What this covers

Sean Carroll talks with Colin Camerer, a Caltech behavioral economist and pioneer of neuroeconomics, about how people actually choose in strategic and intertemporal settings. The conversation moves between lab games, brain measurement, and field data, and keeps returning to the question of what a model must do to predict behavior rather than just label deviations from it.

The game material is concrete. In the two-thirds-of-the-average game, choices cluster near 33 and 22 rather than the equilibrium of zero, and even groups of game theorists pick fives and tens. Chess players do no better, which Camerer reads as evidence that strategic depth is domain-specific. One participant picked 18 because they were 18. The 11-20 game and the repeated Prisoner's Dilemma (via the folk theorem) extend the picture. Field examples include movies released without critic screenings, which scored lower yet earned about 15 percent more, and survival probabilities that fail to sum to 100 percent. Camerer also disputes Kahneman's lists of System 1 and System 2 traits as pre-scientific.

The neural half covers the mentalizing circuit, including the temporoparietal junction, and how a "how" versus "why" picture contrast isolates it. It also covers striatal habit systems in rodents and delay discounting. The more contested claims involve intervention: optogenetic "neural gaslighting" in rats, brain stimulation that made people somewhat more generous, and whispered food names during sleep raising valuations by about 10 percent, which Camerer calls preliminary. A late detour asks whether gym attendance and queue waiting times follow a power law. Camerer reports they do not quite, and that a log-normal fits somewhat better.

Sharpest takeaway

Camerer argues that the rational-utility-maximizer model is only a benchmark, and that economics can be made predictive by modeling bounded computation, willpower, and social reasoning with experiments and neuroscience, which also suggests some of these limits may be modifiable.

  • Utility maximization is a limiting case that holds for some experts but is not an accurate description of typical decision-makers
  • Lab games (two-thirds of the average, 11-20) show people reason only one or two levels deep, and this depth is domain-specific
  • Brain measurements such as mentalizing circuits, habit systems, and delay-discounting regions show mechanisms behind these limits, and some can be stimulated to change behavior

The argument · threads11 threads · 64 claims
0.85

People's strategic reasoning is shallow, domain-specific, and driven by learned schemas, not general ability.

12 pointscentrality 5/5
  • Chess players do not reason more strategically than ordinary people in the beauty-contest game, indicating that depth of strategic thinking is domain-specific with very limited transfer of learning across domains.

    “Depth of thinking seems to be somewhat domain specific, which is true of a lot of psychological things. There's very limited what we call transfer of learning”

  • People can understand deeply nested statements about what others know, but Camerer argues this apparent high-level reasoning comes mostly from learned schemas and familiar scripts rather than general strategic depth, so very high levels are rare and specialized.

    “I think the reason people might seem to be doing a lot of reasoning is because of just learning and familiarity or what cognitive scientists call schemas.”

  • In the 11-20 game, players earn their number but get a bonus for undercutting the other by one; this produces a cycle of reasoning with a focal level-zero choice of 20, and the typical depth found is two or three levels.

    “in 11-20, you pick an integer from 11 to 20. And what's neat about it is you earn whatever number you picked.”

  • Caltech students, who have very high math SAT scores and are selected on verbal scores in a homogeneous population, showed about half a level of extra strategic thinking compared with Pasadena City College students a mile away, a smaller gap than the researchers expected.

    “these were about one-half level extra thinking compared to the PCC, which was a little less than we thought”

  • The shallow choices in beauty-contest games could reflect either limited depth of reasoning or a belief that others are irrational, and Camerer argues that different games and by changing the opponents can separate the two explanations; experts choosing among themselves would still pick 22 or 15, not 0 or 1.

    “a lot of the game theorists who would choose zeros or ones among themselves will choose 22 or 15 or some number”

  • Scores on multi-game strategic reasoning tasks, analyzed in the psychometric way with several games, correlate with working memory capacity and education, and Caltech students reasoned about half a level deeper than Pasadena City College students.

    “those are correlated with working memory, like how many digits can people remember, with education.”

  • Even groups of game theorists rarely pick 0 or 1 when playing the two-thirds game; they pick fives and tens, because they expect other experts to also be sophisticated and share common knowledge of that expertise.

    “unless you do it with an audience of game theorists, and even then, you don't get that many ones and zeros. You'll get fives and tens.”

