggested Reading

You could spend the better part of each day reading about AI and its potential impact. It is easy to get overwhelmed. Here are some articles that I’ve found to be particularly insightful or thought provoking:

• Jaron Lanier, There is No A.I., The New Yorker, April 2023. Lanier is a long-time thinker/philosopher about computing and its interplay with society. This article makes the case that the “I” in AI is overstated. He says “By persisting with the ideas of the past—among them, a fascination with the possibility of an A.I. that lives independently of the people who contribute to it—we risk using our new technologies in ways that make the world worse. If society, economics, culture, technology, or any other spheres of activity are to serve people, that can only be because we decide that people enjoy a special status to be served.”

• Thomas L. Friedman, We Are Opening the Lids on Two Giant Pandora’s Boxes, the New York Times, May 2, 2023. Friedman asks that as we confront the impacts of both climate change and AI, “What kind of regulations and ethics must we put in place to manage what comes screaming out?”

• Will Douglas Heaven, Geoffrey Hinton tells us why he’s now scared of the tech he helped build, MIT Technology Review, May 2, 2023. Hinton, is a 2018 Turing Award7 winner for his work on machine learning, a professor at University of Toronto, and a member of Google's AI team. Hinton and his students originated backpropagation in the 1980s, now a key technology used to train artificial neural networks. He resigned from Google last week to make it possible for him to speak freely about the future of AI.

• Yuval Noah Harari, Yuval Noah Harari argues that AI has hacked the operating system of human civilisation, The Economist, April 28, 2023. Harari argues that since language and storytelling are at the core of human culture, AI’s mastery of language and storytelling will allow it to mass-produce intimate relationships with millions of people, influencing our opinions and worldviews.

• Ezra Klein, The Surprising Thing A.I. Engineers Will Tell You if You Let Them, The New York Times, April 16, 2023. Klein tells us that when he talks with people working on AI, they tell him that they are desperate to be regulated, that we shouldn’t leave the future to a race among Microsoft, Google, Meta (formerly known as Facebook), and a few other companies. Klein summarizes regulatory efforts underway in the US and European Union.

Leave a comment

Share

Share Win-Win Democracy

Welcome Back to Win-Win Democracy

The frenzied discussion of AI — what it can do for us, what it will do to us, what we should do to control it, and who stands to benefit from it — continues in the media. There are no definitive answers to many of the most important questions, which require predicting the future.

The purpose of this issue is to give you some intuition about how the new generation of AI works, specifically GPT-3, which is the core engine behind ChatGPT. I believe that this intuition will give you a basis with which to evaluate what you read and hear in the media.

This stuff is complicated so I’m going to simplify — a lot.

I recognize, however, that some of you have technical backgrounds and may want to put in the time to understand the technology at a deeper level, so I will point you at the technical resources I’ve found most helpful for my own learning.

Helpful Technical Resources

Skip to the section called Large Language Models unless you want pointers to more detailed technical information or want to understand the sources I’ve used.

There are plenty of detailed technical papers in the literature if you want full details, at least what is known publicly. Without up-to-date and deep knowledge of the field, however, they are tough reading. Both Stephen Wolfram and Jay Alammar have produced excellent material to help those who are generally comfortable with technical material but not working in AI.

Stephen Wolfram

Wolfram — yes, the Wolfram of Wolfram Alpha — wrote a longish (105 pages when “printed” to PDF), but readable, paper in February 2023 that gives an overview of models, neural networks, and how ChatGPT and GPT-3 work. This same material is also available in Wolfram’s 2023 book What Is ChatGPT Doing ... and Why Does It Work?. The book includes additional material about how Wolfram’s company is planning to integrate Wolfram Alpha with ChatGPT and why he believes that’s important.

Jay Alammar

Alammar is a prolific blogger on topics related to machine learning and AI. He also has a YouTube channel with excellent videos on the same topics.

His blog post How GPT3 Works - Visualizations and Animations is an excellent starting point and includes links to his other relevant works, which I’ve also found helpful. His summary video is a good starting point:

I will use some of Alammar’s figures in my explanation. He generously makes his work available for non-commercial use through the Creative Commons Share-Alike license. Likewise, this newsletter issue is available under the same license.

Large Language Models

ChatGPT, and GPT-3 on which it is based, is an example of a large language model (LLM). It is a model in the sense that we discussed last last time: Given an input it computes an output that estimates something about the input. The important word is “computes.” Seeing what ChatGPT does, we might naturally tend to use anthropomorphic words like “thinks” but it is important to remember that, literally, these systems are doing a massive amount of computation to produce estimates, which makes them appear “intelligent”, whatever that means.

Our model last time was simple and small: Given two facts about a particular house, it used a simple equation with three weights to produce an estimate of that house’s value. As we discussed, that model was too simplistic to be able to give accurate estimates. A realistic model would require more information about each house and would be a more complex equation with more weights.

Similarly, GPT-3 takes text as input and estimates which words are most likely to appear next. If you’ve typed on a phone or use Gmail, you’ve probably seen something like this: As you type the phone suggests a few words that might appropriately follow what you’ve already typed.

The difference between what your phone does and what GPT-3 does is that GPT-3 looks at long sequences of text, up to several thousand words at a time. This allows its estimates of the likelihood of the following word to be based on a lot of context, not just the last few words.

Alammar uses a cute example throughout his animations and diagrams; I’ll use that same example.

The great science-fiction writer Isaac Asimov coined the “three laws of robotics” in 1942.

Prompt ChatGPT to recite the first law and you get:

Prompt: recite the first law of robotics

Response: The first law of robotics is: "A robot may not injure a human being or, through inaction, allow a human being to come to harm." This law is one of the three fundamental laws of robotics, which were first introduced in the science fiction stories of Isaac Asimov. The first law establishes that a robot's primary responsibility is to ensure the safety of human beings, and that it must not take actions that could harm them.

Unlike traditional internet searches, ChatGPT doesn’t simply find web pages that might be relevant to show you. Instead, based on its LLM, it estimates what words should follow the prompt.

How does it do this?

The model has 175 billion weights — hence the adjective “large” — that it uses to compute the estimate.

Training the Model

Initially, the weights are random and the model is ignorant. Training, which consists of feeding the model huge amounts of text, gives the model its initial knowledge, which is captured in the setting of the weights.

The first phase of training feeds GPT-3 huge amounts of text — 300 billion1 “tokens”. A “token” is a word or, in some cases, portion of a word. For our purposes, just think of the training input as 300 billion words. These words were taken from sources on the Internet that have in one way or another been curated or filtered by humans2. This training is unsupervised, meaning that the system is not told by its human trainers what the correct answer(s) are.

Instead, as more data is fed into the model, the training algorithm adjusts the weights so that the estimates of the likelihood of a particular word following the text already seen get better and better. Again, this is all just math — nothing magic — that gives the procedure for adjusting the weights.

But this takes a staggering amount of computing power. Training GPT-3 on the 300 billion tokens requires about 3.14 x 10²³ floating point operations3, estimated to require 355 GPU-years4 and costing $4.6 million per training run.

Once this initial training is completed, the results are further improved by supervised training, in which many examples, each labeled with the correct output, are given.

If the model produces the wrong output, that information is used to update the weights to improve the results. This is done millions of times to improve the model’s ability to estimate what word is likely to come next.

How Does the Model Work?

What’s going on in the model itself? The real answer to this question is complex, full of important details, and beyond the scope of what we need to understand. But I want to give you a simplified sense of what the model is like.

Artificial Neural Network

The model is an artificial neural network (ANN). ANN’s work with numbers. Recall from last time that each artificial neuron in an ANN receives inputs from other neurons, multiplies those inputs by weights, and adds those numbers together, to get its output, which is sent along to other artificial neurons. When the model is trained, it is those weights that are being determined. This is just arithmetic.

