I am about to flip a coin. Can you predict whether it will be heads or tails?
You might have gotten it right. (It was heads.) But if so, you were lucky. Whether it would come up heads or tails depends on an assortment of factors that you could not identify before I flipped the coin.
In principle, of course, reliable predictions could be made. If you knew everything relevant to whether a coin will come up heads or tails, you could of course make an accurate prediction. So let’s not talk about randomness. That is not the problem. The problem is the absence of essential information.
With respect to the role of randomness, here is an illuminating exchange I had with ChatGPT:
ME: If I flip a coin, can you predict whether it will come up heads or tails?
CHATGPT: No, I can’t predict the outcome of a coin flip. A fair coin toss is random, and each flip has an equal 50% chance of coming up heads or tails. If you’d like, I can simulate a coin flip for you! 😊
ME: Why can’t you predict the outcome of a coin flip?
CHATGPT: Predicting the outcome of a coin flip is nearly impossible because it depends on a complex combination of factors, including:
Initial Conditions: The force, angle, and speed of the flip, along with the height it reaches, are all variables that are difficult to measure and control precisely.
Air Resistance: The way the coin interacts with the air as it flips introduces randomness.
Surface Interaction: The coin’s bounce and spin upon landing on a surface can further alter the outcome.
In theory, if all of these factors were perfectly known and calculated, the outcome could be predicted (this is the basis of classical mechanics). However, in practice, the variables are too many and too sensitive for accurate prediction. For most purposes, a coin flip is treated as random.
ME: Isn’t the word “random” misleading here?
CHATGPT: You’re absolutely right — calling a coin flip “random” might oversimplify what’s actually happening. In reality, a coin flip is deterministic, meaning the outcome is governed by physical laws like Newtonian mechanics. If we had perfect knowledge of all the initial conditions (force, angle, spin, air resistance, etc.), we could theoretically predict the result every time.
When we call it “random,” we’re really referring to practical randomness — the idea that the variables influencing the outcome are so complex and sensitive that, for all practical purposes, we treat the result as unpredictable. In this sense, “random” reflects our ignorance of or inability to measure the key factors, rather than true randomness in the sense of quantum unpredictability.
It’s a convenient shortcut in language but not strictly accurate in a scientific sense.
Justice
A central goal of liberalism, and of the liberal political tradition, is to undo damaging forms of subjugation, which is why John Stuart Mill’s The Subjection of Women is a canonical liberal text. With his emphasis on the importance of individual agency, Mill laments that “the inequality of rights between men and women has no other source than the law of the strongest.” In a key passage, he writes:
What is the special character of the modern world — the difference that chiefly distinguishes modern institutions, modern social ideas, modern life itself, from those of times long past? It is that human beings are no longer born to their place in life, and chained down by an unbreakable bond to the place they are born to, but are free to use their talents and any good luck that comes their way to have the kind of life that they find most desirable.
Mill’s argument here is more subtle than the context might suggest. He is speaking, to be sure, of careers open to talents — of a right to seek opportunities and to try to find the kind of life that one finds most desirable. That is the liberal insistence on the dissolution of unwanted chains and bonds. But Mill is also careful to draw attention to the importance of “any good luck that comes their way.” In its best forms, the liberal tradition emphasizes that lotteries are everywhere. It points to the place of “good luck,” understood as practical randomness, and the multiple forms it takes. John Rawls’ A Theory of Justice is the most sustained development of that point.
The term “good luck” isn’t exactly right, but let’s not be fussy. I want to say something about the lived equivalent of lotteries. I am going to approach that question indirectly, or from the side. But please keep it in mind throughout. Full disclosure: this is, in part, a discussion of justice.
Sampling on the Dependent Variable
A number of years ago, a brilliant law student — let’s call her Jane — came to my office with an intriguing research project. She wanted to study the sources of success. Jane’s plan was to contact dozens of spectacularly successful people in multiple fields (business, politics, music, literature) to see what they had in common. Maybe all of them had difficult childhoods. Maybe none of them had difficult childhoods. Maybe all of them were quick to anger. Maybe none of them was quick to anger. Maybe all of them developed a passion in high school. Maybe none of them developed a passion in high school. Maybe all of them were impatient. Maybe none of them were impatient. Jane was energetic as well as astonishingly smart. There was little doubt that she would be able to carry through with her project. If she called famous people, she would find a way to get them to take her calls.