  • Cognitive reflection test results show MIT students averaging about 2.28 out of 3, Claremont around 1.6, and Harvard slightly below a sample of people on Boston Common, though the Boston Common sample may not be representative of Bostonians.

    “The MIT students, I think, were 2.28 or something like that. We did some at Claremont, they were around maybe 1.6.”

  • In the two-thirds-of-the-average game, where players pick numbers from 0 to 100 and the target is two-thirds of the average, typical choices cluster at one or two steps of reasoning, such as 33 and 22, with few choices at the Nash equilibrium of zero.

    “what typically you see is more 33s than 22s and not that many two-thirds of 22s, which is around 15. So it kind of peters out.”

  • On the cognitive reflection test's algae problem, where algae doubles daily and covers a pond in 48 days, the intuitive answer of day 24 is wrong because the correct answer is day 47, showing a snap judgment answering an almost-right question.

    “And the right answer is 47, because if it doubles every day, it has to be half of full one day before 48”

  • Some players do not reason strategically at all; in Rosemarie Nagel's data, one participant picked 18 because they were 18 years old, which is a level-zero choice rather than three steps of reasoning, and this was only discovered by asking participants why they chose.

    “I picked 18 because I'm 18 years old.”

  • More strategic thinking may have evolved through a mental arms race in which each individual tries to detect deception and avoid being fooled, but this pressure is strong only in very social species, so most species would not evolve deeper reasoning.

    “Probably it's something like there's a bit of a kind of mental arms race where if I could do more thinking, I can figure out if you're fooling me and I can fool you.”

0.56

Bounded computation, not perfect utility maximization, better describes how people actually behave and decide.

6 pointscentrality 4/5
  • Perfect utility maximization, which requires complete rankings of all goods and bundles, is a limiting case that some highly experienced people such as wine experts approximate, not a description of how people generally behave.

    “I think it's a limiting case, like you might think about in physical science”

  • Economics teaching and journals treat utility maximization as the default starting point, even though it is not believed to describe everyone at all ages and experience levels, so it leaves much out.

    “I think of the model of utility maximization as a benchmark of rationality as sort of a starting point, like maybe Newtonian mechanics”

  • When asked separately how likely they are to live to 75 and how likely they are to be dead by 75, people give probabilities that do not sum to 100 percent, because the wording shifts attention and memory so that people do not retrieve a stable internal number.

    “How likely do you think you'll live until 75?”

  • Asked how rational people are, Camerer says about 75 percent, while acknowledging that the technically correct answer is 'it depends' because rationality varies with context and boundary cases such as wine experts show near-perfect consistency.

    “I think I said 75% rational.”

  • The research program aims to incorporate natural limits on computation, willpower, and self-interest to improve economic predictions and policy, and Camerer adds 'natural' to signal that biological mechanisms and neuroscience belong in the analysis alongside psychology and sociology.

    “we're trying to incorporate natural limits on computation, willpower, and self-interest to improve economic predictions and make better policy recommendations”

  • Early behavioral economists faced critics who used single anecdotes to disprove rationality, but the productive scientific question is what the replacement theory is, and the replacement has become a set of boutique theories built around computational limits, such as cognitive hierarchy and level-k modeling.

    “Well, my cousin bought the wrong insurance plan. Therefore, we've disproved everyone being rational all the time. Well, like, great. What's the replacement theory?”

0.41

Habitual behavior follows lawful power-law patterns, explained by simple, testable autopilot mechanisms.

11 pointscentrality 2/5
  • Using a statistical test designed by Gabaix that asks whether a power law can be rejected without a specific alternative, Camerer and colleagues found that the data are not quite a power law and that a log-normal distribution fits somewhat better, especially with fewer very short gaps and a cutoff in long gaps.

    “And indeed, they're not quite power law.”

  • Gym attendance records from about 20,000 people over several years show bursty patterns with many short and many long gaps between visits, which is inconsistent with a homogeneous Poisson arrival process that would produce a single rate parameter.

    “we have some data on about 1,500 days of gym attendance. These are people who joined the gym”

  • Plotting the frequency of gym-attendance gaps against gap length in logarithmic form produced a near-straight line with R-squared of about 0.97, a power-law relationship that suggests a simple generating process.