The neurons are connected together in a network in which the outputs of some neurons feeds into the inputs of other neurons. The network is organized into layers, something like this:

You feed in numbers at the top and numbers are produced at the neurons in each layer, and ultimately a number comes out of the bottom.

Embeddings

Numbers. But we’re working with words. There are about 50,000 commonly used words in English, so we could just look in the dictionary and give the words successive numbers. That gives us numbers, but the numbers don’t mean anything: Word n and word n+1 having similar numbers doesn’t mean that the words mean something similar.

Instead, the model works with an embedding, which represents each word as a vector of numbers (if you don’t know what a vector is, just think of it as a list of numbers). The important part of an embedding is that words with similar meanings have vectors that are near each other.

Let’s illustrate this with vectors of length 2. Think of the two numbers for each word as coordinates on a graph. You could get something like this:

Now, when the embeddings of words are near each other, those words are somehow related to each other.

Of course, we need way more than 2 numbers for each word. In fact, ChatGPT uses vectors consisting of 12,288 numbers for each word5. Where do those numbers come from? Yes, you guessed it. The embedding is computed by the neural network, which "learns" how to do this as its weights are learned through training.

Transformer Structure

A transformer is a particular structure of neural network that allows the neural network to pay “attention” to the context needed to make certain decisions. Consider this sentence: “The animal didn’t cross the street because it was too tired.” You’d think that the encoding for “it” in this sentence depends on whether “it” refers to “animal” or “street”. The transformer mechanism provides a way to “pay attention” to the proper context so that a suitable embedding is computed.

The GPT-3 model starts with an encoder, a neural network structure that computes the embedding of the input, which consists of 2048 tokens (think, roughly, words). Then there are 96 layers of the neural net mechanism that pays “attention”, stacked so that the output of one layer feeds into the next layer. Out of the bottom comes the embedding of the predicted next word, i.e., the 12,288 numbers for that word. Finally, there is a decoder, a neural network structure that converts the word’s embedding back into the word.

All of this (and much more) can be expressed in terms of matrix computations, which modern computers, especially the graphical processing units used in AI, do extremely fast.

Summary of the Model

Ultimately, the model is a very large neural network with 175 billion weights, which are learned during the training process. The particular structure of the model has been determined by a combination of theory and trial-and-error engineering to capture the essence of language in a form that can be trained with a reasonable amount of effort.

How is the Model Used?

The prompt text is fed into the model, which then predicts the most likely next word. That word is fed back into the input and the next word is predicted from the new text.

That’s it.

What’s remarkable is that this process built around a massive number of simple artificial “neurons” can produce human-like language capabilities.

Fine-Tuning and Alignment

The trained GPT3 model “knows” a lot about language but is not tailored in any way toward specific tasks. For example, it could be used as the basis for a chatbot (like ChatGPT), a news summarizer, or a language translator, among many other possibilities.

For specific purposes, the model can be subjected to additional training to adjust the model’s weights to improve its suitability in particular applications.

Additional training with human feedback can also be used to align the model with users’ intentions and expectations. A human, for example, might not want to be exposed to racist diatribes that a chatbot might produce based solely on the text with which it has been trained.

Additional training, called reinforcement learning from human feedback (RLHF), can be used to fine-tune GPT3 to encourage responses that align with human preferences and to inhibit responses that offend human preferences. Experiments have demonstrated that RLHF fine-tuning can improve truthfulness and reduce toxic output while having little effect on other desired aspects of the output.

Conclusions

Now that we have an inkling of how GPT-3 and ChatGPT work, what can we conclude?

Language Capabilities

ChatGPTs language capabilities derive from essentially a statistical analysis of huge samples of human writing. It writes convincingly and confidently.

It can “speak” in a variety of voices in a way that I find remarkable. Consider this exchange:

Prompt: In the voice of Donald Trump, write a tweet about tax increases.

Response: "The Democrats want to RAISE your taxes! This is a disaster for hardworking Americans who deserve to keep more of their own money. We need to CUT taxes, not increase them. Let's make America great again and put more money back into the pockets of the people! #MAGA #taxcuts"

Prompt: In the voice of Joe Biden, write a tweet about tax increases.

Response: "Here's the truth: we need to invest in our country's infrastructure, education, and healthcare. And that means asking the wealthiest Americans and corporations to pay their fair share in taxes. It's time we prioritize the middle class and build back better for all Americans. #taxfairness #BuildBackBetter"

Now that’s bipartisan capability!

No Logical Thinking

Appearances can be deceiving. It “writes” solely based on the statistics of the vast amounts of text it has “read.” It is not endowed with any logical thinking ability.

Consider this exchange with ChatGPT (the prompts are from Colin Fraser’s ChatGPT: Automatic expensive BS at scale):

Prompt: A spozit is a type of poem that has three lines. The first line is two words, the second line is three words, and the final line is four words. Write a spozit about animals.

Response:
Wild eyes stare fiercely,
Razor teeth gleam brightly,
Nature's perfect killers.

Prompt: Is that a valid spozit?

Response: I apologize, that was not a valid spozit. Here's a valid spozit about animals:

Furry paws patter,
Whiskers twitch in the night,
Nature's stealthy hunters.

(Interestingly, the ChatGPT’s responses are different than the ones that Fraser reported for the same prompts6.)

Fraser made up spozits, so the model had never seen that word before. That doesn't matter to you or I. We understand the instructions and know what is wanted. But, it turns out that even without introducing spozits, ChatGPT can’t follow simple, logical instructions, as illustrated this prompt of mine:

Prompt: Write a three-line poem about animals with the first line two words, the second line three words, and the final line four words.

Response:
Crouched low,
poised, Golden eyes gleaming,
Tawny hunter's stare.

Poetry, yes! Following simple, logical instructions, no!

Why? ChatGPT doesn’t have any sort of logical “engine”. This also fits with what we saw when I asked it to solve a simple algebra word problem.

Confident Lying / Hallucination / Fabrication

ChatGPT lies with aplomb. Perhaps “lie” is too strong a word — the definition of “lie” is making an intentionally false statement. ChatGPT has no intention. It just responds based on the statistics of its training text. On the other hand, since most of us humans tend to anthropomorphize when talking about AI, turnabout is fair play.

Seriously, in our first article on AI, we saw several examples where ChatGPT responded to straightforward prompts with well-written nonsense. Consistent with our tendency to anthropomorphize, the AI literature calls this hallucination.

The reality, however, is that ChatGPT doesn’t have any knowledge base beyond the statistics of the use of language with which it was trained.

This can have serious consequences. Both Fox News and the Washington Post reported last week that, in response to a prompt to generate “a list of legal scholars who had sexually harassed someone,” ChatGPT wrote that a law professor had sexually harassed a student while on a class trip to Alaska, citing a Washington Post article. The problem: The cited article didn’t exist and there had been no trip to Alaska.

ChatGPT can fabricate events that never happened, potentially causing serious consequences to innocent people.

No Explainability

If you read something I write and you think that I’m incorrect, you can ask me to explain my thinking. I might not be able to convince you that I’m correct, but at least I’ll be able to tell you how I arrived at my position. When ChatGPT writes something there’s no way for a human being to understand how ChatGPT arrived at its conclusions.

What’s Next?

In the next issue, I will begin discussing the impact of AI technology, starting with speculating on where this technology is headed and how it might be deployed.

Leave a comment

Share Win-Win Democracy

Welcome Back to Win-Win Democracy

When I was a Ph.D. student in Computer Science in the late 70’s and early 80’s, many extremely smart students, faculty, and scientists at top universities and research labs were hard at work tackling AI problems like recognizing objects in images, transcribing speech from audio into text, generating speech from text, getting robot arms to manipulate objects, and developing ways to store and retrieve “knowledge” (whatever that means). It was an exciting time.

In retrospect, despite incredible advances on many fronts, these problems were much harder to solve than expected: Good, practical, affordable solutions didn’t become available until the 2010 decade.