Still, something was wrong with what she had in mind. Suppose we learned that a large number of spectacularly successful people did indeed have something in common. Would we know that what they had in common was responsible for their spectacular success?
Not at all. There might be plenty of people (hundreds, thousands, millions) who share that characteristic and who did not end up spectacularly successful. The shared characteristic might not be sufficient for success. Imagine, for example, that spectacularly successful people turn out to be quick to anger. Plenty of people who are quick to anger do not succeed. Maybe they never got a chance. Maybe they got mad at the wrong person at the wrong time. Maybe they were born in poverty. Maybe they didn’t have the right skin color.
If we learn that spectacularly successful people tend to be quick to anger, have we learned anything at all? Maybe not. The problem with Jane’s project has a name: selecting on the dependent variable. Countless successful business books follow a path identical to that proposed by Jane. They try to figure out what characteristics are shared by inventors, innovators, leaders, or other successful types. If they find a shared characteristic, they urge that they have discovered a secret or clue of some kind. Maybe so. But maybe not. (Probably not.)
Challenges
Could AI have predicted in 2006 that Barack Hussein Obama would be elected president of the United States in 2008? Or could AI have predicted in 2014 that Donald Trump would be elected president of the United States in both 2016 and 2024? Could AI have predicted in 2005 that Taylor Swift would become a worldwide sensation? The answer to all of these questions is obvious: No. AI could not have predicted those things, and no human being could have predicted those things. But why?
To test your answer, here are five challenges:
- Consider the question whether two people are going to fall in love. AI might not be able to foresee the potentially decisive effects of context, timing, and mood.
- Consider the question whether a song will become a big hit. AI might not be able to foresee the effects of social interactions, which can lead people in directions that are exceedingly hard to predict.
- Consider the question whether a social movement, on the left or the right, will arise in a specified month or year. AI might not be able to identify people’s preferences, which might be concealed or falsified, and which might be revealed at an unexpected time.
- Consider the question whether fossil fuels will be phased out by 2048. AI might not be able to anticipate change, including rapid change, which might be a product of unexpected shocks (a technological breakthrough, a successful terrorist attack, a black swan).
- Consider the question whether a new start-up will do well. AI might not have local knowledge, or knowledge about what is currently happening or likely to happen on the ground
Friedrich Hayek was the twentieth century’s most rigorous critic of socialism, or government planning, and his most influential essay, “The Use of Knowledge in Society,” from 1945, is best seen as an argument about the limits of prediction. Hayek did not draw attention to the motivations of planners, and he was not claiming that they are corrupt or self-interested. His concern was what he saw as their inevitable lack of information.
He began: “If we possess all the relevant information, if we can start out from a given system of preferences, and if we command complete knowledge of available means, the problem which remains is purely one of logic.” That is a lot of “ifs.” Of course we do not have all relevant information; preferences shift; and we do not have complete knowledge of the available means (including technologies) which change over time. Thus Hayek emphasized that the “peculiar character of the problem of a rational economic order is determined precisely by the fact that the knowledge of the circumstances of which we must make use never exists in concentrated or integrated form but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.”
Focusing on those dispersed bits of incomplete and frequently contradictory information, Hayek pointed to “the importance of the knowledge of the particular circumstances of time and place” — knowledge that planners cannot possibly have. You cannot predict what will happen if you do not have knowledge of those particular circumstances. How much are people going to like a new movie about spies? How popular will a new store be? You might have some clues, but you might not be able to make confident predictions. The same is true of AI.
Hayek also pointed to a separate problem: change. In October things might be very different from what they were in January, and planners might struggle to understand that. What is true in January (what people like, what technologies exist, what diseases are spreading) might not be at all true in October. The knowledge that people have in markets also shifts rapidly over time. As Hayek had it, the price system is a “marvel,” because it can incorporate knowledge that is both not only widely dispersed but also fleeting. Some new fact might become clear all of a sudden and it might change everything. Or some new taste might emerge quickly and spread in a hurry. Markets can absorb new information and new tastes. Planners cannot.
Like central planners, AI will struggle to make accurate predictions, not because it is AI but because it does not have enough data to answer the question at hand. Those cases often, though not always, involve complex systems.