    “It was an R-squared of 0.97.”

  • Standard first-in-first-out queuing theory produces exponential-type waiting times, but if some customers are served out of order with a fixed probability, the queue produces power laws whose exponents depend on the priority rule, such as minus 1.5 for Cobham's 1954 form and minus 1 for Barabási's form.

    “if you tweak queuing theory to what's called priority queuing, you can get a power law.”

  • Action slips, such as driving to a usual destination when a different one was intended, or a flight attendant forgetting a cookie seconds after serving ice, show that habitual sequences run on autopilot without conscious monitoring.

    “So one bit of evidence for habits is what are called action slips.”

  • Bursty behavior shows fewer very short gaps than a power law predicts because of recharging refractory periods, such as satiety after eating or muscle soreness after exercise, which imposes a biological minimum between repeated actions.

    “after you eat and you're full, you don't eat again. Sex, exercise, sleep, you know, all these things, you don't expect to see a lot of really small things close to each other”

  • In the Neural Autopilot model, people toggle between optimization mode, which compares prices and trade-offs, and habit mode, which retrieves the last choice and a single reward reliability signal based on absolute reward prediction error, so repeated choices require little executive function and allow multitasking.

    “In habit mode, you do something much simpler. What you do is you retrieve a memory, what did I pick last time? And then you pick up one number, which is the reward reliability”

  • The Neural Autopilot model generates streaks of consecutive days alongside long gaps and produces power-law slopes near minus two, close to the empirical gym data, though the model is less mathematically elegant than existing statistical-physics models.

    “it generates a slope of the power laws. The slopes in the power laws we saw in our data were mostly like minus two.”

  • When a behavioral pattern is empirically this lawful, there must be a simple underlying process, though in practice several different simple mechanisms can generate the same power-law pattern.

    “if it's that lawful empirically, there must be a simple explanation.”

  • An R-squared near 0.97 is rarely seen in social science, typically only when regressing a variable against itself or against a highly persistent series such as GDP, so such a fit prompts the search for a simple underlying process.

    “in social science, you only get that when you're regressing by mistake one thing against the same thing, or something that's really persistent like GDP.”

  • Habit and deliberative control are well-established as dissociable neural systems that are part of broader circuits connecting to motor systems, which is why simple habit models can be tested against neural data.

    “I think the idea that the, that habit and more deliberative choice are fundamentally dissociated is well established.”

0.40

Distributed brain circuits, especially mentalizing, underlie strategic thinking and prosocial choice.

5 pointscentrality 3/5
  • Very few brain functions are localized to one region, so neuroscience studies of economic decisions look for circuits, although some functions such as language are left-lateralized and the brain's plasticity lets function reorganize after damage, especially early in life.

    “very few things are highly localized to one part of the brain. Right? So we do think about certain regions, like amygdala or DLPFC I mentioned, but typically we're looking for kind of a circuit.”

  • Using transcranial direct current stimulation over right temporoparietal junction, Alex Smith and Sheryl Ball found that participants became somewhat more generous to charity and more consistent in trade-off choices, suggesting that mentalizing-related circuits can make choices more rational and prosocial.

    “when you upregulate right temporal parietal junction, they're a little bit more generous and a little bit more rational. They're more consistent.”

  • Researchers isolate the mentalizing circuit by showing participants 'how' versus 'why' pictures matched for visual content and button presses, then subtracting 'how' activity from 'why' activity, which reveals right temporoparietal junction and dorsomedial prefrontal cortex as key regions.

    “And then you subtract the how activity from the why, and what's left over we think is mentalizing.”

  • Goal-directed, deliberative action in rodents depends on dorsomedial striatum, while habitual action depends on dorsolateral striatum, and lesioning these regions can remove habits or make goal-directed actions habitual, showing that the two systems can be selectively turned on and off.

    “goal-directed... is in dorsomedial striatum, habit in dorsolateral striatum”

  • In one study of the two-thirds game, participants who did two steps of strategic thinking showed more activity in the mentalizing circuit, suggesting that extra reasoning levels recruit the same network that tracks others' mental states, though Camerer calls for many more studies.

    “people who do two steps of thinking in two-thirds of the average have more activity in mentalizing circuitry.”

0.38

Neuroscience experiments can directly test economic mechanisms like reward learning and deliberative choice.

