AI, *Muybridge*, and the Mind

Graphic novelist Guy Delisle is most famous for his memoir Pyongyang: A Journey in North Korea. I’ve read virtually every word he’s ever written, and seen virtually every picture he’s ever drawn. After two readings, I’m ready to affirm that Delisle’s new Muybridge is his very best book.

Muybridge (pronounced MY-brij) is at once a graphic biography of photographer-inventor-entrepreneur Eadweard Muybridge, a history of 19th-century photography, and a crash course on cross-species anatomy. At the meta-level, more importantly, it is a profound treatise on how technology changes the worldview of humanity — opening new vistas and creating novel confusions at the same time.

The titular character is famous for proving that horses at full gallop get all four hooves off the ground simultaneously. A topic I couldn’t care less about, honestly. Indeed, I’d never ever heard of Muybridge before I bought the book. But Delisle’s narration totally won me over. After introducing his central character, the writer-artist gives a quick history of photography, then shows how Muybridge built his brand, crisscrossing the mid-19th-century U.S. to capture its natural wonders. Early photography was perfect for landscapes and other stills, but struggled to take clear pictures of living — hence fidgety — beings.

Muybridge in Yosemite. (A self-portrait because his assistant was too scared to pose).

Once Muybridge became a top nature photographer, Leland Stanford — mogul, one-time California governor, and horse enthusiast — hired Muybridge to decisively prove that horses at full gallop get all their hooves off the ground. The project takes years, interrupted by an unrelated capital murder trial. (Muybridge was a cool guy, but a bad man). Standing on the shoulders of other inventors, Muybridge finally develops the technology for sequential high-speed photography. Which solves the galloping horse question — and unlocks untold secrets of the anatomy of man, beast, and bird. And also quickly leads to motion pictures. Read the whole book for details.

The most spell-binding part of the book for me, however, comes near the beginning, when Muybridge is first learning about photography. I keep thinking about these three panels.

“Can the people in the picture see us?” was a common question in the early days of photography. No one at the time wondered if paintings of human beings were human or conscious. But photographs seemed eerily different. When a picture superficially looks exactly like a real human being, it’s easy to start wondering if the picture also has the deeper characteristics of a human being. Since the person in the portrait is looking right at you, maybe he can see you. If he can see you, he’s conscious. He has a mind, like you or me. Hence, alive!

After sympathetically putting yourself in the shoes of people first encountering photography, I hope you’ll still agree that such musings were utterly absurd. You don’t have to understand how eyes work to realize that a drawing of an eye can’t actually see. Nor do you need to understand how vision works to realize that the reason why a drawing of an eye can’t see is not that the drawing is insufficiently realistic. The visual awareness possessed by a cartoon, a quick sketch, a painted masterpiece, and a photograph is always the same — and always equal to zero. Furthermore, even though no one has ever built a new eye, getting the right biological building blocks is crucial. Fleshy material at least has the potential to see, but ink and paper never will.

So what? Humanity is once again at the dawn of an amazing yet unsettling technology: AI. Since human nature remains the same as in the mid-19th-century, many of us feel the same way our forebears felt when they first saw photographic portraits. AI visionaries see the potential to make life better. Good for them! But lots of other folks are going crazy, attributing literal consciousness to software that expertly simulates human communication.

My point: Once you realize that an infinity of perfect pixels don’t make a portrait come to life, you should also conclude that an infinity of perfect words don’t make software come to life, either. Both are technologies of superficial simulation. A photograph is designed to look just like the human it captures. An AI is designed to talk just like the data upon which it is trained.

Superficial simulations are immensely useful for some purposes. I can only imagine how Muybridge would have swooned to use a modern camera phone. But for other purposes, they’re useless. Taking a photo of your dying friend may make you feel better, but it doesn’t actually keep him alive. The same goes for building a chatbot of your dying friend. To simulate a mind is not to save it. Or to update Woody Allen: “I don’t want to achieve immortality through my bot; I want to achieve immortality through not dying.”

What makes me so sure that AIs aren’t even slightly literally conscious? The same thing that makes you sure that portraits of staring people can’t slightly literally see you. Sure, you can pretend that biology or physics let you resolve the latter question. But that’s too clever by half. You don’t need to study STEM for years to realize that photographs or AIs aren’t actually conscious. All you need is common sense philosophy of mind — and if you don’t have that common sense, all the STEM in the universe won’t save you.