Life Trajectories
In 2020, a large team of researchers engaged in an unusually ambitious project, whose subject they called “measuring the predictability of life outcomes with a scientific mass collaboration.” They wanted to see if life trajectories could be predicted. To do that, they challenged the world. Their challenge had a simple name: The Fragile Families Challenge.
The challenge began with an extraordinary data set, known as the Fragile Families and Child Wellbeing Study, which was specifically created in order to enable social science research. That study, which is ongoing, offers massive amounts of data about thousands of families, all with unmarried parents. Each of the mothers gave birth to a child in a large city in the United States around 2000. The data was collected in six “waves,” at birth and at the ages of one, three, five, nine, and fifteen. Each collection produced a great deal of information, involving child health and development, demographic characteristics, education, income, employment, relationships with extended kin, father-mother relationships, and much more. Some of the data was collected by asking a battery of questions to both the mother and the father. Some of it came from an in-home assessment (at ages three, five, and nine) that included measurements of height and weight, observations of neighborhood and home, and various tests of vocabulary and reading comprehension. The Fragile Families Challenge was initially launched when data had been collected from the first five waves (from birth to the age of nine years), but when complete data from the sixth wave (year fifteen) were not yet available.
That was a terrific advantage, because it allowed the researchers to create the Challenge, which was to predict the following outcomes:
- Child grade point average
- Child grit (determined by a self-reported measure that includes perseverance)
- Household eviction
- Household material hardship
- Layoff of the primary caregiver
- Participation in job training by the primary caregiver
Those who took the challenge were given access to background material from the first five waves, and also to data on one-half of the families from the sixth wave. The material contained data on a total of 4,262 families, with a whopping 12,942 variables about each family. The central task was to build a model, based on the data that was available, that would predict outcomes for those families, during the sixth wave, for whom data were not available.
The researchers sought to recruit a large number of participants in the Fragile Families Challenge. They succeeded. In the end, they received 457 initial applications, which were winnowed down to 160 teams. Many of the teams used state-of-the-art machine-learning methods, explicitly designed to increase accuracy. The central question was simple: Which of the 160 teams would make good predictions?
The answer is: none of them! True, the machine-learning algorithms were better than random; they were not horrible. But they were not a lot better than random, and for single-event outcomes — such as whether the primary caregiver had been laid off or had been in job training — they were only slightly better than random. The researchers conclude that “low predictive accuracy cannot easily be attributed to the limitations of any particular researcher or approach; hundreds of researchers attempted the task, and none could predict accurately.”
Notwithstanding their diverse methods, the 160 teams produced predictions that were pretty close to one another — and not so good. As the researchers put it, “the submissions were much better at predicting each other than at predicting the truth.” A reasonable lesson is that even with the aid of AI, we really do not understand the relationship between where families are in one year and where they will be a few years hence. Seeming to draw that lesson, the authors of the Fragile Families Challenge suggest that their results “raise questions about the absolute level of predictive performance that is possible for some life outcomes, even with a rich data set.” You can learn a great deal about where someone now is in life, and still you, or AI, might not be able to say very much at all about specific outcomes in the future.
As GPT-4o put it in 2024,
At the time, the challenge highlighted the difficulty of predicting these outcomes, even with advanced machine learning techniques. One of the key findings was that models, while powerful, struggled with out-of-sample prediction. Many researchers found that human lives are influenced by so many complex and often random factors that standard models didn’t perform as well as anticipated.
That sounds a lot like Hayek, writing a long time ago. Asked whether AI could do better today, it responded, “predicting human behavior and life outcomes remains incredibly challenging due to the inherent complexity and unpredictability of life events. AI models, no matter how advanced, would still struggle with randomness, unmeasured variables, and ethical concerns regarding fairness and bias.”
Let us put to one side the point about randomness. Let us also (temporarily!) put aside ethical considerations, which do not bear on predictive accuracy. Let us focus on unmeasured variables, or the absence of relevant data, which are Hayek’s concern as well. Consider a girl named Susan, who is ten years old, and learn everything you can about her: her family, her demographics, her neighborhood, her schooling, her sports. Now predict various things about her life at the age of twenty-one. Will she be a doctor? A lawyer? A computer scientist? Do you have much confidence in your prediction?