4 pointscentrality 3/5
  • By optogenetically upregulating dopamine reward prediction error neurons while rats press a lever that delivers salty water, Camerer's team made the animals behave as if salty water were rewarding, which he calls neural gaslighting, though the result is a proof of concept and not a knockout study.

    “you optogenetically upregulate the reward prediction error neurons. I call this neural gaslighting.”

  • Camerer's transition from economics to neuroscience was driven by his preference for novelty and doing things others do not, combined with tenure that allowed risk-taking on unpublishable work, and Caltech subsidized a human scanner that enabled the work.

    “I don't care if I can't publish these papers. I have tenure. That's what tenure's for.”

  • When dieting participants choose healthy foods over tempting ones, dorsolateral prefrontal cortex activity increases, suggesting that this region represents future health consequences, though the findings may change with GLP-1 drugs.

    “when they're able to resist temptation, there's activity in DLPFC.”

  • A likely explanation for why people do not reason more levels deep is energy cost, since the brain runs on about 40 watts and evolution favors approximations and offloading, though the fossil record and evolutionary constraints remain open questions.

    “Energy usage is always a good candidate, right? You know, because the brain runs on 40 watts”

0.38

Dual-system labels must give way to brain-defined neural systems rather than verbal trait lists.

4 pointscentrality 3/5
  • When Kahneman presented the System 1/System 2 idea to neuroscientists, they were not receptive, partly because they do not believe in that much reductionism and felt two systems was too small a number compared to modular proposals like Cosmides and Tooby's 78 modules.

    “And part of it was neuroscientists don't believe in that much reductionism.”

  • If the brain truly split into two systems, it would be a striking finding, and System 1 might be present in infants while System 2 emerges around puberty or age 25, though the notion remains untested.

    “if it is two, that would be wild. And it may well be, for example, that System 1 is available in infants and System 2 shows up in puberty or when you're 25”

  • Kahneman's list of System 1 traits (associative, fast, effortless) and System 2 traits (serial, slow, deliberative) is a pre-scientific description, and a neural system should be defined by brain evidence rather than by a list of words.

    “We don't want a neural system to be defined by a list of words. Right? We want to see it in the brain.”

  • Camerer prefers 'mentalizing' to 'theory of mind' because it describes the process of constructing a representation of another person's mind, and the neural mentalizing circuit centers on right temporoparietal junction, dorsomedial prefrontal cortex, and posterior cingulate.

    “So when it comes to strategic thinking, the terms that I prefer is mentalizing.”

0.38

Economics must drop denial and status anxiety, embracing experiments and interdisciplinary learning.

4 pointscentrality 3/5
  • Samuelson and Nordhaus's widely taught textbook said economics cannot do experiments like astronomy or geology, even though hundreds of experiments had already been published, which Camerer describes as a kind of collective denial.

    “Sadly, like astronomy and geology, we can't do experiments.”

  • In economics there is a proud tradition, with famous exceptions, that treats saying 'I don't really understand this' as low status, which led economists at a 1997 neuroeconomics meeting to listen silently rather than ask questions.

    “in economics, there's a little bit of a proud tradition, with some famous exceptions, of it's considered low status to say, 'I don't really understand this.'”

  • Scientific disputes with critical referees are productive, and the best way to answer skeptics is to run another experiment or regression, such as testing the same economic games with larger stakes or in a poorer country where the amounts are equivalent to days of wages.

    “as Thaler likes to say, the easiest way to win an argument is to run another experiment or run another regression.”

  • Successful interdisciplinary collaboration requires top experts who are willing to learn each other's language and toggle from expert to student, because words like 'rational' carry seven different meanings across fields.

    “You have to be in kind of student mode. Like you have to toggle from expert to student.”

0.35

Framing and missing signals mislead people, and markets exploit these predictable reasoning failures.

3 pointscentrality 3/5
  • Tversky and Kahneman coined framing effects to describe cases where logically identical descriptions produce different choices, such as a salad dressing sells far fewer units when labeled 10 percent fat than when labeled 90 percent fat-free, and advertisers and political consultants exploit this.

    “If you call a salad dressing 10% fat, you sell a lot fewer than if it's 90% fat-free.”