Qatar and Reverse Causation

A clear-cut case of GDP causing education, rather than the other way around

[I’m at the University of Wisconsin, Madison the rest of the week! You and all your friends and family are invited to a BBQ at my sons’ apartment on Saturday night at 6:30 PM. 624 West Washington Ave., Madison, WI 53703. 5th floor balcony. Please RSVP acceptances ONLY in the comments.

When I started writing The Case Against Education, I expected a chorus of critics to appeal to the empirics of economic growth: “If education is mostly signaling, why do macroeconomists find a large effect of education on GDP?” From the start, I planned to rebut this argument with an appeal to “reverse causation”: Much of the alleged effect of education on GDP is in fact an effect of GDP on education. After all, regardless of education’s actual economic benefits, any government that gets an infusion of cash is likely to spend a good chunk of the windfall on education.

When I delved into the empirical growth literature, I was surprised to learn that macroeconomists struggle to detect much effect of education on growth. In most data sets, the national education premium is a small fraction of the private education premium. I also discovered a handful of papers testing for reverse causation from GDP to education, suggesting that the economic benefits of education are even smaller than they look. I was glad the reverse causation papers existed, but they were only moderately convincing. I’d hardly call them a smoking gun.

During my trip to Qatar this summer, however, I saw the smoking gun in all its glory. The gun’s name is Education City. It’s a vast complex, populated by eight different universities and about 11,000 students, in a country with under 3 million people, a supermajority of them guest workers. Estimates of the total cost of building Education City are hard to pinpoint, but there is little doubt that the Qatari government spent billions on the project. The six foreign universities based in Education City get about $400M annually.

One small part of Education City

To belabor the obvious, the overwhelming cause of Qatar’s immense wealth is not education, but fossil fuels. Annual industry profits vary widely with global fuel prices, but they exceeded $40B in 2022. There is no chicken-and-egg problem here. Qatar got rich decades before it got educated. Phase 1: Import foreign energy talent to capitalize on abundant natural resources. Phase 2: Spend billions on a vast range of prestige projects. Qatari GDP caused Qatari education; Qatari education did not cause Qatari GDP.

Will Education City eventually yield economic benefits? The university spokesmen assured me that higher ed was already working economic wonders, but I could only roll my eyes. Most of the programs they’re offering are totally disconnected from Qatar’s comparative advantage:

Our higher education offerings—ranging from undergraduate to postdoctoral degrees—include programs in media, international affairs, business, computer science, medicine, engineering, cultural heritage, knowledge management, and arts.

Granted, Education City is gorgeous. Dipped in gold, really. I had fine fun touring around, witnessing the wonders that Qatar, like United Arab Emirates, has built by the power of mass migration. But the whole project screams not “investment,” but “vanity project.” Also known as “waste.”

At minimum, Qatar is an existence proof: reverse causation from GDP to education definitely happens. But on reflection, it is much more. Almost any country that strikes oil — literally or figuratively — would act like Qatar. Education is drenched in Social Desirability Bias, so almost every demagogue wants to fund it generously. Not “whatever the cost,” because education is hardly the only product drenched in Social Desirability Bias. But education is so beloved that the U.S. has had three “education presidents” in my lifetime alone.

To their great dismay, most of the world’s ruling demagogues aren’t swimming in petro-dollars. They can’t afford to build another Education City, or anything comparable. So they make do. If and when they get their hands on extra cash, they loudly give education a sizable cut of the loot. Like the Qataris, they loudly promise immense long-run benefits. And like the Qataris, they carefully avoid asking pointed questions about the curriculum.

Not all curricula are equally terrible, so not all education is equally economically wasteful. But standard estimates of education’s economic benefits are always even smaller than they look. Why? Because demagogues run the world. “Every citizen of our great country deserves the best possible education” is music to humanity’s ears. Anyone who seeks to rule fellow humans is well-advised to blast such music so loud they can’t even hear the wonkish words “reverse causation.” Education isn’t wonderful, but spending taxpayer money on education is a wonderful way to gain and hold power.

Bambauer Contra BS

My friend Jane Bambauer at the University of Florida wants your input on her new misinformation project. Here’s her pitch. Enjoy!

---

I am running a research program at the University of Florida aimed at getting policymakers and intellectuals to accept that the "misinformation problem" is (1) a demand-side problem rather than a supply-side one, (2) much more prevalent in mainstream media than scholars and elite society tend to recognize, and (3) is best tackled by making confirmatory bullshit socially embarrassing.