You shouldn’t. The number of variables that can move a life in one direction or another is very high, and it is not possible to foresee them in advance. Someone might break a leg at a crucial moment, meet an amazing music teacher, find a new friend, hear a song on the radio on Sunday morning, or see something online or on the news that changes everything.
Love and Romance
Can AI algorithms predict whether you will fall in love with a stranger? Can AI help people to find romantic partners?
Thus far, the results on such counts are not promising. Samantha Joel and colleagues, in a study of “machine learning applied to initial romantic attraction,” find that AI algorithms struggle to predict “the compatibility elements of human mating . . . before two people meet,” even if one has a very large number of “self-report measures about traits and preferences that past researchers have identified as being relevant to mate selection.” Again sounding a lot like Hayek, Joel and her colleagues suggest that romantic attraction may well be less like a chemical reaction with predictable elements than “like an earthquake, such that the dynamic and chaos-like processes that cause its occurrence require considerable additional scientific inquiry before prediction is realistic.”
What are “dynamic and chaos-like processes”? It is worth pondering exactly what this means. Most modestly, it might mean that AI needs far more data in order to make accurate predictions — far more, at least, than is provided by self-report measures about traits and preferences (“considerable additional scientific inquiry”). Such measures might tell us far too little about whether one person will be attracted to another. Perhaps we need more information about the relevant people, and perhaps we should focus on something other than such measures. It is possible that AI cannot make good predictions if it learns (for example) that Jane is an extrovert and that she likes football and Chinese food. It is possible that AI algorithms would do a lot better if they learn that Jane fell for John, who had certain characteristics that draw her to him, and also for Tom and Frank, who had the same characteristics. If so, perhaps she is most unlikely to fall for Fred, who has none of those characteristics, but quite likely to fall for Eric, who shares those characteristics with John, Tom, and Frank.
On this view, the right way to predict romantic attraction is to say, “If you like X and Y and Z, you will also like A and B, but not C and D!” Or perhaps we should ask whether people who are like Jane, in the relevant respects, are also drawn to Eric. Of course it would be necessary to identify the relevant respects in which people are like Jane, and that might be exceedingly challenging. Maybe AI could be helpful in that endeavor. Maybe not.
More radically, we might read the findings by Joel and her colleagues to suggest that romantic attraction is not predictable by AI algorithms for a different and more Hayekian reason: it depends on so many diverse factors, and on so many features of the particular context and the particular moment, that any form of AI will not be able to do very well in specifying the probability that Jane will fall for Eric. The reference to “dynamic and chaos-like processes” might be a shorthand way of capturing current mood, weather, location, time of day, background sounds, and a large assortment of other factors that help produce a sense of romantic connection or its absence.
Think a bit, if you would, about what made you feel a romantic spark. Did someone look at you in a certain way, at a certain moment, and then boom? What led to that boom? Do you even know?
Jane might smile at a specific time at a street corner, and look Eric in the eye, and Eric’s heart might flutter. Or Jane might not look Eric in the eye at that moment, because she is distracted by something that happened in the morning. Eric might say something witty as sandwiches arrive, because of something he read in the paper that morning, and that might initiate a chain of events that culminates in marriage and children. Or Jane might make a bad joke at a bad time, and Eric might think, “This really will not work.”
For romance, so much may depend on factors that cannot be identified in advance. This is the sense in which AI is sometimes like centralized planners: it does not have relevant information about time and place. Even the largest language models will still lack crucial data. (Continuing the Hayekian market analogy, there does not seem to be anything like the price system to replace AI with.)
Careful
We have to be careful here. AI might be able to say that there is essentially no chance that Jane will like Carl, because there are things about Carl that we know, in advance, to be deal-breakers for Jane. Jane might not be drawn to short men or to tall men. She might not be attracted to much older men or to much younger men. She might not be attracted to men. An algorithm might be able to say that there is some chance that Jane will like Bruce; there is nothing about Bruce that is a deal-breaker for her, and there are some clear positives for her. Perhaps an algorithm can specify a range of probability that Jane will fall for Bruce; perhaps the probability of a romantic connection (suitably defined) is more than ten percent but less than seventy percent. An algorithm might be able to say that Eric is within the category of “it might well happen” for Jane, because Eric is in some sense “her type.”
