Paul Bloom
What this covers
Paul Bloom and Tyler Cowen discuss the innateness of human moral and social capacities, tracing evidence from infant cognition to artificial intelligence. Bloom argues that much of what we treat as learned or culturally constructed—moral judgment, social preference, religious belief—emerges early and emotionally rather than through reasoned deliberation. The conversation ranges across psychology, philosophy, and technology, examining how babies evaluate others as good or bad through simple physical cues like helping versus hindering, how children sort themselves into groups by language before race, and how utilitarian calculus conflicts with visceral moral responses. Bloom contends that emotions like empathy and disgust lead systematically astray in moral reasoning, producing choices that diverge from what cold-blooded cost-benefit analysis would suggest.
The discussion turns contentious on several points. Bloom uses LLMs' success at language to challenge Chomsky's poverty-of-stimulus argument—that language structure is too complex to learn from input alone—claiming the evidence no longer makes this a logical necessity, though children may retain some innate apparatus. He explores us-versus-them psychology in children, religious belief as a system of conflicting mental modules rather than hypocrisy, and how cultural group selection explains monotheism's dominance. On AI, Bloom notes that people rate GPT-4o responses as more empathic than identical human ones—yet harbor prejudice against AI when told the source. He challenges common assumptions about psychopaths, prestige bias in expert judgment, and the overrated status of Jean Piaget's specific claims versus his organizing insight that children are explorers of the world.
Bloom argues that much of human moral and social cognition is innate and emotionally driven rather than rational, and that emotions like empathy and disgust are poor moral guides that we should consciously discount in favor of reasoned cost-benefit thinking.
- Babies show sophisticated social/moral understanding (helper/hinderer, us-vs-them) before culture could teach it
- Empathy and disgust produce systematically biased moral decisions that diverge from a utilitarian calculus
- LLM success at language and therapy challenges prior nativist assumptions about cognition
Psychologists possess no special ability to understand human nature or parent better than other educated people.
- Psychologists have no special ability to understand or manipulate people in day-to-day, on-the-spot situations; their training does not translate into superior real-time understanding of human nature compared to other smart, educated people.
“We have no special ability to manipulate people, to understand people. You could say that that reflects a weakness of the field, and maybe it does, but I don’t think psychologists are any savvier about human nature in that sort of day-to-day, on-the-spot way than anybody else.”
- Psychological expertise exists in specific subdomains (memory and its fallibility, mental disorders, the timing of child development, group behavior) but does not generalize into improved everyday interpersonal understanding.
“I think we understand certain subdomains. I think I have a better understanding of memory and the fallibility of memory than a lot of people. I think I have more of a grasp of mental disorders.”
- Psychologists tend to be good parents because the kind of person who becomes one is more sensitive to others' feelings, well-educated and affluent, not because of psychological expertise—no more so than an anthropologist or biochemist.
“I think we tend to be good parents because the sort of person who becomes a psychologist is probably more sensitive to other people’s feelings, well-educated, and tends to be affluent enough to provide kids with good stuff. But no, not any more so than an anthropologist or a biochemist.”
- The claim that psychologists have more neuroses than non-psychologists is mostly a joke, though it could hold to some degree in clinical psychology where people may enter the field to address their own problems, making cause and effect hard to distinguish.
“It could be true in some degree in clinical psychology, where people might go into the field in order to address issues and problems that they have. It’s hard to distinguish some issues of cause and effect here.”
- Measured intelligence may be largely built on a scaffolding of accumulated knowledge, such that the more knowledge a person has, the less raw general intelligence (G) they need to reach the same point, while someone with less knowledge needs more G.
“Maybe there’s a tradeoff, which is, the more knowledge you have, the less G you need to get to the same point, where somebody with less knowledge would need a lot of G to get to that point.”
Infants possess sophisticated innate understanding of physics, social categories, and moral judgment.
- The most misunderstood thing about children is that they are blank slates; in fact a large body of evidence shows infants possess a sophisticated understanding of the physical and social world from birth.
“Probably the most misunderstood thing, period, is that they’re blank slates. A lot of people walk around thinking that babies know nothing, and that’s not true. There’s a tremendous body of evidence now showing a quite sophisticated understanding of the physical and social world was there from the get-go.”
- Children divide the world into us versus them based on language at a very young age, before they reliably do so by race—American children prefer to affiliate with and give toys to English speakers over French speakers, and over English speakers with even a faint accent.