  • Studios use analytics and test screenings to forecast box office performance days before opening and sometimes cold-open the films they expect to do badly, so the public does not see reviews of films they expect to flop.

    “often a couple of days or a week before opening, you'll hear, 'We expect it to open this much.'”

  • Of about 600 wide-release US movies from 2000 to 2009, roughly 10 percent were cold-opened, meaning not screened for critics; these films later received much lower Metacritic scores yet earned about 15 percent more box office, because some moviegoers fail to infer that a missing review signals a bad film.

    “about 10% of movies were cold open, which means they were not shown to critics”

0.24

Patience and discounting are measurable and can be shifted by credibility or vivid future events.

4 pointscentrality 2/5
  • People may discount future rewards partly because they doubt they will be paid, and researchers addressed this by giving post-dated checks or by paying in early sessions so participants saw the payments were real, which builds credibility.

    “this thing about credibility of future rewards a lot of people have tackled”

  • When a delayed reward is tied to a vivid future event, such as appearing on a podcast in a month, people become more patient, because the imagined event pulls the future forward mentally and changes the effective discount rate, though a credibility explanation remains possible.

    “they said, 'You can have $10 now or $15 when you're going to be in Sean Carroll's podcast in a month.'”

  • Therapeutic brain methods such as deep brain stimulation for Parkinson's and transcranial magnetic stimulation for depression could eventually be applied to improve patience and self-control in economic decisions, though this requires moving from one-hour proof-of-concept studies to durable interventions, possibly in early childhood.

    “I think there's a good chance if we worked on it, we could get people to be more patient.”

  • Delay discounting, measured with smaller-sooner versus larger-later choices calibrated by titration, produces a discount factor that is exactly analogous to an interest rate, and psychology, animal learning, biology, and economics now share standard tools for measuring it.

    “So we call it a discount factor, but it's exactly like an interest rate.”

0.24

Game theory adds strategic modeling and repeated-game cooperation, unifying social science under one framework.

4 pointscentrality 2/5
  • What game theory adds to earlier economics such as Smith, Ricardo, and Marx is the ability to model other agents who are acting strategically alongside you, which changes outcomes when there are only a few players.

    “it sounds like what game theory adds to that is the ability to model the other agents acting with you.”

  • Game theory is arguably the closest thing social science has to a unifying theory, because it applies to people, genes, organisms, and nation states, and it became central to economics even though much of its foundation came from applied mathematics.

    “I think it's in a way it's the closest thing we have to a unifying theory of social science”

  • Chad Jones's study of how much to spend on limiting existential AI risk uses the standard economist approach of multiplying a value of life by a probability, and Camerer argues that this is a reasonable starting point with sensitivity analysis as the key check.

    “he does a very typical economist study, which is to think of a value of a life and this probability”

  • In a one-shot Prisoner's Dilemma people tend to defect, but when the game is repeated and players know the history and expect future periods, as with two companies in a small town, cooperation can emerge that would not arise in a single round, as formalized by the folk theorem.

    “if it's a repeated game and people know the history and know that there's more periods coming, like two companies in a small town, then the results really change”

0.10

Neuroscience evidence for economics is thin and preliminary, needing skilled cross-field researchers.

4 pointscentrality 1/5
  • Caltech biology has few researchers who study humans because human systems are too complicated, so they use model organisms such as rodents for nicotine addiction and zebrafish larvae for sleep and development, which is why the human brain scanner was underused.

    “very few people are interested in humans. Because we're too complicated a biological system.”

  • Paul Bloom is right that interdisciplinary conferences fail when they invite the top people in each field rather than people in one field who care about the other, and productive participants must be good at their field and willing to toggle from expert to student and learn each other's language.

    “you need people who are good at what they do, like really good, willing to learn language and communicate across languages.”

  • Cognitive neuroscience provides only a thin web of relevant evidence for economic questions such as self-control, and the field is not yet a dense, well-developed framework, so most of the evidence is preliminary.

    “it's just a very thin web of tissue. It's not a nice, dense plan.”

  • In Lucia Zhu's sleep study, whispering food names during sleep, at timings identified by sleep equipment, increased participants' valuations of those foods by about 10 percent relative to unwhispered foods, though the effect is preliminary and Camerer does not regard it as a marketing technique.

    “they value them about 10% more.”