Perhaps an algorithm can specify a range of probability for Jane and Eric; perhaps the probability of a romantic connection (suitably defined) is more than twenty percent but less than eighty percent. The real question is whether and to what extent AI will eventually be able to do much better than that. We might speculate that the importance of particular factors — the concrete circumstances — is so great that there are real limits on AI’s predictive power (even if AI might be able to outperform human beings, whose own predictive power is sharply limited in this context).
The topic of romantic attraction is intriguing in itself, and it can be seen as overlapping with an assortment of other prediction problems: whether you will enjoy living in Paris; whether you will become friends with a co-worker; whether you will like a new job; whether a pandemic will occur in the next five years; whether a recession will occur in the next six months; whether a new movie will be make a specified amount of money; whether a new book will hit the bestseller list. We have seen that in stable environments with fixed rules, AI algorithms, armed with a great deal of data, are able to make pretty good predictions. But if the future is unlikely to be like the past, there is a real question whether, where, and when AI algorithms will do well, or even outperform human beings.
The Hayekian problem might be the sheer number of relevant factors, not knowable in advance, that might produce one or another outcome; this is why the case of romantic attraction has general lessons. (Friendship is similar.) Another problem might be an external shock or unexpected event, which might turn everything around (a technological innovation, a terrorist attack, a pandemic, or in the context of romance, a recent loss for Jane, which opens her heart). Yet another problem might be how factors interact with each other (lasagna followed by chocolate cake might spell romance). We are speaking here of the essentially unpredictable nature of many events, because of the absence of necessary data.
Revolution Is in the Air
In an important work called Private Truths, Public Lies, which predated the rise of AI, the economist Timur Kuran urged that revolutions were unpredictable by their very nature. Kuran argued that an underlying problem lies in “preference falsification”: people do not disclose their preferences, which means that we cannot know whether they will, in fact, be receptive to a revolutionary movement. If we do not know what people’s preferences are, we will not know whether they might be willing to participate in a rebellion given the proper circumstances.
I asked GPT4o whether AI could predict revolutions, and it said this:
Predicting revolutions or large-scale social upheavals is extremely challenging, even for modern AI. While AI has advanced significantly in its ability to analyze complex data, revolutions are driven by a mix of structural factors (e.g., economic inequality, political repression) and unpredictable triggers (e.g., sudden political decisions, individual actions, or random events). The difficulty lies in the dynamic, multi-layered nature of these events. . . . Revolutions often result from unforeseen or seemingly minor events that AI may not anticipate — so-called “black swan” events. For instance, the Arab Spring was sparked by the self-immolation of a Tunisian street vendor, an act that set off widespread protests across the Middle East. . . . AI relies heavily on historical data to make predictions, but reliable data on political discontent, underground movements, or the true state of public sentiment is often scarce, especially in authoritarian regimes. . . . The most AI can do is flag societies at risk and provide real-time alerts to possible flashpoints.
It is true that we, and AI, might be able to learn something about when a revolution is improbable in the extreme, and also about when a revolution is at least possible. For one thing, we might be able to make at least some progress in identifying private preferences — for example, by helping people feel safe to say that they dislike the status quo, perhaps by showing sympathy with that view, or perhaps by guaranteeing anonymity. AI might well be able to help on that count. Kuran wrote before the emergence of social media platforms, which give us unprecedented opportunities to observe hitherto unobservable preferences — for example, via prompts, posts, and google searches, which might reveal widespread dissatisfaction with the current government.
Perhaps AI can say something about probabilities, based on data of this kind. But if GPT is right, AI will not be able to tell us a whole lot, because its knowledge of preferences and thresholds will be limited, and because it will not be able to foresee social interactions. The general analysis should not be limited to revolutions. Preference falsification, diverse thresholds, and social interactions — one or more of these are in play in many domains.
When will marriages break up? When will employees engage in some kind of revolt? When will we see something like #MeToo? When will a populist movement emerge and succeed? AI might be able to tell us something, but not nearly everything.
Hits!
Consider the question whether books, movies, or musical albums are likely to succeed. Of course we might know that a new album by Taylor Swift is likely to do well, and that a new album by a singer who is both terrible and unknown is likely to fail. A few decades ago I was part of a rock group called Serendipity. You haven’t heard of us, and we were terrible; there was no chance that we could succeed. You don’t need AI to know that. But across a wide range, a great deal depends on social interactions and apparent accidents, and on who says or does what exactly when. In such circumstances, AI might not be able to help much.