“Children at a very young age, as an American child, prefer to affiliate with an English speaker than a French speaker and also prefer to affiliate, to give toys to, to play with somebody who speaks English without an accent than someone who speaks English with a faint accent.”
- In one-act puppet experiments, nine-month-old babies shown a character helping another up a hill versus one shoving it down later prefer the helper, indicating a rudimentary moral evaluation capacity emerges very early.
“you show babies little one-act plays where you have this little thing trying to get up a hill, and one character helps it up the hill, and the other one shoves it down. Later on, when exposed to these two characters, the helper and the hinderer, babies prefer the helper.”
- Part of the human moral capacity is hardwired by natural selection, with infants using physical-act cues like helping versus hindering or caressing versus hitting to evaluate others as good or bad.
“the origin of this, I think, is natural selection. I think some part of our moral capacity is hardwired.”
- Jean Piaget is underrated: nearly every specific empirical claim he made about children has been found mistaken, yet his great encompassing theory—particularly the notion that children are like scientists exploring the world—has stood the test of time.
“He is wrong in just about everything he said about kids... but he was a great scientist with a great encompassing theory. A lot of his theoretical foundation, particularly the notion that children are like scientists exploring the world, has stood the test of time.”
Moral judgment arises from visceral responses to specific acts rather than aggregate harm calculation.
- There is a flattery shortage in the world: flattery is low-cost and effective even when the recipient knows it is researched and strategic, yet people undersupply it—likely because of awkwardness flattering someone of higher status when seeking a job—while a false theory of mind leads people to instead offer unhelpful blunt criticism.
“I fully knew — even my ego is not so . . . I fully knew he read up on me, but it still felt great.”
- We often delude ourselves into thinking our views are grounded in rational analysis when they are actually driven by emotions, as in Haidt's account—though Bloom maintains some views are still more rational than others, and a person reasoning 'what harm does it do?' is better than one driven by raw disgust.
“I’m sympathetic enough to Jonathan Haidt’s line that often we delude ourselves and we’re driven by these emotions.”
- Adult sibling incest, while not obviously wrong on adult-autonomy grounds, is problematic because the prospect that sex could become possible once they are adults corrodes the family relationship during childhood—a corrosion that is worse where power asymmetries exist, as between a stepfather and stepdaughter, and some valuable human relationships are better off remaining sex-free.
“the idea that sex is a possibility in the future when they’re adults corrodes the relationship early on.”
- Human morality is built on visceral responses to specific behaviors rather than a cold-blooded assessment of aggregate pain and pleasure; people demand more money to strangle a cat than to painfully pull out their own tooth, even though the same payment could be used to save many lives or many cats.
“We have a visceral response to certain behaviors. It doesn’t translate to a cold-blooded assessment of pain and pleasure.”
Empathy is unreliable moral guidance and people hold inherent non-egalitarian preferences they cannot transcend.
- Even if sociopaths can learn to behave according to incentives and fit in, what we want in the people we choose to be around is that they are nice to us even when no incentives are present—a quality sociopaths lack.
“the thing about the people we want to be around is, we want them to be nice to us even when the incentives aren’t present.”
- No human is truly egalitarian across all dimensions; to be indifferent between present and future, between one's own child and a stranger, and to disregard others' looks, intelligence, or kindness would mean ceasing to be a person—such a being is conceivable as an LLM program but we would find it repellent.
“To become indifferent between the present and the future, indifferent between your child and a stranger, to not care about somebody’s looks or intelligence or kindness would be . . . You could imagine an LLM program for that, but I think when we encountered the LLM, we’d find it repellent.”
- Empathy is a poor moral guide that leads to bad decisions; Bloom's argument against empathy is limited to its role as a guide to moral action, and he reports having made moral mistakes by being too caught up in empathy.
“My argument against empathy is limited. It argues that empathy is a poor moral guide. It leads to bad decisions.”
In-group bias, sex differences, and other negative tendencies are largely hardwired and resistant to parental elimination.
- Without cultural support a child raised in isolation (the 'Would Tarzan Believe in God' thought experiment) would become polytheistic, seeing invisible agents and gods triggered by nature everywhere, because children early on are biased to believe invisible agents cause things; coherent belief in a single God requires later cultural scaffolding.
“We came to the conclusion that he’d end up being polytheistic, that he’d see gods everywhere. He’d see them as triggered by nature and everything.”
- Some negative human tendencies—such as us-versus-them psychology and certain sex differences in play—are largely hardwired and present early, so parents should take them in stride rather than blaming themselves or their teaching, while recognizing innate does not mean unchangeable.
