The moral mind why your ethical intuitions are less reliable than you think — and what to do about it

A brother and sister, both adults, are traveling together and decide one night to have sex. They use contraception. They enjoy it, but agree not to do it again. They keep it to themselves. No one is harmed. Neither is hurt or troubled by it afterward.

Is it wrong?

Most people say yes, immediately. Then the interviewer starts in on the reasons. What about genetic harm? Ruled out — they used contraception. Social confusion for children? There are no children, and they agreed to keep it private. Might one of them have felt coerced? The scenario stipulates mutual consent. What if it damages their relationship? It didn’t — they’re closer.

And one by one, the reasons fall. The person being interviewed grants each concession, sometimes reluctantly but honestly. Yes, if you stipulate no genetic harm, I accept that. Yes, if you stipulate secrecy and mutual agreement, I accept that. And at the end of this process — reasons exhausted, every stated objection acknowledged and ruled out — the verdict doesn’t move.

“I still think it’s wrong. I just know it is.”

Sit with that for a moment. The person saying this is not stupid, and they’re not being obtuse. They’re describing something true about themselves with unusual accuracy. The conviction didn’t begin with the reasons. The reasons were consulted after the conviction arrived. And when the reasons failed, the conviction remained standing on its own — not because there was something the person hadn’t yet articulated, but because the verdict was never being held up by reasons in the first place.

Haidt called this moral dumbfounding: the persistence of a moral judgment after every supporting reason has been stripped away. Understanding it requires understanding what moral judgment actually is — which turns out to be quite different from what most people assume.

The emotional dog and its rational tail

The dominant Western tradition in moral psychology runs from Piaget through Kohlberg: moral development is the progressive acquisition of reasoning capacity, a movement from primitive rule-following toward principled reflection on justice and rights. In this picture, moral judgment is — if it’s working properly — the output of moral reasoning. You think it through. You reach a conclusion. If you’re at a sufficiently advanced developmental stage, you can even explain why.

Jonathan Haidt’s 2001 paper in Psychological Review argued that this picture is substantially backward.

Haidt’s social intuitionist model holds that moral judgment is the product of fast, automatic, affective processing — something that fires quickly, produces a verdict, and feels like a conclusion even though it was never reasoned to. What follows the verdict is rationalization: the process of generating reasons that could support it, conducted by a system whose task is not to evaluate whether the verdict is correct but to defend it. The metaphor Haidt reaches for is a dog and its tail. The emotional system is the dog. It acts. Reason is the tail — it moves in response to what the dog does. Watch only the tail and you’d conclude the dog was reasoning. But the tail isn’t driving anything.

This feels counterintuitive from the inside, and deliberately so. When you reason about a moral question, it genuinely feels like reasoning — you’re considering, weighing, sometimes revising. Haidt’s response is to distinguish between what the conscious mind can observe about itself and what’s actually generating the outputs. You can observe the reasoning process. You can’t observe the affective verdict that arrived before the reasoning started.

The affective primacy claim draws on Robert Zajonc’s work showing that affective responses to stimuli can precede cognitive evaluation and operate independently of it. Zajonc’s 1980 paper in American Psychologist documented that people form preferences before forming beliefs. Haidt applies this directly to moral judgment: the “wrongness” of the incest scenario arrives like a physical sensation before the justification machinery has started.

Haidt’s 2001 paper marshaled four converging lines. Moral judgments are fast and automatic — the reaction precedes deliberation and feels like perception rather than inference. When subjects generate justifications, those justifications serve a social function: they’re offered to persuade others, not derived as premises for a conclusion. Moral reasoning is predominantly motivated reasoning — subjects almost never revise their initial verdict downward from a counterargument; upward revision happens only when other people endorse the opposing view. And moral intuitions propagate through social networks before individuals reason their way to them — people adopt the moral views of their communities through social transmission, then generate the supporting reasons afterward.

The mechanism, from the subject’s perspective: reasoning feels like it’s working because it’s all you can observe. But the machinery that delivered the verdict is running beneath the level of observation, and the reasoning process isn’t evaluating the verdict — it’s building the case for it.

The dual-process debate

Haidt's social intuitionist model does not go uncontested. Joshua Greene's dual-process model, developed in tandem with neuroimaging research, holds that moral judgment involves two distinguishable systems: a fast, emotion-driven process and a slower, more deliberate one, and that the deliberate system can and sometimes does override the emotional one. Paxton and Greene (2010), in Topics in Cognitive Science, pushed back on what they saw as the overly deflationary reading of reason's role, arguing that deliberate reasoning does sometimes produce genuine moral revision, not just post-hoc rationalization.

Haidt's response, and the position that has held up better in subsequent research, is not that reason never produces genuine revision but that it does so far less often, and under far narrower conditions, than the rationalist tradition assumed. The core claim — that moral judgment is predominantly affective, that reason operates mostly downstream, and that the rationalist picture of moral development substantially misfires — has survived this debate. What remains contested is the proportion and the conditions. This is live science, not settled doctrine, and the honest position acknowledges real uncertainty about exactly how much the deliberate system can override the emotional one. But the baseline: your moral verdicts arrive before your reasons, and the reasons' primary job is defense, not discovery.