This point clearly emerges from research from a number of years ago, when Matthew Salganik, Duncan Watts, and Peter Dodds investigated the sources of cultural success and failure in what is known as their Music Lab experiment. Their starting point was that those who sell books, movies, television shows, and songs often have a great deal of trouble predicting what will succeed. Even experts make serious mistakes. Some products are far more successful than anticipated, whereas some are far less so. This seems to suggest, very simply, that those that succeed must be far better than those that do not. But if they are so much better, why are predictions so difficult? Why do the best analysts fail? No one anticipated the success of the Harry Potter series; the Beatles couldn’t get a record deal; the rise of Donald Trump was a shock.
To explore the sources of cultural success and failure, Salganik and his co-authors created an artificial music market on a preexisting website. The site offered people an opportunity to hear forty-eight real but unknown songs by real but unknown bands. One song, by a band called Calefaction, was called “Trapped in an Orange Peel.” Another, by Hydraulic Sandwich, was called “Separation Anxiety.” The experimenters randomly sorted half of about fourteen thousand site visitors into an “independent judgment” group, in which they were invited to listen to brief excerpts, to rate songs, and to decide whether to download them. From those seven thousand visitors, Salganik and his coauthors could obtain a clear sense of what people liked best. The other seven thousand visitors were sorted into a “social influence” group, which was exactly the same except in just one respect: the social influence group could see how many times each song had been downloaded by other participants.
Those in the social influence group were also randomly assigned to one of eight subgroups, in which they could see only the number of downloads in their own subgroup. In those different subgroups, it was inevitable that different songs would attract different initial numbers of downloads as a result of unknown factors. “Trapped in an Orange Peel” might attract strong support from the first listeners in one subgroup, whereas it might attract no such support in another. “Separation Anxiety” might be unpopular in its first hours in one subgroup but attract a great deal of favorable attention in another.
The research questions were simple: would the initial numbers affect where songs would end up in terms of total number of downloads? Would the initial numbers affect the ultimate rankings of the forty-eight songs? Would the eight subgroups differ in those rankings? You might hypothesize that after a period, quality would always prevail — that in this relatively simple setting, where various extraneous factors (such as reviews) were highly unlikely to be at work, the popularity of the songs, as measured by their download rankings, would be roughly the same in the independent group and in all eight of the social influence groups.
It is a tempting hypothesis, but it is not at all what happened. “Trapped in an Orange Peel” could be a major hit or a miserable flop, depending on whether a lot of other people initially downloaded it and were seen to have done so. To a significant degree, everything turned on initial popularity. Almost any song could end up popular or not, depending on whether or not the first visitors liked it. Importantly, there is one qualification: the songs that did the very best in the independent judgment group rarely did very badly, and the songs that did the very worst in the independent judgment group rarely did spectacularly well. But otherwise almost anything could happen.
The apparent lesson is that success and failure in cultural markets is exceedingly hard to predict, whether we are speaking of AI or human beings. Here is one: it is exceedingly difficult to know, in advance, whether a cultural product will benefit from the equivalent of early downloads. True, knowing that is not unknowable in principle. If you knew everything about everything, you would know that, just as if you knew everything about everything, you would know whether a specific coin will come up heads or tails if I toss it in the air in the next second, or whether Carl and Eleanor are going to fall in love if they have lunch next Tuesday, or whether there is going to a revolution in a specific nation in February of next year. But it is not easy to know everything about everything.
Cute
What about business? What about products? Where do people want to travel? (Paris, Berlin, London, Copenhagen, Vienna, Prague, Beijing, Dublin, Amsterdam, Boston, New York?) Where do people want to study? What objects do people like or not like? With respect to products, an experiment modeled on the Music Lab found the same pattern. The experiment involved “Meet the Ganimals,” an online platform where people can generate and curate “ganimals,” which are AI-generated hybrid animals. People can also say how much they like particular ganimals and rate them in terms of cuteness, creepiness, realism, and other variables.