“you shouldn’t be too surprised when your child has an us-versus-them psychology. It’s very quick to split the world into the people that they care about and the rest of the people, and you should take that in stride. It’s not because you taught them anything bad.”
Cultural and psychological mechanisms genuinely shape human behavior and cognition.
- Mirror neurons were once hugely overrated as the solution to everything—explaining autism, language learning, and empathy—and the field crashed; they remain overrated because some people still don't know it crashed, though they may have a limited role in motor learning.
“it was once hugely overrated, in that this was the solution to everything. Explained autism, explained language learning, explained empathy. And the whole field, I think, crashed. It’s still overrated because some people don’t know it crashed.”
- Although polytheism is the cognitive default, monotheism has come to dominate because a single omniscient moralizing God serves moral and coordinating functions—binding communities and making people behave better even when unobserved—giving such societies an advantage through cultural group selection.
“If you believe in a single omniscient God that looks at you and looks at me, it could govern our behavior in certain ways, coordinate it, cause us to be nicer when we’re not being observed, and so on. The claim is that societies that have that have an advantage over societies that have a hundred mischievous little gods running around.”
- Personality psychology and the Big Five have proven surprisingly robust and stood the test of time, even though people often throw out all of psychology when complaining about the replication crisis; the idea of characterizing people by five trait numbers carries real sense, but should be held flexibly and not taken too seriously.
“Often, when people complain about the replication crisis, they throw away all of psychology, but personality psychology has proven surprisingly robust, and I think there’s a lot of sense in the idea.”
- Bloom's unusual work habit is to work in timed six-minute bursts—switching among writing, reference letters, laundry, email, and social media when a timer beeps even mid-word—which prevents boredom and leaves residual energy to resume each task, allowing hours of productive work.
“My oddest work habit is that I work — not always, but often — in six-minute bursts... by stopping mid-word or stopping wherever I’m stopping, when I go back to it, the energy is still there to want to continue.”
- The doctrine called wokeness leads to a lot of moral problems and makes the world worse in certain ways, so Bloom would not want 'woke' LLMs (citing Gemini), though he would not mind left-wing ones.
“Because I think the doctrine called wokeness leads to a lot of moral problems and makes the world worse in certain ways, but I wouldn’t mind left-wing LLMs.”
- Group selection is implausible as a genetic evolutionary mechanism (as Dawkins argued, it could happen in principle but does not in practice), yet it is a reasonable framework when applied to cultures—some cultures do better and thrive—where the unit is communities of people rather than genes.
“Very critical of it as an evolutionary mechanism. I’ve been convinced, as Dawkins viewed it, that it doesn’t happen in practice. But it seems like a perfectly reasonable thing to talk in terms of cultures, to say that some cultures do better than others, those cultures thrive.”
- Honesty (along with disagreeableness) is the trait hardest for us to value properly, because being truthful with someone who asks about their book, fitness for a job, or relationship can do them a huge service, yet the reward you receive is hatred and resentment.
“To be honest with somebody when they say to you, “What do you think of my book?” “Do you think I’m suited for this job?” Stuff like that can be doing them a huge service... and the reward you will get is hatred and resentment.”
- Sociopaths and psychopaths are overrated and romanticized: many studies find they are often below normal intelligence, tend to be depressed, more likely to kill themselves, and live lousy lives, contradicting the glamorous conscience-free image—suggesting non-sociopaths are sensitive to and avoid them.
“Many studies of psychopaths and sociopaths find that they’re actually often below normal intelligence. They tend to be depressed. They’re more likely to kill themselves. They live lousy lives.”
- Education raises measured IQ a bit each year; while skeptics like Bryan Caplan attribute this to getting better at taking tests and scholastic thinking, education genuinely makes people smarter by providing practice with abstractions and hypotheticals—the very things used as markers of intelligence.
“I do think, though, that education, period, does play a role in making you smarter. It gives you practice in dealing with abstractions, with hypotheticals, the sorts of things that we often use as markers of intelligence.”
- Bloom decides which book to write by drafting a proposal for his agent: if he can write a good proposal the material is there, and ultimately he knows a book is right for him if he couldn't live without writing it—he would read it himself even if nobody else wanted it.
“by the time I’m done with that, I know it’ll be a good book for me if I couldn’t live without writing it. If nobody wanted it, I read it anyway.”
Parenting and development are shaped by material conditions, culture, and community context as much as individual choice.