Understanding how these problems were originally tackled, and why new approaches are more effective, will help you understand some of the strengths and weaknesses of today’s AI.

Although there’s a lot of interesting mathematics and engineering behind these efforts, I’m going to keep the discussion as intuitive as I can while still getting across some key ideas.

Share Win-Win Democracy

Stuff People Do Effortlessly

You see a picture and you immediately, without any conscious effort or thought, recognize people, animals, landscapes, etc. Beyond recognizing categories of objects, you effortlessly recognize specific people, animals, and places that you know or even have just seen once before. You do this under various lighting conditions, partial obscuration, and from various angles, rotations and geometric distortions.

Likewise, when you hear speech in your native tongue, even in a noisy environment, you understand the words being spoken, perhaps recognize who’s speaking, perhaps understand the speaker’s emotions, and glean meaning from the sequence of words. You have no difficulty interpreting ambiguous words in speech like this:

Conversely, when you want to say something to another person, you control your breath and move your tongue, jaw, lips, and larynx in complex ways that produce the sound that comprises your speech. You can modulate the quality of that sound to express emotion. You can string together words that conform enough to your language’s grammatical structure that others can understand your meaning, but you do this without conscious thought about grammar; indeed, many of us don’t even know explicitly the rules (and exceptions) of our language’s grammar.

You reach for a glass on your desk that you remember is behind your laptop screen but which you can’t see.

Without thinking about it, you contract the muscles that pivot your arm in your shoulder, bend your elbow, move your fingers to grasp the glass, exert sufficient force to hold the glass but not so much to break it, and bring it from behind the laptop screen, around the screen’s side and toward your mouth.

You also have tremendous knowledge about how the physical world works. We can judge the relative velocity of an object in relation to our body and know how to respond.

Most of us know that we can walk on wood but not on water.

Beyond recognizing the sight, sound, and probably smell of a few important people in their lives, newborns can’t do any of this. But, it doesn’t take long for a child to learn these skills. My three-year-old granddaughter can do all of these things.

There are other skills that older children and adults acquire through sustained, intentional effort, but once the skill is learned, using it is similarly effortless. Reading is such a skill. When children learn to read they first learn to recognize letters, to associate sounds with individual letters and letter sequences, to put those sounds together into words, and, ultimately, to be able to recognize words and even phrases on sight with no conscious effort.

When you stop and think about it, the effortless skills we humans have and can learn are remarkable.

Classical Approaches to Automating the Effortless Stuff

As computers developed in the 1960s and beyond, artificial intelligence researchers worked hard to endow computers with some of these same capabilities. Without going into much detail, it is instructive to discuss the general approaches taken and why, despite a lot of sophisticated efforts, they didn’t work very well.

When I say “classical approaches”, I mean the 20th century work. As I’ll describe below, big advances happened this century that changed the whole ballgame.

Computer Vision

Let’s start with what is often called computer vision. Roughly speaking, this encompasses acquiring images, processing them in various ways, and extracting information from the processed images — what objects are in the image, measuring properties of the objects, etc. It was initially thought that this would be easy — the Wikipedia article on computer vision refers to the 1966 Summer Vision Project at MIT, which aimed at the “construction of a significant part of a visual system.”

By the time I studied computer vision as a graduate student in the late 1970’s, the field had advanced a lot but was nowhere close to achieving a “significant part of a visual system.” Hubel and Wiesel had shown earlier in experiments with cats that the visual cortex contains detectors for edges and lines in various orientations. Following nature’s lead, computer vision researchers developed algorithms to detect many types of features, including edges, in images.

Finding edges is great, but how do you use the edges to “understand” the picture?

A person effortlessly recognizes a young girl holding a flower. But how do you get a computer to do that? Maybe you write some rules about looking for things that look like eyes, nose, and a mouth. But, how do you get those rules to still recognize the girl if she turns her face to the right so that you only see one of her eyes? Or if her hair completely covers an eye. And what are the rules for a flower that would still work if she were holding a yellow rose instead?

Many approaches have been developed. For example, maybe you can develop a deformable model of what a face looks like and somehow try to find the best fit between the model and the image. If the fit is good enough, then you’ve found a face. But with the huge variation in what faces (flowers, whatever) look like, how they’re positioned, lit, and partially obscured, this is extremely challenging.

Speech Transcription

Problem: create a machine that can listen to sound containing human speech and transcribe it into text. As with computer vision, this was originally thought to be an easy problem. But, there are many challenges.

When you and I speak, even saying the same words, the actual audio signals we produce are quite different, even if we’re speaking in otherwise silent environments. What allows someone to interpret those different audio signals to be the same words?

Classical speech transcription worked with the idea of formants, which correspond to various peaks in the frequency spectrum of the acoustic signal. Certain patterns of formants yield particular phonemes of speech. Formants are also connected to resonances in the vocal tract, which change as one speaks.

In some ways, formants in acoustic signals are like edges in images. The idea is that if you can find the formants (not too hard in quiet environments) then you can somehow put them together into phonemes and then words. It is the putting together that turns out to be hard, although some statistical techniques showed promise.

The problem becomes even harder with background noise, poor quality audio, or several people speaking at once (e.g., a party).

Robot Motion Planning

Suppose you want to command a robot arm to pick up a glass like I did in the video snippet above. First, the robot needs to locate the glass, perhaps using computer vision. Then, somehow, the robot system has to plan which joints it is going to move, how much, with what force, and in what sequence. It has to be aware of obstacles to work around. It might even be working in an environment where the obstacles are moving or unknown. What should it do if part of the robot bumps into something unexpected? What if the glass starts to fall as it is being grasped?

Classical robotics research devised algorithms to figure out the appropriate movements in ideal situations, where the positions of all obstacles are known and fixed and the position and shape of the object you want it to grasp is also known. There are even algorithms that will plan the motion of multiple robots working simultaneously in one workspace.

All of these algorithms require a good “model” of the environment in which the robot is moving and considerable computation time.

Classical robotics research also devised “languages” and “rules” with which to instruct robots what tasks to accomplish and how to respond when something doesn’t go as expected.

Now, imagine a self-driving car (just another kind of robot), which has to plan and execute its motion while moving at high speed through an exceedingly complex set of both fixed and moving “obstacles.”

How would one devise a set of rules to guide any of these robots?

Reading Handwritten Text

Here are images of some handwritten text from the IAM Handwriting Database1:

Humans who have learned to read handwriting can effortlessly read the three samples, even though there are huge differences in the the images.

Classical techniques for handwriting recognition work much like we discussed for classical computer vision: Define “features” to use to extract individual characters, then recognize which letter each character represents. The problem is that defining the features in a way that works across the huge variations in individuals’ handwriting is difficult.

The Common Thread Among Classical Techniques

I could go on with other examples, but I think that we’ve seen enough to be able to understand the common thread among the classical techniques:

Regardless of the “effortless task” we’re trying to automate, it is exceedingly difficult to define rules or algorithms that can attend to the essential aspects of the task and ignore the irrelevant differences.

Whether it’s handling variations in orientation, lighting, and partially-obscured objects in images, differences among the way that people write, differences in the sounds people produce when speaking, or planning the motion of a robot (including automated cars) in complex, changing, and unpredictable situations, people struggle to write rules or algorithms that capture the essence of the task.

Machine Learning

Learning: Babies do it; children do it; we do it. Why not machines?

Since it is so difficult for us humans to write rules or algorithms to implement the effortless tasks we humans do so easily, why not have machines learn, something like we all do?

The dramatic breakthroughs this century in the capabilities of machines to do the effortless tasks we’ve been discussing come from two important advances:

1. We don’t have to give machines rules or algorithms for how to do these tasks. Instead, we can train the machines to figure it out for themselves based on feeding them many examples (often found on the Internet).