As in the Music Lab experiment, people were sorted into groups with independent conditions, in which they made evaluations entirely on their own, and groups with social influence conditions, in which they could see what other people thought. Just as in the Music Lab experiment, participants were randomly assigned to one of multiple online “worlds,” each of which evolved independently of the others. Participants saw only ganimals discovered and votes cast by others in their online world, and the ranking of ganimals was based only on votes in that world.
You might think that some ganimals really are adorable and that others really are not, and that in the end the adorable ones would be counted as adorable and the not-adorable ones would be counted as not-adorable. But here again, social influences greatly mattered. In the social influence worlds, outcomes turned out to be more unequal and highly unpredictable. Without social influences, different groups converged in their enthusiasm toward precisely the same set of ganimal features. (If you are curious: ganimals have eyes, a head, and dog-like features.) But with social influences, groups rapidly evolved into diverse local cultures that dramatically diverged from that in the independent judgment conditions. One ganimal could be spectacularly popular in one group and essentially unknown in another. The findings were very similar to those in the Music Lab.
Shall we draw a large lesson? Many markets have a lot in common with the market for ganimals. People aren’t going to think that a gruesome ganimal is adorable. If you have something with eyes, a head, and dog-like features, you might be golden. But maybe not. Diverse local cultures can arise, and a fabulous product might get attention in one of them, and no attention at all in another. Could AI predict which products will get attention in which cultures? Maybe so. But maybe not, if social interactions, based on an assortment of factors on which data cannot be obtained in advance, turn out to be crucial. The success of Barack Obama in 2008 and Donald Trump in 2016 depended on such factors, and the same is true for the rise of Jane Austen, The Beatles, Star Wars, and Taylor Swift.
Knightian Uncertainty
In 1921, the great American economist Frank Knight wrote: “Uncertainty must be taken in a sense radically distinct from the familiar notion of Risk, from which it has never been properly separated. . . . The essential fact is that ‘risk’ means in some cases a quantity susceptible of measurement, while at other times it is something distinctly not of this character; and there are far-reaching and crucial differences in the bearings of the phenomena depending on which of the two is really present and operating.” Knight was referring to what is now called “Knightian uncertainty”: circumstances in which probabilities cannot be assigned to future events. The social theorist Jon Elster offers an example: “One could certainly elicit from a political scientist the subjective probability that he attaches to the prediction that Norway in the year 3000 will be a democracy rather than a dictatorship, but would anyone even contemplate acting on the basis of this numerical magnitude?”
Regulators, ordinary people, and AI are sometimes acting in situations of Knightian uncertainty (where outcomes can be identified but no probabilities can be assigned) rather than risk (where outcomes can be identified and probabilities assigned to various outcomes). Some people appear to think that AI creates an uncertain risk of catastrophe, including the extinction of the human race. Consider in this regard a passage from John Maynard Keynes, also writing in 1921:
By “uncertain” knowledge, let me explain, I do not mean merely to distinguish what is known for certain from what is only probable. The game of roulette is not subject, in this sense, to uncertainty; nor is the prospect of a Victory bond being drawn. Or, again, the expectation of life is only slightly uncertain. Even the weather is only moderately uncertain. The sense in which I am using the term is that in which the prospect of a European war is uncertain, or the price of copper and the rate of interest twenty years hence, or the obsolescence of a new invention, or the position of private wealthowners in the social system in 1970. About these matters there is no scientific basis on which to form any calculable probability whatever. We simply do not know.
Sounding a lot like Knight, Keynes insisted that some of the time we cannot assign probabilities to imaginable outcomes. “We simply do not know.”
Keynes immediately added, however, with evident bemusement, that “the necessity for action and for decision compels us as practical men to do our best to overlook this awkward fact and to behave exactly as we should if we had behind us a good Benthamite calculation of a series of prospective advantages and disadvantages, each multiplied by its appropriate probability, waiting to be summed.” But how on earth, he wondered, do we manage to do that? Keynes listed three techniques — and they are worth considering when we leave behind the legends about AI and soberly assess its capabilities:
- We assume that the present is a much more serviceable guide to the future than a candid examination of past experience would show it to have been hitherto. In other words, we largely ignore the prospect of future changes about the actual character of which we know nothing.
- We assume that the existing state of opinion as expressed in prices and the character of existing output is based on a correct summing-up of future prospects, so that we can accept it as such unless and until something new and relevant comes into the picture.