- In wealthy societies a young baby's life is probably difficult because needs are not immediately satisfied and they experience frequent pain, fear and hunger; well-being rises markedly around age two when the child gains more autonomy.
“I would think that roughly around the age of two, you get much more autonomy, and then life starts looking up, and then it just gets better and better.”
- In wealthy countries babies often sleep apart in a separate room and are left to cry it out, which is hard on the child; in less wealthy countries co-sleeping in the same room is better for the child even if rougher on the parents.
“In less wealthy countries, the kid co-sleeps and sleeps in the same room. I think that’s better off for the kid. Rough on the parents, but better for the kid.”
- Asian tiger parenting is underrated: although it doesn't work for every kid or parent, many tiger parents end up with children who are not merely successful but happy and love their parents, so the approach should be taken more seriously even by permissive parents.
“a lot of these tiger parents end up with kids who are not merely successful, but happy and love their parents. It doesn’t work for everybody... but I think we should take it more seriously.”
- Parenting approaches—like tiger parenting, or rules around cell phones, smartphones and social media—are extremely difficult to sustain as an individual within a community doing the opposite; it is much easier to be any kind of parent when everyone around you parents the same way.
“if you’re one individual within the community doing it, it’s extremely difficult. It’s much easier to be any parent if everybody else around you is the same sort of parent.”
Emotions like disgust and empathy drive necessary action even when rational analysis supports different policy.
- Even a decision grounded in rational cost-benefit analysis may require an emotion like disgust or empathy to motivate ordinary people to act on it—so disgust mixed with other views could be what motivates necessary limits on otherwise-beneficial policies like immigration.
“You might want to argue that without disgust, Canadians wouldn’t be motivated to a policy that, on balance, would be better for them.”
- Disgust in the moral realm is the one emotion Bloom would remove from people if he could, because it drives hatred toward gay people, opposition to interracial marriage, and exclusion of immigrants and other races; he would build LLMs with anger, pity and gratitude but keep moral disgust out, though he is wary of removing any emotion given uncertain consequences.
“If there’s one emotion I would take away from people, it would be disgust, at least disgust in the moral realm. They could keep their disgust towards rotten food and that sort of thing. That’s the one thing I wouldn’t put into LLMs.”
- Overt displays of empathy by politicians (like reading a letter from an eight-year-old to sell a healthcare plan) should be discouraged through social disapproval the same way overt racism is now booed and penalized, because empathy is easily used as a manipulative tool and crowds should refuse to be treated like children.
“I would like to see that when a politician comes up and says, “I want to talk about the healthcare plan,” and says, “I want to read you a letter I got from an eight-year-old who told me the story about her father,” I want the crowd to boo and to shout, “Don’t treat us like children.””
- While Bloom refuses to discount his favoritism toward his own children, he holds that for the treatment of animals we should try to talk ourselves out of our visceral reactions and reason our way to making the world better, even if it involves uncomfortable acts.
“For animals and treatment of animals and cat strangling, I think we should try to talk ourselves out of this. I think we should, in some way, try to reason to make the world a better place even if it involves strangling some cats.”
People hold multiple conflicting mental systems that produce contradictory beliefs simultaneously.
- Religious people cry at loved ones' deaths despite believing in heaven not because they are hypocrites but because humans are not unitary beings—a confident or even 100 percent belief in heaven coexists with other mental systems that produce terrible grief and fear of death, and these conflicting beliefs both genuinely exist.
“We’re not unitary beings. It’s possible, upon learning my son has a terminal disease, for part of me to think, “Oh, he’ll be reunited with God in heaven... and feel horrible grief. Yes, they conflict with one another, but that doesn’t negate the existence of either one.”
- The Hebrew Bible is a great text in social, personality, and even cognitive psychology, starkly capturing us-them biases, prejudice, and in-group/out-group thinking, and portraying God as a psychologically complex character—for example fiddling with the Pharaoh's mind to make him refuse, to show off divine power.
“It captures in a very stark way the us-them biases that are present, that social psychology makes so much out of. It captures the idea of prejudice.”
Cognitive biases like speciesism and prestige bias become entrenched and shape irrational preferences over time.
- Research with Matti Wilks shows children are less speciesist than adults—they will favor ten pigs over one human, a trade adults in the sample never make—and speciesism is an irrational prejudice that creeps in gradually with age until adults treat a human as vastly more valuable.
“kids are less speciesist than adults. They even will favor 10 pigs over one human. Adults in our sample never do that.”
