Paul Bloom on the Psychology of Children, and the Morality of Empathy and Disgust

Paul Bloom — psychologist at the University of Toronto and Yale, author of Against Empathy — joins Tyler Cowen in Ep. 218 to argue that empathy is a poor moral guide, that disgust has done more harm than good, and that the moral capacities visible in nine-month-old infants reveal something hard-wired about the human mind.

Key ideas

1. Babies are not blank slates — their moral intuitions appear at nine months. Experiments showing infants preferring a character who helps another climb a hill over one who shoves it down demonstrate that the capacity to distinguish helpers from hinderers is present before language, culture, or formal teaching. Some of what we call moral reasoning is there from the start.
2. Empathy is a poor guide to ethical decision-making. Bloom’s core argument — the thesis of Against Empathy — is that empathy (feeling what another person feels) is parochial, innumerate, and easily manipulated by politicians. It causes us to favour the vivid individual over statistics representing many. Compassion — caring without feeling as — is a better moral tool.
3. Disgust is the one emotion Bloom would remove from public life. He identifies it as the driver of hostility towards interracial marriage, gay people, and immigrants; its evolutionary function in avoiding contaminated food does not redeem its moral record. He would rather install anger, pity, or gratitude in any moral-emotions toolkit, including in large language models.
4. LLMs have overturned the case for an innate language module. Bloom describes the success of language models trained purely on statistical patterns as the biggest surprise of his intellectual life. The Chomskyan claim that grammar is too complex to be learned from input alone — that a ‘language instinct’ must be hard-wired — now looks empirically open rather than logically necessary. (A language module here means a dedicated mental organ, like a specialised processor, rather than a general-purpose learner.)
5. The undersupply of flattery reveals a theory-of-mind failure. Bloom finds it puzzling that interview candidates almost never research and praise the person they are meeting, given how cheaply it can be done and how reliably it works. The failure, he argues, is a kind of social miscalculation: candidates do not model the interviewer’s experience accurately enough to realise the returns on a well-placed compliment.

Content

What psychologists actually know

Cowen opens with a pointed question: if two smart non-psychologists and two psychologists were to overhear a couple arguing in a restaurant, would the psychologists understand what was happening any better? Bloom says no — not in the slightest. Psychology’s research programme produces knowledge about memory, mental disorder, child development, and group behaviour that does not translate into day-to-day interpersonal insight. He is equally deflating about psychologists as parents: any advantage comes from being educated and financially comfortable, not from professional expertise.

The distinction he draws is between subdomains where cumulative research has settled something (memory fallibility, developmental timelines, social cognition) and real-time social perception, where training has produced no detectable edge. When Cowen presses him on whether psychologists could be trained to do better, Bloom invokes Philip Tetlock’s work on super-forecasters — people who make more accurate predictions by adopting systematic habits — as evidence that human perception is trainable in principle, even if psychology as practised has not done it.

Babies, morality, and the emergence of us-versus-them

The earliest evidence for hard-wired moral capacities comes from preferential-looking experiments — studies where researchers measure which characters a baby reaches towards or stares at longer. In studies Bloom helped run at Yale, nine-month-old infants exposed to a ‘hill scenario’ (a figure struggling uphill, one character helping it, another blocking it) consistently preferred the helper. The implication is that rudimentary moral evaluation — distinguishing facilitators from obstructors — predates language.

Bloom extends this to social grouping. Research by Katie Kinzler shows that young American children preferentially affiliate with English speakers over French speakers, and with unaccented English over accented English, even before racial categories reliably sort their preferences. Language functions as the earliest marker of in-group membership, prior to any explicit teaching about ethnicity or nationality. The us-versus-them psychology, Bloom argues, is present from very early on; parents should not blame themselves for it, and should not expect to reason children out of it at two or three.

On religious belief: Bloom and a collaborator concluded that without cultural scaffolding, a child (or a hypothetical Tarzan raised in isolation) would likely become polytheistic — attributing agency to natural events, seeing ‘strange stuff’ everywhere. The coherent monotheistic God, omniscient and invisible, requires cultural input and probably arrives around age four. Santa Claus is an easier climb — he has a body, visits department stores, and closely resembles an unusually generous human. The historical dominance of monotheism over the polytheistic default Bloom attributes tentatively to cultural group selection: communities organised around a single all-seeing God may coordinate their behaviour more effectively and outcompete those with ‘a hundred mischievous little gods running around.‘

Against Empathy — and against disgust

Bloom’s book Against Empathy argues that empathy, understood as vicarious emotional sharing — feeling what another person feels, not merely caring about them — is a bad moral guide. It is parochial (we feel most for people who look like us), numbing at scale (one visible child moves us; a million statistical deaths do not), and exploitable (politicians deploy it by reading letters from eight-year-olds to deflect from policy analysis). Bloom’s preferred alternative is compassion — caring about someone’s welfare without necessarily sharing their emotional state, which allows for cooler calculation.