I have a few research projects in the works that could benefit from the input of thoughtful economists, philosophers, lawyers, and other Caplanites. You can use this form for easy input and to express interest in attending convenings or getting involved, but I’ll also pay attention to the comments section.

Using LLMs to Measure Bullshit / Grade the News

1. Grading old predictions. We are building an LLM tool that can scan old news stories to identify concrete-enough-to-grade predictions, identify a "due date," and then grade the prediction if the due date has passed (and otherwise store for future grading.) We are also building a model that works backwards from existing prediction markets: we look at the betting markets that Polymarket or Kalshi or whatever have set up in the past, then we go search for old news items that made an overt or implied prediction, and we imagine that the speaker or author had placed an appropriate bet in that market on the date that the claim was made. We have some ideas for dealing with tentative predictions (“might” language), and we also have ideas for penalizing speakers who imply predictions but use vague or weaselly language that avoids falsification. When we are happy with performance, we will use UF's supercomputer to do this at scale.

Why:
(1) More trustworthy than fact-checking. Fact-checking applies to claims that are already contested. By contrast, usually there is little debate about whether a prediction has or has not been met—hence the reasons that prediction markets are able to clear without constant controversy.
(2) Reputation scores for experts. Philip Tetlock used predictions to see whether experts do a good job aggregating and synthesizing knowledge. His studies found “meh, a bit better.” But it couldn’t be done at scale before LLMs.
(3) Expert reputation is really important! Most knowledge is built from testimony of others rather than direct experiences and observations, so to quote Michael Huemer, “The key question would then be why do you think that person is reliable about p [the proposition]. Perhaps you have heard this person speak about p-related matters before, you have frequently checked on her p-related assertions, and you have generally found them to be correct. Thus, you inductively inferred that this person is reliable about things related to p… Sad to say, I really haven't checked on very many people at all. I haven't taken a large, random sample of assertions by various humans and verified their truth, as would seem to be required for an inductive inference to the general veracity of people's statements.”

2. Grading for Cherry-picking. We are also building a similar LLM tool that grades news stories based on when misleading examples are used to support a narrative. This program scans a news story to identify explicit or implicit factual claims, and also identifies the evidence/examples used to support the claim. It then does an independent assessment of available data to see if the examples are representative given the claim. This almost always boils down to comparing two ratios, so usually cherry-picking can be detected by having good data on 4 numbers (2 numerators, 2 denominators). Bryan's post on the "Chinese robbers" fallacy would get picked up by this.

3. Proportion of epistemically useless content. We have also just started to build LLM tools that parse content into (a) premises, (b) claims, (c) arguments, (d) evidence, and (e) something that is none of that stuff. (e) is bad. It's usually unjustified narrative control or emotional appeals. What proportion of a news story is (e)? Also, we are hoping to make a tool that looks for logical fallacies among the (c) arguments, and this might be a good alternative to the fact-checking approach because you don't need any source of ground truth.

Market-Based Institutions Beyond Betting/Prediction Markets

4. News guarantees / bounties. Marshall van Alstyne (BU), Yonathan Arbel (Alabama), and a few others have written about using contractual agreements to let people sue you if they think you are wrong to create reliable signals of quality. There are a few ways to set up such a program, but no content creator has taken this up (other than Bryan, of course.) If we were able to do field experiments, how should we structure a bounty? Example: the content creator guarantees factual claims made in a story and deposits $10k in an escrow. A challenger would have to pay the costs of a challenge, and then the two sides would put on a trial for an online jury. Alternatively, the two sides could each select a judge who is blinded to who selected them, as is the system at Rootclaim. There are probably other models, too. What type of system would be administrable and widely trusted? If Meta could introduce a new revenue stream by serving as the escrow for content creators that agree to a bounty system, and participation in the system would boost their content in the curation algorithm, could this provide a decentralized and credible signal for quality? (Marshall gets credit for this Meta idea.)

2. Prediction Market-Funded Investigative Journalism. Why don’t news companies that put tireless work into collecting and analyzing new information place bets on the prediction markets before publishing their articles? The authorities might consider this illegal but that would be bad policy and also probably unconstitutional under the First Amendment. I suspect that’s not the main reason journalists don’t do it, anyway. What are the other reasons? Can this be a way to establish a new business model for careful journalism? (Thanks to my colleague Peter Molk for this idea.)

Again, the link to the input form is here. Thanks very much to Bryan for allowing me to reach his “beautiful bubble” of regular readers.