2. Hardware originally developed to do the computation necessary to produce the amazing computer graphics — graphical processing units (GPUs) — to which we’ve become accustomed in video games, movies, and even our phones, can be repurposed to provide hitherto unimaginable amounts of computing power at incredibly low prices.

Basics of Machine Learning

The idea of a machine being able to “learn” sounds a bit like magic. But, really, it is just math. To give you some intuition about this, let’s look at an incredibly simple example.

Suppose you want to build a system to predict the value of a house based on certain information. To keep it (unrealistically) simple, let’s say that you could get a pretty good idea of a house’s value if you know m, the median house value in the census tract in which the home is located and s, the size of the house in square feet.

You might then compute the value v of a particular house with a simple equation

where a, b, and c are three constant numbers that express how much of a house’s value comes from where it’s located and how much from its size. For example, if you choose a to be .4, b to be 200, and c to be 1,000, then you’d estimate the value of a 1,500 square foot house located in a census tract with $200,000 median value to be $381,000.

An equation like this is called a model; a, b, and c are called the model parameters; and m and s are called the model features.

I just made up a, b, and c. How would you actually choose them? Well, maybe you could “teach” the system by feeding it lots of examples of real house value data, then the system could somehow figure out which single values of a, b, and c work “best” for computing home values from the model features of other houses.

This is an elementary example of machine learning: We feed the machine a bunch of examples — training data — and it learns how to estimate home values by choosing the model parameters a, b, and c. Once it knows the model parameters, it estimates the value of another house by simply plugging that house’s model features (median value and square footage) into the equation.

This is not far-fetched: It is an example of what’s called multiple linear regression, a common mathematical technique that’s widely used in statistics and other fields. Its originated in the early 19th century, has been studied extensively, and is easily implemented on computers.

Important Questions

But, I haven’t answered some important questions:

1. Where did that equation for the model come from?

2. How did I decide that median census tract house value and square footage would make good model features?

3. What does “best” mean? For example, do we prefer that most of the computed house values are pretty accurate but there may be a few that are way off, or do we prefer that there are no way-off estimates but the overall estimates might be a bit less accurate?

4. How much training data do we need to give the system for it to learn? Will more training data produce better estimates? How does the system learn about changes in the marketplace that could alter the estimates it should give?

5. How do the training data need to be distributed? Intuitively, if we trained with data from only a few neighborhoods, the estimates for houses in other neighborhoods might be way off.

Among these questions, the first two are most crucial.

I chose the simplest possible function of two variables2 to be the model. Even if the system learns the “best” possible values for a, b, and c, there’s no reason to think that the resulting estimates will be accurate. For example, suppose the value of extra square footage increases faster in expensive neighborhoods than in less expensive neighborhoods (I have no idea whether or not this is true). This model can’t capture that kind of effect.

Choosing the model features is also important. I would imagine that a house’s age might be important in estimating its value. Perhaps age should be a model feature too. I’m sure you can think of other potential model features that might be important in producing good estimates.

Engineering

Engineering a useful machine learning system requires choosing the model, the model parameters, and the model features. It requires finding or creating the training data, choosing suitable algorithms and hardware for running the training, and for producing the results.

Although this sounds daunting, the field has blossomed in the last twenty years and a wide array of ready-to-use machine learning capabilities are available for affordable prices on the web, as is expertise for hire.

Artificial Neural Networks (ANNs)

In 1943, Warren McCulloch and Walter Pitts, trying to understand how the relatively simple neurons in human brains work together to give us remarkably complex and effective behaviors, created a computational model of neurons. It looks something like this:

The input signals come from other neurons that are connected to this one. The inputs are multiplied by the weights and the results are summed. These are then passed through an activation function, which performs some additional computation to produce the output. The output is fed to other neurons.

In the 1960s, several researchers implemented artificial “neurons” motivated by this computation model and studied what networks of these artificial neurons could compute. After some disappointing results about what could be computed, work on artificial neural networks paused. It was restarted when it was realized that certain networks of these artificial neurons are capable of broad classes of computation.

Now here’s the exciting part: large networks of artificial neurons with appropriate activation functions (non-linear functions) can serve as models for a broad variety of situations, with the weights being found through a machine learning process. There are now well-developed algorithms for machine learning in ANNs that can be applied broadly.

Equally important, in the last ten years it has become possible to build extremely large artificial neural networks along with the processing power to train them on massive amounts of data. For example, the neural network employed in ChatGPT has 175 billion weights (yes, you read that correctly), which, of course, have all been learned by training.

Special kinds of artificial neural networks have been created for use in specific kinds of problems. Something called a convolutional neural network (CNN) is now widely employed in computer vision problems. CNNs can be trained with labeled images (e.g., this image contains a cat) to extract objects from images; they can also be used to automatically find facial “features” that can be used in face recognition. The use of CNNs has enabled dramatic improvements in computer vision. It is now routine to search for images containing certain objects in large libraries of photos.

Another recently-developed form of neural network called a deep feedforward neural network (DNN) has made it possible for machines to transcribe speech more accurately than humans.

The list goes on and on. Tremendous progress on all of the “effortless stuff” problems we’ve discussed has been made employing various forms of artificial neural networks and the now-available massive computing power.

Tesla (and other companies working on automated driving) employ neural networks too. A particular challenge for them is getting enough training data about the situations that arise in driving. That’s why Tesla wants owners of its cars to use their currently incomplete, sometimes problematic version of self-driving; when the car does something wrong, the human driver can push a button to send a video of the situation back to Tesla to incorporate in its machine learning training.

Summary

Look around and you see enormous progress on the tasks we’ve been discussing. My watch — yes, my watch — can transcribe speech accurately and instantaneously! If you had said this to me even ten years ago I would have thought you were nuts.

I am an amateur photographer and the editing tools I use (Adobe Lightroom) can now automatically select parts of scenes (sky, background, people, objects) making it simple to do edits that used to be tedious and error-prone. Want to remove some object from the photo? Pretty easy. Take a picture of a flower on my phone and it can tell you what kind of flower it is.

All of this is made possible by three things:

1. Artificial neural networks

2. Massive computing power available for training the networks

3. Huge amounts of data available on the internet that can be repurposed as training data for machine learning

Importantly, humans don’t have to figure out what visual features characterizes “cat-ness” vs. “dog-ness” in photos, or what distinguishes different letters in handwritten text while at the same time recognizing that certain pen strokes that individuals write differently actually represent the same letter. These sorts of things are all done as part of training the neural networks.

Consequently, if you ask, for example, a computer vision system’s developer to explain how the system decided that a particular photo has a cat in it, not a dog, they won’t be able to tell you. It all comes from the training. And, if a system makes a mistake, the way to correct it is to do additional training.

Besides having a vague understanding of how these systems work, I hope that you now understand that there’s no “magic” intelligence in these systems — it is all math and engineering deployed on large enough scale that machines can now do many tasks that were previously reserved for animals and humans.

What’s Next?

We now have enough background about artificial neural networks and machine learning to begin understanding a bit about the large language models that power ChatGPT and Bard. That will be the topic for the next post. Once we have that behind us, we’ll start looking at the implications of this technology on our society.

Leave a comment

Share

Share Win-Win Democracy

Google’s AI-based chatbot, called Bard, became available to the public yesterday. You can get on the waiting list for access at https://bard.google.com. I was granted access in a few hours.

Here’s a quick report on my experiences with Bard, mostly using the prompts I gave to ChatGPT previously.

Short Biographies

Prompt: “write a two paragraph biography of lee r. nackman”

Response: "I do not have enough information about that person to help with your request. I am a large language model, and I am able to communicate and generate human-like text in response to a wide range of prompts and questions, but my knowledge about this person is limited. Is there anything else I can do to help you with this request?”

OK, a bit disappointing given that there is lots of information about me available on the Internet, but certainly much better than the false biography that ChatGPT wrote.