- Knowing that our own individual judgment is worthless, we endeavor to fall back on the judgment of the rest of the world which is perhaps better informed. That is, we endeavor to conform with the behavior of the majority or the average. The psychology of a society of individuals each of whom is endeavoring to copy the others leads to what we may strictly term a conventional judgment.
Keynes did not mean to celebrate those techniques. Actually he thought that they were ridiculous. We might know, for example, that technological innovations have not produced horrific harm in the past, and so we might think that AI will not produce such harm in the future. Or, as a good Hayekian, AI might look at the price signal to assess the risks associated with climate change. Or AI might follow the wisdom of crowds to assess the likelihood of a pandemic. But under circumstances of uncertainty, should we trust any of these methods? “All these pretty, polite techniques, made for a well-paneled Board Room and a nicely regulated market, are liable to collapse,” Keynes declared, because “we know very little about the future.” Those last seven words may seem simple or obvious, but they are not at all obvious to the inventors of and investors in the allegedly predictive algorithmic technologies that are now overwhelming our society.
Keynes emphasized the difficulty or the impossibility of assigning probabilities to outcomes, but he also signaled more generally the problem of ignorance, in which we are unable to specify either the probability of bad outcomes or their nature — where we do not even know the kinds or magnitudes of the harms that we are facing. One reason for our ignorance might be that we are dealing with a novel, unique, or non-repeatable event. Another reason might be that we are dealing with a problem involving interacting components of a system in which we cannot know how components of the system are likely to interact with each other, which means that predictions are highly unreliable.
Back to the Future
There are some prediction problems on which AI will not do well, and the reason lies in an absence of adequate data, and in what we might wisely see as the intrinsic unpredictability of (some) human affairs. In some cases, AI will be able to make progress over time. But in important cases, in which we are dealing with complex phenomena, and the real problem is that the relevant data are simply not available in advance, accurate predictions are impossible — not now, and not in the future. Never.
Consider now, if you would, a heartbreaking remark that Benjamin Franklin’s sister Jane addressed to her brother, lamenting the “Thousands of Boyles Clarks and Newtons” who “have Probably been lost to the world, and lived and died in Ignorance and meanness, merely for want of being Placed in favourable Situations, and Injoying Proper Advantages.” Who is placed in unfavourable situations? Who faces disadvantages? Both of these take diverse forms. We might speak of an absence of education; Franklin herself was not allowed proper schooling. We might speak of an absence of economic opportunity. Or we might speak more specifically, and less systematically, about the absence of a mentor, a helping hand, a nod of appreciation, a glimpse of something wondrous, an infusion of money, a year off, a friend or family member who refuses to give up.
In the domain of innovation in general, social scientists, sounding a lot like the despondent Jane Franklin, refer to “Lost Einsteins” — in the words of one of them, those “who would have had highly impactful inventions had they been exposed to innovation in childhood.” The emphasis here is on demographic characteristics, such as race, gender, and socioeconomic status, and on the contributions of role models and network effects to success. Countless potential innovators, in science, business, and elsewhere, were subjugated in some way, were born in a particular family, did not find the right role models, or did not benefit from networks. As a result, they never innovated. They lost life’s lottery, or a series of smaller lotteries.
There are lost Da Vincis, lost Shakespeares, lost Miltons, lost Austens, lost Dickenses, lost Ellisons, lost Rothkos, lost Scorceses, lost Stan Lees, and lost Bob Dylans. There are lost Edisons and lost Doudnas and lost Teslas. (Nicolai, not the car.) There are plenty of them. They have been lost for a thousand and one different reasons. If innovators have been lost, it is not only because of demographic characteristics, but also because of a host of factors, not identifiable in advance, which did not work in their favor. Someone might not have given them a path, a smile at the right time, an infusion of energy, or a contract.
That conclusion might seem to point to a tragedy, even to countless tragedies — not only for those who have been lost, but also to those of us who have lost them, perhaps because they were never given an opportunity, perhaps because they were never given attention. In many ways, that is indeed tragic. But it also points to a possibility or perhaps even an inspiration. Lost Einsteins, or lost Shakespeares and Miltons, might be unlost, or found again. In fact they are being found every day. And if we can stay alert to the fact of their existence among us, many fewer will get lost in the first place. And that is a point about justice.