"Aging Out" of Drug Addiction is the Norm

Abusers really do give vice a bad name.

The stereotypical drug user is a life-long addict. While stereotypes are usually good statistical approximations of the truth, I’ve long had the sense that this particular stereotype is false. The infamous heroin study, for example, found that the vast majority of U.S. soldiers who used heroin in Vietnam quit when they came home. Only recently, however, did I discover much more systematic evidence.

The source: Gene Heyman’s Addiction: A Disorder of Choice, published by Harvard University Press in 2010. In chapter 4, “Once an Addict, Always an Addict?,” Heyman tracks down nationally representative samples of drug users, finding that people in drug treatment programs are far less functional than the much larger segment of the population that has used drugs. I’ve said it before and I’ll say it again: Abusers give vice a bad name.

What exactly does Heyman have to teach us? Let’s start with the data:

There are four large national studies that recruited representative populations and provide data relevant to relapse and remission rates for addiction (Anthony & Helzer, 1991; Kessler et al., 2005a, 2005b; Stinson et al., 2005, 2006; Warner et al., 1995). Although these studies were conducted by leading researchers and funded by various national health institutes, the findings have not become a staple of discussions of the nature of addiction. For example, none of the clinical texts and journal articles cited in the introduction to this chapter reference these epidemiological studies. Thus, one of the goals of this chapter—and this book—is to introduce this information to addiction researchers as well as the public, although the research itself is not new.

The Epidemiologic Catchment Area Study (1980–1984) is the earliest study. Results:

Remission rates in the ECA survey. The next figure summarizes the ECA findings on remission. It shows the percentage of individuals who reported no drug-related problems for at least twelve months prior to the survey, but had met the lifetime criteria for drug dependence or drug abuse (based on Table 6.22 in Anthony & Helzer, 1991). The data are organized in terms of gender and age. For example, according to the biographies, remission rates should top 50 percent by age 30.

Figure 4.1 shows that at approximately age 24 more than half of those who ever met the criteria for addiction no longer reported even one symptom, and that by about age 37 approximately 75 percent of those who ever met the criteria for dependence were no longer reporting any symptoms. Since dependence requires at least three symptoms over a twelve-month period, it is likely that the proportion of those in their thirties who were still dependent was actually less than 25 percent. Also note that it must be the case that most of those who quit did so outside of the purview of drug treatment clinics.

Next, Heyman reviews the National Comorbidity Survey (NCS), 1990–92 and replication (2001–02). Background:

In the early 1990s and then again in the years 2001–2002, Ronald Kessler directed two large, nationwide surveys of mental health and mental health services (Kessler et al., 2005a, 2005b; Warner et al., 1995). As was the case with the ECA study, the goal was to provide an unbiased scientific account of key characteristics of psychiatric disorders. The research reports emphasized the associations between different disorders (e.g., the correlation between depression and addiction), and hence the project was titled the National Comorbidity Survey. In addition to comorbidity, the measures included lifetime prevalence rates, current prevalence rates, age of onset, and demographic indices. But, in contrast to the ECA survey, nonmetropolitan as well as metropolitan areas were sampled. There were about 8,100 and 9,300 subjects in the two NCS studies.

Results:

Figure 4.2 compares the remission rates for the ECA and 1990–1992 National Comorbidity surveys. Not counting addiction, the two surveys were in close agreement. The average (absolute) difference in remission rates was just 4 percent. For addiction, however, the two surveys came up with very different estimates of the percentage of “ex-addicts.” The NCS reported that 74 percent of lifetime addicts were in remission, an even higher rate than the ECA study (59 percent). This is by far the largest difference between the two studies.

The difference is due to two factors. First, the ECA survey used a much more liberal standard for counting current cases. Instead of the typical three symptoms, only one symptom was needed.

There were two more follow-up studies. All the results together:

In a nutshell: “Replication is the gold standard for scientific findings. By this criterion addiction is not chronic. Indeed, it could be said that it is just the opposite: self-limiting. Note that nothing has been said about treatment. This is because most of those who quit were not in treatment.”

Next Heyman considers the possibility that “remission” rarely lasts. News flash: Most heavy users really do “age out.”