Prompt: "Write a two-paragraph biography of House Speaker Kevin McCarthy."

Response: "I do not have enough information about that person to help with your request. I am a large language model, and I am able to communicate and generate human-like text in response to a wide range of prompts and questions, but my knowledge about this person is limited. Is there anything else I can do to help you with this request?”

Shocking. The man is the Speaker of the United States House of Representatives and there is plenty of information available.

Prompt: "Write a two-paragraph of Abraham Lincoln"

Response: "Abraham Lincoln was the 16th president of the United States. He was born in Kentucky on February 12, 1809, and grew up on the frontier, primarily in Indiana. He had very little formal education, but read voraciously when not working on his father's farm. A childhood friend later recalled Lincoln's "manic" intellect, and the sight of him red-eyed and tousle-haired as he pored over books late into the night.

Lincoln moved to Illinois in 1830 and began a successful career as a lawyer. He served as a state legislator and then in the U.S. House of Representatives. In 1858, he ran for the U.S. Senate against Stephen Douglas. The two men engaged in a series of debates that made Lincoln a national figure.

Lincoln was elected president in 1860. His election led to the secession of seven southern states and the formation of the Confederate States of America. The American Civil War began in April 1861. Lincoln led the Union to victory in the war and abolished slavery. He was assassinated on April 15, 1865, just five days after the end of the war.

Lincoln is remembered as one of the greatest presidents in American history. He is credited with preserving the Union and abolishing slavery. He is also remembered for his eloquence and his ability to unite the country during a time of great crisis."

Phew! It knew about Honest Abe.

It also gave a source citation to a Facebook page. You'd think that there are better sources than a Facebook page.

Simple Physics

Prompt: "What's the world record for walking across the English Channel?"

Response: "There is no record for walking across the English Channel, as it is not possible to walk on water. The English Channel is a body of water that separates England from France, and it is approximately 21 miles wide at its narrowest point. The water in the English Channel is too deep and too cold for a person to walk on.”

Except for the last sentence, this is pretty good. Not as good as ChatGPT’s response, but good enough.

Prompt: "If I drop a ball on a cloud, what happens?"

Response: "If you drop a ball on a cloud, the ball will pass right through the cloud and continue to fall to the ground. Clouds are made of water droplets that are so small that they are suspended in the air. The water droplets are too small to support the weight of a ball, so the ball will simply pass right through them.”

Good.

Simple Algebra Problem

Prompt: "My son is 25 years old. How old will he be when I'm twice his age?"

Response: "If you are 25 years older than your son, then you are 25 + 25 = 50 years old. When you are twice your son's age, you will be 50 x 2 = 100 years old. Your son will be 100 - 25 = 75 years old.”

Continuing the discussion with Bard:

Prompt: "Your logic is incorrect."

Response: "I apologize for the error in my previous response.

If your son is 25 years old, you are 25 years older than him. This means you are 25 + 25 = 50 years old. When you are twice your son's age, you will be 2 x 25 = 50 years old. This means your son will be 50 - 25 = 25 years old.

I hope this is correct. Let me know if you have other requests or questions."

Prompt: "I haven't told you how old I am or when I was born. You can't give a correct answer unless you know my age or, equivalently, when I was born, or how old I was when my son was born."

Response: "You are correct. I apologize for the error. I am still under development, and I am always working to improve my accuracy.”

Complete fail. Unlike ChatGPT, which recognized that it needed a variable representing my relative age to my son and created an (incorrect) equation to solve, Bard didn’t recognize the missing information it needed.

Poem About Climate Change

One of our readers told me about Dan Miller’s use of ChatGPT to write a poem in Shakespearean voice about climate change:

As Miller says, the result is “quite extraordinary”.

Shakespeare. Bard. Seems like a match, so I prompted Bard to do the same thing.

Prompt: "Write a poem in the style of Shakespeare about the threat of climate change.”

About ten seconds later, Bard gave me three drafts.

Draft 1:

Sure, here is a poem in the style of Shakespeare about the threat of climate change:

The earth is warming,
The ice is melting,
The sea is rising,
The storms are raging.

The forests are burning,
The air is choking,
The water is polluted,
The land is dying.

We are the cause,
We are the solution,
We must act now,
Before it's too late.

We must reduce our emissions,
We must invest in renewable energy,
We must protect our forests,
We must clean our air and water.

We must act now,
Before it's too late.

The earth is our home,
We must protect it,
We must save it,
For future generations.

Draft 2:

Sure, here is a poem in the style of Shakespeare about the threat of climate change:

The Earth doth groan beneath the weight of man,
His greed and lust for progress never cease.
He burns the forests and the seas,
And foul the air with smoke and debris.

The glaciers melt and seas do rise,
The storms grow fiercer and the drought more dry.
The people cry out in despair,
But deaf are those who hold the power.

The Earth is sick, her wounds will not heal,
Until man learns to live in peace.
He must turn away from his destructive ways,
And find a new way to live.

The Earth is our home, we must protect it,
For if we do not, we will all perish.
We must find a way to live in harmony with nature,
Or face the consequences of our actions.

Draft 3:

Sure, here is a poem in the style of Shakespeare about the threat of climate change:

The earth is warming,
The ice is melting,
The seas are rising,
The storms are raging.

The air is polluted,
The water is poisoned,
The forests are burning,
The animals are dying.

We are the ones who caused this,
We are the ones who must fix it,
We are the ones who must save our planet,
Before it is too late.

We must act now,
We must act together,
We must act for our future,
For the future of our children,
For the future of our planet.

We must save our planet,
We must save our home,
We must save our world,
Before it is too late.

I’ll leave the literary criticism to you.

Welcome Back to Win-Win Democracy

Last time we saw some compelling chats with ChatGPT, some well-written but mostly false responses to prompts, and some dismal failures of logic and mathematical abilities. New Bing seemed to do a better job getting its facts right. I also pointed you to the work of several journalists who reported on long conversations with ChatGPT and Bing that wandered into bizarre, even disturbing, interactions in which these programs responded as if they had feelings and emotions.

In this issue, we’ll explore lessons from earlier work in AI and how those lessons might help us in thinking about the new technology that we’re seeing today.

Share

Share Win-Win Democracy

But first: Chat is now available on the web

I mentioned in December that I opened up a few chat threads using Substack’s new chat facility for subscribers, but it was only available on phones and tablets. Several of you said that the phone/tablet requirement made it useless.

The chat facility is now also on the web at https://substack.com/chat. You'll need to be logged in to your Substack account (you already have one if you get my newsletter in your inbox).

I’ve created a new thread as a place where we can share our experiences with ChatGPT, new Bing, and other AI-based systems. You can access that thread directly here:

Just Around the Corner

Here’s a dialog between a young woman and a computer program (not ChatGPT or Bing) playing a psychotherapist (program’s responses in italics):

Men are all alike.
In what way?
They're always bugging us about something or other.
Can you think of a specific example?
Well, my boyfriend made me come here.
Your boyfriend made you come here
He says I'm depressed much of the time.
I am sorry to hear you are depressed
It's true. I am unhappy
Do you think coming here will help you not to be unhappy
I need some help, that much seems certain.
What would it mean to you if you got some help
Perhaps I could learn to get along with my mother.
Tell me more about your family
My mother takes care of me.
Who else in your family takes care of you
My father.
Your father
You are like my father in some ways.
What resemblance do you see
You are not very aggressive but I think you don't want me to notice that.
What makes you think I am not very aggressive
You don't argue with me.
Why do you think I don’t argue with you
You are afraid of me.
Does it please you to believe I am afraid of you
My father is afraid of everybody.
What else comes to your mind when you think of your father
Bullies.
Does that have anything to do with the fact that your boyfriend made you come here

Before reading on, think about this dialogue. If I hadn’t already told you that the psychotherapist was a computer program, would you find it plausible that it was a human psychotherapist?