- People overestimate the intelligence of someone with prestige markers—being a Yale professor, an older white man—taking him seriously even on domains outside his expertise, illustrating prestige bias in judgments of competence.
“Prestige bias. I’m a university professor. I’m at Yale. Maybe I benefit from other forms of privilege, being an older white man, but I don’t know. I show all the markers of a smart guy.”
Large language models expose flaws in classical language theories and require different reasoning than human minds.
- Effective use of current LLMs requires a distinct theory of mind from the one used for people, plus flexibility and patience, because LLMs will hallucinate answers (e.g., fabricating papers) to give you what you signal you want, so you must 'game' them—for instance by stating that false information will cost you your job to elicit accuracy.
“They said, “Yes, it is. Here are three papers that show it decisively.” None of the papers were real. It hallucinated it because it’s what I wanted.”
- The success of large language models at learning language through statistical analysis of input shows the Chomskyan claim is wrong—that language structures are too complex and underspecified in input to be learned without innate machinery; it is no longer a logical truth that language requires a language instinct, though children may still use some innate apparatus.
“If it’s true — and I think it is true — that the LLMs learn language through statistical analysis, this shows the Chomskyan view is wrong. This shows that, at least in theory, it’s possible to learn languages just by observing a billion tokens of language.”
People prefer human contact and enjoy fiction with negative emotions over equivalently empathic AI.
- The cognitive dissonance effect—where choosing one of two options makes you like the chosen thing more and the unchosen less—appears in Capuchin monkeys (in work with Laurie Santos) and in three-year-olds, and choice-based preference change is a textbook example of dissonance even if the endowment effect has alternative explanations.
“In some work with Laurie Santos, who had a thriving monkey lab, we found that Capuchins did the same thing. We also found three-year-olds did the same thing.”
- People enjoy horror and negative emotions in fiction partly because the mind is drawn to worst-case scenarios as safe practice—analogous to a flight simulator programmed for difficult situations to rehearse coping with them—though there is wide individual variation in which negative emotions and intensities people prefer.
“it might reflect what’s sometimes been called safe practice, where our minds are drawn to worst-case scenarios as a way to practice encountering them. An analogy I like is that of a flight simulator.”
- In experiments pitting GPT-4o against people in short-term interactions where someone shares troubles, people repeatedly rate the AI's responses as more empathic and understanding than the human's, though it remains unproven whether AI can sustain a long therapeutic relationship holding memory across sessions.
“Over and over and over again, people rate the AI responses more empathic, more understanding.”
- People hold a prejudice against AI: when given the exact same response, those told it came from an AI rated it as less empathic and understanding than those told it came from a person.
“There was a paper recently that gave the same sort of response, the exact same response to two groups of people. Told one of them it’s a person, the other one it’s an AI. The people told it was an AI said, “Oh, that wasn’t very empathic.””
- Santa Claus is an easier belief for children to acquire (around age two or three) than God, because Santa is just a capable, kind big man who climbs down chimneys, whereas God is invisible, omniscient and all-good—a much bigger conceptual climb.
“I think in some ways, Santa Claus is an easier climb because he’s just a big guy who looks like a big beard... God, in the sense that we’re talking about it, is invisible and omniscient and all-good. So, I think God is a much bigger climb. I’ll say Santa Claus at two or three.”
Theory of mind is best tested through sustained unpredictable conversation rather than controlled tasks.
- The best test of whether something has a theory of mind is sustained conversation rather than technical tests like the false-belief task or Reading the Mind in the Eyes, because those are easily gamed by systems trained on thousands of instances; a good test is showing it half a novel or TV episode and asking what happens next, which can't be cheated on novel material.
“There are technical tests, like the false-belief task or Reading of the Mind in the Eyes task, but those are easily gamed. LLMs will succeed on them because they’ve seen thousands and thousands of instances on them.”
- It may be possible to train people to better understand others or predict the future—Tetlock's super-forecasters suggest you can either weed out and find the most naturally gifted people or train specific techniques—though this takes us far from psychology as normally studied.
“Phil Tetlock talks about super-forecasters and talks about, how do you get people who could better predict the future? You could imagine training people, either by weeding them out... or training them in certain techniques”
- Within not too many years, regular conversations with an LLM about one's problems will outnumber human therapeutic interactions, because there is currently a prejudice against AI responses that will fade over time—especially among younger people aged 18 to 20 versus older adults.
“I think that human interaction within not too many years will be the minority of interactions. I think people will be very drawn to AI therapists.”