The idea that addiction consists of relatively short periods of heavy drug use followed by relatively longer periods of abstinence is easily tested. It predicts that one-time surveys will show that the overall rate of addiction remains approximately constant as a function of age. The next graph, Figure 4.4, tests this prediction…

For both cohorts, the age trend is sharply decreasing, suggesting that when those who met the criteria for lifetime drug dependence quit, they usually quit for good… To be sure, addicts may relapse the first few times they try to quit, but, according to the surveys, by about age 30 most have quit for good. Thus, for addiction, quitting drugs is more accurately described as “resolution,” not “remission,” when the ex-drug user is more than 30 years old.

Note that when these studies were conducted, overdose deaths were nowhere near high enough to explain the rarity of older users. While young users have elevated death rates, the vast majority survived long enough to quit.

Why do even many experts hold unrealistically negative stereotypes about drug users? Selection bias!

The simplest explanation of the discrepancy between the research findings and received knowledge regarding the nature of addiction is that experts are basing their understanding of addiction on addicts who show up in the treatment clinics, whereas the research reviewed in this chapter is based on studies that selected subjects independent of treatment history, with the goal of obtaining a representative sample. Both approaches would lead to similar results if most addicts ended up in treatment. But most addicts do not seek treatment. (emphasis added) Given that the clinic studies support the claim that addiction is a chronic disorder, this means that addicts who end up in treatment keep using drugs, treatment notwithstanding, and those not in treatment quit using drugs. This interpretation fits all the data presented so far, and it also suggests an interesting hypothesis…

First, it is reasonable to suppose that differences in pharmacological history distinguish the two groups.

Second, it is just as plausible that individual differences distinguish the two groups. These are not mutually exclusive explanations and both promise to increase our understanding of the determinants of drug use in addicts.

The first story is wrong: “The graph shows that the remission rates for the three drug classes did not differ markedly. The implication is that the higher remission rates for clinic addicts is not a function of using drugs that are more addictive.”

The second story, in contrast, is right. People in treatment programs are strongly selected for dysfunction.

The most important is that addicts in treatment are much more likely to suffer from additional psychiatric disorders than those not in treatment. In the ECA survey the ratio was more than two to one (Regier et al., 1990). About 64 percent of those in the clinics suffered from at least one additional psychiatric disorder, whereas for those who said they had not sought clinic help for drug use, the comorbidity rate was about 30 percent, a number that is closer to the psychiatric prevalence rate for nonaddicts (see Figure 4.7).

Note: You don’t have to believe that “nondrug psychiatric disorders” are literally diseases to believe that people diagnosed with psychiatric disorders are especially prone to poor life choices! As Thomas Szasz quipped, “They are not disturbed; they are disturbing.”

Last question: How does “aging out” work in practice?

According to the biographies, financial and family concerns were among the main reasons addicts quit using drugs. Economics and family responsibilities are also the concerns of adulthood. This suggests that the pressures that typically accompany “maturity” bring drug use to a halt in many addicts. Assuming this to be the case, then one of the reasons that comorbidity promotes drug use in addicts is that it gets in the way of adult roles. It seems reasonable to suppose that those who are very depressed or very anxious are less likely to be engaged in activities that are incompatible with heavy drug use. Put another way, psychiatric impairment renders the drug experience relatively more valuable by undermining the ability to engage in and enjoy competing activities. The more general message is that whether addicts keep using drugs or quit depends to a great extent on their alternatives.

What difference does all this make? I don’t know Heyman’s view, but here’s mine:

While I would be horrified if any of my kids tried hard drugs, I’d also be horrified if my kids tried climbing mountains, racing cars, worshipping Satan, visiting Haiti, or even getting visible tattoos. Yet my horror does not show that these decisions are literal “diseases,” nor does it justify depriving my children of their freedom when they reach adulthood.

Medicalizing bad decisions is at best a Noble Lie. But once you realize that most people who make bad decisions eventually course-correct, the lie looks ignoble indeed. Why should everyone be deprived of their freedom to marginally help a tiny self-destructive minority? The righteous path is to harshly punish drug abusers for violence, threats of violence, theft, and trespassing, firmly stigmatize abusers, and stop losing sleep over the tens of millions of casual users who surround us.

Israetel's Favorite Economist

An hour-long intro to my work

Mike Israetel is a exercise scientist and fitness guru, with over 3 million subscribers to his main Youtube channel. He’s also in the running to be my biggest fan. He recently released this detailed introduction to the work of “his favorite economist”… who turns out to be me. Enjoy!

P.S. Mike and I will probably be doing a podcast together in the foreseeable future. Got questions for either of us? Topics you’d like us to cover? Please share in the comments.