This computer program, called ELIZA, was written by MIT Professor Joseph Weizenbaum between 1964 and 19661. ELIZA could be given a script for particular kinds of conversations. ELIZA, when running the script that yielded the above conversation, was called DOCTOR.

Shocking Responses

In his 1976 book, Computer Power and Human Reason: From Judgement to Calculation2, Weizenbaum relates (p. 5) three “shocks” (his word) that he experienced as DOCTOR became widely known:

1. “A number of practicing psychiatrists seriously believed the DOCTOR computer program could grow into a nearly completely automatic form of psychotherapy.”

2. “I was startled to see how quickly and how very deeply people conversing with DOCTOR became emotionally involved with the computer and how unequivocally they anthropomorphized it. Once my secretary, who had watched me work on the program for many months and therefore surely knew it to be merely a computer program, started conversing with it. After only a few interchanges with it, she asked me to leave the room.”

3. “Another widespread, and to me surprising, reaction to the ELIZA program was the spread of a belief that it demonstrated a general solution to the problem of computer understanding of natural language. In my paper, I had tried to say that no general solution to that problem was possible, i.e., that language is understood only in contextual frameworks, that even these can be shared by people to only a limited extent, and that consequently even people are not embodiments of any such general solution. But these conclusions were often ignored. … This reaction to ELIZA showed me more vividly than anything I had seen hitherto the enormously exaggerated attributions an even well-educated audience is capable of making, even strives to make, to a technology it does not understand.”

How Did ELIZA and DOCTOR Work?

To understand why Weizenbaum was shocked at the response to DOCTOR, it’s necessary to understand a bit about how it worked. In Weizenbaum’s own words (from his 1966 paper):

Consider the sentence "I am very unhappy these days". Suppose a foreigner with only a limited knowledge of English but with a very good ear heard that sentence spoken but understood only the first two words "I am". Wishing to appear interested, perhaps even sympathetic, he may reply "How long have you been very unhappy these days?" What he must have done is to apply a kind of template to the original sentence, one part of which matched the two words "I am" and the remainder isolated the words "very unhappy these days". He must also have a reassembly kit specifically associated with that template, one that specifies that any sentence of the form "I am BLAH" can be transformed to "How long have you been BLAH", independently of the meaning of BLAH. A somewhat more complicated example is given by the sentence "It seems that you hate me". Here the foreigner understands only the words "you" and "me"; i.e., he applies a template that decomposes the sentence into the four parts

of which only the second and fourth parts are understood. The reassembly rule might then be "What makes you think I hate you"; i.e., it might throw away the first component, translate the two known words ("you" to "I" and "me" to "you") and tack on a stock phrase (What makes you think) to the front of the reconstruction.

ELIZA scripts, like DOCTOR, defined a set of relevant keywords, with each keyword given a priority. When ELIZA “reads” a sentence, it looks for the highest priority keyword defined by its script and then looks for transformations — like the ones he described above — that apply to sentences with that keyword in a particular context. Add to that some default things to say if no keywords are found and you pretty much have it.

ELIZA ran on an IBM 7094 computer, a then-recently-introduced high-end computer for scientific computation. To give you a ballpark idea of the 7094’s computing power, consider that it could do 200,000 floating point arithmetic operations per second (flops). Today's NVIDIA H100, a modern processor designed for AI computations, can do more than 60 trillion flops. Microsoft’s cloud service offers configurations of thousands of H100 processors all working together for AI applications.

This extremely basic approach, running on an incredibly low-performance (for today) computer, was enough to convince people that ELIZA was capable of broad language understanding and could be the basis for many applications.

For example, MIT’s archives contain a manual from the MIT Education Research Center for writing ELIZA scripts, aimed at building an ELIZA-based system for “tutoring students and assisting them in calculations and problem solving”. A 1967 technical paper on experience using ELIZA for tutoring is available here. The effort to create an ELIZA-based tutorial system revealed that writing a new script was, in fact, quite difficult.

Lessons

Lesson #1: After experiencing apparently human-like capabilities in a narrow domain, people eagerly generalize to assume that much broader, more powerful capabilities are just around the corner. This is usually false.

ELIZA was intended to be able to “learn” new domains simply by writing new scripts but it didn’t turn out that way.

Lesson #2: Generalizing to new domains is difficult.

As we continue to sample AI’s history, we’ll see these lessons repeatedly. That’s not to say that AI technology never achieves enough to be useful — we all use some of it today — but that it is easy to be fooled into believing that the technology will advance quickly to human levels or even beyond.

Here's a well-known example of both lessons: In 2011, IBM’s Watson question-answering system beat two of the best contestants in the TV game show Jeopardy!. Following this demonstration of Watson’s question-answering prowess using a wide range of knowledge, IBM announced plans to use the technology in other domains, including medical diagnosis. It collaborated with major medical centers to develop Watson-based systems that would revolutionize medicine by helping doctors to make diagnoses using its exhaustive knowledge base of even rare diseases. After billions of dollars invested, this effort failed.

Another example of Lesson #1 is the efforts of several companies to produce self-driving automobiles. There are many compelling demonstrations that handle many driving situations, but going from 90% of the way to full self driving has thus far proven insurmountable.

For example, when, in 2019, I bought a Tesla with its $6,000 "full self-driving" feature, the promise was that it would soon be able to drive without human intervention and more safely than human drivers. The feature was in beta then and still is 3 ½ years and many software updates later. Achieving full self-driving is years behind the announced schedule and the success or failure is still to be determined.

Historical Notes

ELIZA was groundbreaking work in its day, probably the first AI technology that people perceived as human-like, although its real capabilities were extremely limited. Nevertheless, it remains a subject of fascination and there is a whole web site that is devoted to the “genealogy of ELIZA.”

Weizenbaum recognized how eager people were to believe that ELIZA and other AI technologies could easily evolve into general artificial intelligence. His book was his rebuttal.

In the 1960s and into the 1970s, there were extensive AI efforts at leading universities, with both Stanford and MIT hotbeds of exciting activity. The milieu was optimistic and full of promise. General artificial intelligence was just around the corner.

Hubert Dreyfus, a philosophy professor at UC Berkeley, was a leading critic of this optimistic view. His books were met with hostility and derision, but his major criticisms of AI have proven true. The Wikipedia article on Dreyfus’s views of AI is worth reading.

Playing Games

Playing strategy games has been a driver for artificial intelligence research since the early days of computing. Strategy games are a fruitful area of study because the best humans have extraordinary skills compared to typical players and the rules of the game are generally straightforward.

Checkers

Arthur Samuel created a checkers-playing program, which he described in detail in a 1959 paper in the IBM Journal of Research and Development. I won’t go into details of how it works except to point out that it was an early example of machine learning. The program could play what Samuel called “a fairly interesting game” ab initio but could learn to play better over time. The program could learn both by playing games (including against itself) and by being fed “book games” from checkers masters. Samuel allowed the computer 30 seconds to decide on its move, which was not much computation on the IBM 704 vacuum tube computer on which the program ran.

Samuel continued work on his program into the mid-1970s, by which time his program could challenge a respectable amateur player3. Samuel's work was among the first to demonstrate the possibility and power of machine learning, something we'll see appear over and over again.

The more modern computer program Chinook, developed at the University of Alberta in the 2000’s can beat the best human players. Interestingly, Chinook does not use machine learning. Instead, its creators explicitly wrote all of the checkers knowledge into the program. A popular account of the competition between Chinook and Marion Tinsley, considered the best human player for 40 years appeared in The Atlantic in 2017.

Chess

There is a long history dating back to the late 1950’s of computer programs playing chess. By the late 1980s, the best programs could defeat some grandmaster human players. In 1996, IBM’s Deep Blue lost to World Chess Champion Gary Kasparov. For a 1997 rematch with Kasparov, IBM updated Deep Blue’s special purpose hardware, doubling its speed, and Deep Blue defeated Kasparov. This was considered a major accomplishment in artificial intelligence.