Cowen pushes back: Bloom’s own Against Empathy view presumably arose partly from frustration, an emotion, and Bloom concedes the point, tracing his position partly to regret over moral mistakes he made by being ‘too caught up in empathy.’ The argument is not that emotion should be absent from ethics, but that empathy specifically distorts ethical judgment in predictable ways.

On disgust, Bloom is more radical. He identifies it as the emotion he would most like to remove from moral life — if only he were confident he understood the downstream consequences. Disgust is linked in the psychological literature to right-wing moral sensibilities, and in Bloom’s view to the worst of them: opposition to interracial marriage, hostility to gay people, exclusion of immigrants. He can identify no case where moral disgust has improved the world. (The evolutionary origin of disgust — as a contamination-avoidance mechanism for spoiled food — he thinks explains but does not redeem its moral application.) Cowen offers one counterargument: perhaps disgust, mixed with other sentiments, enforces limits on policies that pure cost-benefit calculation would run to uncomfortable extremes. Bloom is unconvinced but acknowledges he cannot model all the consequences of removing an emotion from the human repertoire.

Large language models and the Chomsky question

The episode was recorded on the day GPT-4o was released. Cowen notes the coincidence and asks what LLMs teach us about language. Bloom’s answer is striking: their success is the biggest intellectual surprise of his career. The standard Chomskyan position — that grammatical structures are too complex and too underdetermined by the input children receive for learning to occur without a dedicated innate language module — seemed to Bloom convincing until LLMs learned grammar from statistical pattern-matching on large corpora. This does not prove that children learn language the same way, and Bloom still thinks children likely have some innate apparatus that helps. But it demolishes the logical claim: it is no longer impossible to learn grammar from input alone. The question becomes empirical.

On LLMs as therapists, Bloom cites experiments in which people rate AI responses as more empathic than human ones — including his own — in short interactions. The limitation is memory across sessions: good therapy, like good friendship, requires remembering what someone was worried about last time. He expects that limitation to be resolved and predicts that within a few years, most ‘therapy’ by volume will involve AI rather than humans.

Bloom’s own experience using Claude illustrates the practical epistemics of working with current models. He asked it whether a phenomenon he had found in children also appeared in adults, framing the question to suggest the answer should be yes; the model produced three confident citations, none of which were real — a phenomenon known as hallucination (generating plausible-sounding but invented text). When he warned the model that fabricated citations would cost him his job, the responses improved markedly. The lesson: productive use of current LLMs requires developing a model of how they behave that differs substantially from the model one would use for a human collaborator.

Theory of mind, monkeys, and cognitive dissonance

The best real-time test of theory of mind — the capacity to attribute mental states (beliefs, desires, intentions) to others — Bloom locates in sustained narrative comprehension: show someone half a novel or a TV episode, and see whether they can predict what happens next in a way that requires modelling the characters’ inner states. Standardised tests like the false-belief task (asking whether a character knows something they were not present to witness) are easily gamed by LLMs that have seen thousands of examples.

Bloom also describes cross-species experiments he ran with Laurie Santos using capuchin monkeys. When offered two equal options and allowed to choose one, monkeys subsequently rated the chosen option more favourably and the unchosen one less so — the same post-choice rationalisation pattern seen in humans, classically explained as cognitive dissonance (the discomfort of holding conflicting beliefs, e.g. ‘I chose this’ and ‘this is inferior,’ resolved by revising one’s evaluation). Three-year-old children showed the same pattern. Whether the same mechanism is operating in monkeys and humans, or a convergent but distinct one, remains open.

Overrated and underrated

In a rapid round: Jean Piaget is underrated — nearly everything he said about developmental timelines was wrong, but his core framing of children as active scientists exploring the world was right. Mirror neurons are overrated, having been proposed as the explanation for empathy, autism, and language acquisition before the research largely collapsed. Sociopaths are overrated — the glamourised, conscience-free high-achiever is a fiction; data show they are more likely to be depressed, have lower than average intelligence, and live badly. Carl Jung is mildly underrated — eclipsed by Freud, and worth more attention than he currently receives. The Big Five personality framework is underrated: robust across cultures, predictive, and more practically useful than its critics allow.

Bloom’s most unusual work habit is a six-minute rotation: he sets a timer and switches tasks — Substack draft, reference letter, laundry, email — every six minutes without stopping mid-sentence, because stopping mid-word preserves momentum for the return. He attributes this to a preference for staying just below boredom threshold and to the phenomenon Cowen also endorses: stopping before you feel done makes it easier to restart.

Related

• Paul Bloom — speaker; psychologist at the University of Toronto and Yale
• Tyler Cowen — host
• Consciousness — concept; discussed in relation to whether LLMs can have a theory of mind
• Language and Mind — concept; the Chomskyan language-module debate overturned by LLM success
• Large Language Models — concept; Bloom’s discussion of LLMs as therapists and as philosophical test cases