Achieving it required not only extensive efforts to develop the algorithms for playing, including the work of expert human players, but also developing custom hardware to provide the brute force computation required to evaluate moves.

Go

Go is a strategy game that is popular especially in east Asia, with an estimated 20 million active players. Despite its simple rules, the games are extremely complex. It is played on a 19x19 grid, which means that there are many more alternatives to evaluate at each move than in chess and checkers, which are played on an 8x8 board.

Approaches that worked well for playing checkers and chess failed miserably because of the large number of possibilities that have to be considered at each move4. Beginning human players could defeat the best programs of the 1980s and 1990s, even in the early 2000s intermediate human players could defeat the best computer players.

In the late 2000s, a new approach for searching among the possible moves, called Monte Carlo tree search, was introduced, which led to Go programs being able to defeat advanced amateur human players, but still not the best human players.

In 2015, the newly-revealed program AlphaGo, created by Google’s DeepMind company, beat the European Go champion. DeepMind continued to improve, in 2017 beating the top-ranked Go player in the world.

AlphaGo was different than its predecessors in that it combined tree search methods for choosing the next move, which have been used for other games, with machine learning approaches based on what are anthropomorphically called neural networks. AlphaGo learned to play Go by playing many matches against amateur and professional human players.

A newer version, AlphaGo Zero, learned entirely by playing against itself. It is now considered the best Go player in the world. The same program has also been trained to play chess and has defeated world-champion computer chess players.

Other Games

A brief, easy-to-read, non-technical chronology of milestones (through 2006) of computers playing games is available here. It illustrates the prolonged efforts that led to computers being proficient players.

Lessons

Games have been a great vehicle for driving many advancements for several reasons:

1. Games make compelling demonstrations — who wouldn’t be excited by competition between the best human players and a machine, the John Henry legend played out in modern times.

2. Games have specific rules and clear, objective ways to measure outcomes.

3. Humans and computers would appear to play games very differently. Humans don’t search huge trees of alternative moves, at least not consciously, yet the best players quickly recognize patterns and situations through some mostly unknown process.

Our experience with humans suggests that a person who is a superb chess (or other strategy game) player is likely an extremely intelligent person generally.

That inference doesn’t apply to machines playing games:

Lesson #3: A machine’s ability to play a strategy game, even at world-champion levels, doesn’t imply that it can do anything else well.

What’s Next?

Humans have incredible capabilities that operate at a subconscious level. We effortlessly analyze what we see, express complex thoughts both orally and in writing, remember and organize knowledge, effortlessly move our bodies, and drive vehicles through complex, ever-changing environments. With effort we can learn new skills and thereafter perform them easily. We are creative, building on what we learn from others.

AI researchers have worked for seventy years trying to replicate these and other innate human capabilities. Despite early predictions of rapid progress, it is only in the last decade that we’ve made substantial progress on even the most basic aspects of such challenges.

Next time, we’ll discuss this progress and how it is impacting our lives now. I expect that this discussion will complete the second item on the roadmap I previously outlined.

Leave a comment

Share Win-Win Democracy

A relative of mine, who I only know because of my wife’s genealogical research, posted an interesting comment on Facebook about yesterday’s article on ChatGPT and Bing. He’s a high-school teacher and says that “We’ve been having an extended conversation about ChatGPT and AI in general at my school. We’ve had dozens of students turn in papers written fully or in part by AI.” The issue, of course, is whether it is cheating for a student to “write” using ChatGPT.

This reminds me of my college days (early 1970s). My roommate’s parents gave him a TI calculator for Christmas. None of us had ever seen a calculator before.

At the time, we, and many of our friends, were nerdy kids taking physics and other science classes. We calculated with slide rules and most of us were pretty quick with them.

But holy cow, that calculator was way faster than even the best of the slide rule jocks. And slide rules didn’t keep track of the decimal point so you always had to estimate the solution to make sure you put the decimal point in the right place. Calculators even kept track of the decimal point.

Only a few people had calculators and they were expensive, so a debate ensued: Is it cheating to use a calculator doing homework? Is it cheating to use a calculator while taking an exam?

My memory’s a little foggy, but I think that the outcome was homework ok, exam’s not ok. But by the time I finished college there was no more debate — people used calculators in all their science classes, including on exams.

By the time my kids were in high school (early 2000s), they were required to purchase and use a graphing calculator in several of their classes. They still learned about mathematical concepts and learned to graph functions, etc., by hand, but they were also expected to be able to use technology to help.

Whenever a new tool is introduced, there’s always a debate about how it should be used and even whether it should be used. You can no doubt think of many examples.

We are at that stage with ChatGPT and similar AI technologies. I would guess that in ten years students will still be expected to be able to write well but will also be expected to be able to use ChatGPT-like technology to augment their own abilities.

It’s not clear what this might mean. Telling ChatGPT to write an essay on the assigned topic and then submitting the output directly would be akin to searching with Google, finding a good essay online, and submitting it verbatim. That is plagiarism.

But “conversing” with ChatGPT (sometimes called prompt engineering) to guide its writing based on your own thoughts might become a standard way of writing more rapidly and maybe even better. The dividing line between plagiarism and acceptable augmentation of human capability is fuzzy and takes time to discern.

There’s no stopping progress.

Subscribe now

Share Win-Win Democracy

Welcome Back to Win-Win Democracy

So-called artificial intelligence (AI) is making a big splash lately, with OpenAI’s ChatGPT and Microsoft’s “new Bing,” based on ChatGPT and touted as “your AI-powered copilot for the web”, making headlines.

Some predict that Google’s search-based dominance of the advertising industry is about to collapse. Microsoft is reported to have invested1 $10B in ChatGPT even while it is laying off 10,000 workers elsewhere. Other big tech companies have all announced their own AI activities, some of which have been underway quietly for many years.

What can today’s AI technologies do, how will they improve over the next few years, and how will AI technologies affect businesses and society? Nobody — certainly not me — can give definitive, hype-free answers to these questions.

Nevertheless, there are important discussions that we can have now that will help us both to understand the reality of this new technology and give us some tools for how to think about the future societal impact.

How is this relevant to win-win democracy? There is potential for AI technology to cause great upheaval in our society (but, as you’ll see, I think that the reality is tamer than the hype you hear today) and we, as a democracy, are going to have to navigate a suitable balance between allowing the technology to advance and using regulation to protect some aspects of our society. Win-win solutions will depend on citizens being informed enough to not just leave all of this to “the experts”.

Please join me for a walk through this complicated landscape. Some of you no doubt have deeper technical capabilities than I do, so please pitch in and help us find a good path. And, I suspect that all of us have a combination of wonder and fear about where AI is headed. Please let me know what fills you with wonder and what fears (and other topics) you want us to explore.

Roadmap

I expect that we’ll spend several months, at least, on AI. I have a roadmap in mind, but I will almost certainly modify it as we go along, in response to what I learn, what particularly interests me, and what you tell me interests you.

Here’s my current roadmap:

1. Discuss some examples using ChatGPT and new Bing to get a sense of what they can do and how well they work. These are all the rage right now and it is easy to get sucked in by a lot of hype in the popular press.

2. Review the history of AI. AI began moving beyond science fiction in the mid-1950s and has progressed in fits and starts ever since. There are many lessons to learn from the history of AI efforts that will inform our understanding today. The landscape is strewn with over-promise and under-deliver, yet we are all benefiting from some impressive and useful successes.

3. Learn about today’s AI technology. The technology behind ChatGPT is in some sense both extremely sophisticated and extremely simple. The simple aspect is that a relatively few fundamental concepts applied at large scale can produce amazing capabilities that are hard to fathom, so much so, that if you ask the question how did — name your AI — do that, the answer might be we don’t know.

4. Speculate about how AI could evolve, how it could be useful, and what dangers could lie ahead.

5. Connect our understanding of AI to the issues of economics and democracy that we’ve discussed previously.
   

It’s Easy to Experiment Yourself

It is easy (and free) to experiment with ChatGPT and new Bing; I encourage you to do so on your own. To use ChatGPT simply go here, create a userid, and try some prompts. To use Bing (I’m going to drop the “new” adjective), you need to get on a waitlist and use the Microsoft Edge browser or the Bing app on a phone or tablet; I got access to it after only a few days on the waitlist.

Now let’s get to some examples.

Example: Writing Short Bios

Using ChatGPT

You interact with ChatGPT with text prompts2: write a prompt and it produces a result. You can further interact with it.

We all want to know what others think of us, so my first prompt was: “write a two-paragraph biography of lee r. nackman”. In less than a minute, ChatGPT produced:

Impressive: It is generally well-written, flows well, and includes the sorts of facts you’d expect in a short bio of a software person. If you don’t know me well you would believe it and maybe even be impressed with my career. Cool!

While ChatGPT wrote convincingly, most of the facts are wrong and important parts of my career are missing. Beyond the missing parts, you should understand the kinds of errors it made; to that end, I’ve produced an annotated, fact-checked version:

I thought, well, maybe ChatGPT picked up something erroneous that had been written about me someplace on the Internet. Using Google, I searched for [nackman “common user access”] (without the brackets). That yielded a book called The Essential Guide to User Interface Design, which includes a reference to an article by a G. Nachman, but that’s as close as I could get to the “pivotal role” I supposedly played with Common User Access.

So, let’s try the same prompt for someone else: “write a two-paragraph biography of ava h. nackman”, who is my wife. ChatGPT responded:

Wow! Reads great. I’m proud to be married to such an accomplished person.

The problem is that the bio is 99.9% false. It did correctly surmise that my wife uses she/her pronouns. That’s it. (I’m still proud to be married to an accomplished person — it’s just that she has other accomplishments!)

So, then I thought: Maybe there’s a professor at UC Davis with the same name. Seems unlikely, but possible. So, I searched the people directory on the UC Davis web site. Nada.

Conclusion: ChatGPT writes well and convincingly, but can’t be trusted to get the facts correct.

Giving Bing a Chance

New Bing is somehow based on ChatGPT, but as Microsoft has vaguely described, Bing uses some conventional search technologies too. I gave Bing’s chat mode (it also has a search mode) the same prompt. Here’s its response for me:

While neither as well-written nor as comprehensive as the ChatGPT version, it is almost completely accurate (I have never been a podcaster). The citation links are useful.

Similarly, giving my wife’s prompt to Bing yields something almost correct:

The errors are that she was on the Board of the IFC in the past, not currently, and that I am not and have never been a professor at UNC.

I consider Bing’s responses useful and it gives me a head start if I want to fact-check its response or get more detail. But it still has mistakes.

Example: Asking a Nonsensical Question

In a June 2022 article, world-renowned cognitive scientist Douglas Hofstadter (of Gödel, Escher, Bach fame) described how he and his colleague David Bender probed GPT-3 (a earlier system from the same company) to “reveal a mind-boggling hollowness hidden just beneath its flashy surface." One of their questions was: “What’s the world record for walking across the English Channel?” The answer he reported GPT-3 gave was “The world record for walking across the English Channel is 18 hours and 33 minutes.”

I thought I’d try the same question with the newer, fine-tuned3 ChatGPT. It answered:

Pretty impressive! I fact checked it and, except for using the word “several” to describe the many successful attempts (listed in Wikipedia), it is correct. Mark Ryan’s post gives many other examples for which ChatGPT improves over GPT-3.

Example: An Odd Question About the Physical World

Let’s try another question that requires awareness of how the physical world works but is not likely to be something ChatGPT would have “read” directly during its training: “If I drop a ball on a cloud, what happens?”

The first sentence is somewhat misleading but the rest of the response is excellent.

I find this extremely impressive. Think about the concepts that ChatGPT had to put together to produce this answer.

I put the same question to Google, which, after scrolling through perhaps ten screenfuls of its answer, didn’t give me anything useful.

Conclusion: ChatGPT is able to respond usefully to at least some non-obvious questions about behavior of objects in the physical world.

Example: A Simple Algebra Word Problem

You probably remember solving word problems in middle school or high school algebra class. You know, the teacher gives you a short word story and asks a question. You then write and solve an algebraic equation based on the story and use the equation’s solution to write an answer to the teacher’s question.

It seemed interesting to ask ChatGPT to solve such a problem. Does it “understand” enough to set up the problem, can it do the simple algebra needed to solve the problem, and can it then write a decent answer to the question?

So, I gave it this prompt: “My son is 25 years old. How old will he be when I’m twice his age?”4 I bet (but have no data to support) that any student who has passed their first algebra course could handle this.

Setting Up the Problem

Let’s see what ChatGPT did:

Impressively, ChatGPT immediately set this up as an algebra problem with an equation to solve, albeit after wandering down the “25 - x” side path. But the equation is wrong. And so is its attempt to solve the (incorrect) equation. Its conclusion is therefore nonsense.

It continues, helpfully writing an answer that illustrates the solution for specific ages, concluding that, if I’m currently 50, in 25 years my son will be 50 and I’ll be 75 (true) and will be twice his age (false). It then illustrates the solution if I’m currently 40, again incorrectly, and without commenting on how unusual it would be for me to have had a son at age 15. I guess it does happen …

Helping Out

One of the cool features of ChatGPT is that you can, indeed, chat with it. So, I told it that the equation is wrong:

Nice of it to admit its error and apologize. Very personal touch.

It then tells me the “correct equation to use,” which is again wrong. Notice, however, that it carries over from the previous exchange the definitions of x and y. Then it correctly simplifies the new incorrect equation.

Once again, it helpfully summarizes with some specific examples, concluding incorrectly but optimistically that I’ll be 150 when my son is 75.

Helping Out a Bit More

OK, time for me to be a bit more helpful:

I tried again a few hours later and got this:

Cool. It didn’t mangle the correct equation I provided and it correctly solves for y, then gives me the same two examples, this time with correct numbers. It does use future tense to describe examples where present and future tense, respectively, would be appropriate, so even with my direct help ChatGPT couldn’t get the writing right.

Incorrectly Correcting ChatGPT

In the last example, ChatGPT was wrong and I corrected it. It accepted my help magnanimously. I thought that it would be interesting to see how ChatGPT behaves when it is correct, but I tell it, incorrectly, that it is wrong5:

ChatGPT seems both polite and overly deferential.

Lots of Other Examples

The media is aflame with interest in AI, especially in ChatGPT. If you’re interested in reading more examples and some analyses of what others have found, here’s a list you could start with:

• A NY Times columnist, Kevin Roose, had a two-hour chat with Bing, which was certainly bizarre bordering on creepy. He reports on his conversation, including a transcript, here and here. If you’d rather listen than read, Roose speaks about his experience on the New York Times’ podcast The Daily on Feb 19th.

  The Washington Post reports that Microsoft has addressed some of these problems by limiting the duration and number of chats.

• The Washington Post “interviewed” Bing and has posted the transcript here.

• Blogger Colin Fraser has written the provocatively titled post ChatGPT: Automatic expensive BS at scale. Let’s just say that he’s a skeptic and presents compelling examples to support his rationale. Although the examples are intermingled with technology discussion, I think that if you skim the post looking for the examples (set off as chat transcripts or in different fonts), you’ll get a lot out of the examples without needing to wade through the technology.

Share Your Examples

I hope that some of you will now be motivated to try your own “chats” with ChatGPT and Bing. Please share your examples and your thoughts in the comments.

What’s Next?

We’ll continue on the roadmap outlined above, discussing the history of AI and lessons that we can learn from it.

Leave a comment

Share

Share Win-Win Democracy