Moral dumbfounding — what the mechanism looks like under pressure

The “Julie and Mark” vignette is Haidt’s own example, designed to demonstrate precisely what the social intuitionist model predicts: a scenario constructed to strip away every standard harm-based justification for condemning incest, such that if the judgment depended on those reasons, removing them should destabilize it. The experimental finding, replicated many times, is that it doesn’t.

The term “moral dumbfounding” — coined by Haidt in collaboration with Fredrik Björklund — refers specifically to this: the stubborn maintenance of a moral judgment in the absence of supporting reasons, accompanied by the subject’s puzzled awareness that the supporting reasons have run out. The dumbfounding isn’t the judgment itself; it’s the persistence of the judgment after the justificatory props have been knocked away, combined with visible bewilderment at the position one is left in.

McHugh, Brophy, McGrane, and colleagues (2017), working with explicit protocols to identify measurable indicators of moral dumbfounding, found that around 17% of participants explicitly admitted having no reasons while maintaining their judgment, and a further 9% produced “unsupported declarations” — assertions of wrongness with no accompanying justification. These numbers are conservative: only participants who actually verbalized the structure of dumbfounding. A significant proportion of participants who had initially offered harm-based justifications ended up, after conceding those justifications, still maintaining the judgment.

What happens when the reasons run out is specific: participants don’t reach for new arguments. They express bewilderment. “I just feel it.” “It’s just wrong, I can’t explain it.” These read like evasions if you approach them as failures of reasoning. They read as accurate self-reports if you take the social intuitionist model seriously: the conviction didn’t depend on the reasons because it didn’t arise from them. The reasons were recruited afterward. Remove the recruits and the conviction stands on its own, where it always was.

The implication: the reasons offered were not the grounds for the judgment. They were post-hoc rationalizations — not deliberate dishonesty, but the reasoning process searching for justifications for a position it didn’t create, using machinery whose job is advocacy, not evaluation.

The Royzman challenge

The dumbfounding research has a serious methodological challenger. Royzman, Kim, and Leeman (2015), in Judgment and Decision Making, argued that many participants who appeared dumbfounded in Haidt's paradigm were not actually failing to provide reasons — they were implicitly rejecting the harmlessness stipulation. They found the incest scenario, and kept finding harms the stipulation hadn't explicitly ruled out: psychological effects, long-term relational damage, harms too remote to see from outside. The apparent dumbfounding, on this reading, was participants holding onto a reason — harm — they weren't willing to fully release, rather than maintaining a reasonless verdict.
This is a real and substantive critique, not a quibble. If most apparent dumbfounding dissolves on closer inspection into implicit harm reasoning, the phenomenon would be far less diagnostic of affective primacy than Haidt claimed.

McHugh et al.'s response is empirical: dumbfounding persists even among participants who explicitly acknowledge and accept the harmlessness stipulation. When you control for this — when you look only at participants who verbally confirmed that the scenario entails no harm — the effect doesn't vanish. It narrows, as any experimental effect narrows when you apply more stringent controls, but it remains. The phenomenon is real. It's narrower than Haidt originally suggested and less universal than his earlier writing implied. But it exists, and its existence is informative about the structure of moral conviction.

The same math, different verdict

The trolley problem is overused, but bear with it for a moment, because the numbers are genuinely strange.

In the switch scenario, a runaway trolley heads toward five people. You can pull a lever to divert it onto a side track, killing one instead of five. Most people say pull it — studies put approval at around 85–90% in Western samples, and Awad and colleagues’ survey of 70,000 participants across 42 countries found a mean approval rate of about 81%.

In the footbridge scenario, the math is identical: five people in the trolley’s path, one death prevents five. But here the mechanism is different. You’re on a bridge. The only way to stop the trolley is to push a large stranger onto the tracks, using his body as an obstacle. In the Awad et al. data, approval dropped to around 51%. In earlier Western-sample studies, the majority said it was not permissible.

Same math. Same outcome. What changes is the causal structure: one case involves a mechanism, the other requires your hands on another person’s body. The verdict flips.

Greene and colleagues’ 2001 fMRI study in 과학 found a neural correlate for exactly this asymmetry. When subjects contemplated the footbridge scenario — personal harm, direct causation, physical contact — brain regions associated with emotional processing showed preferential activation. When they contemplated the switch scenario — impersonal harm, indirect causation, mechanical mediation — deliberative reasoning regions showed greater relative activity. Different neural systems, different verdicts.

The interpretation matters enormously: the moral intuition against pushing the stranger isn’t tracking something morally different about the two cases. It’s tracking something evolutionarily ancient — an aversion to direct physical harm calibrated for an environment where the only people you could harm were close enough to touch. Foot (1967) introduced the philosophical version of this dilemma; Thomson (1985) formalized the distinction between harmful redirection and harmful use. Both noted the asymmetry in moral intuition. Neither had a satisfying explanation. The explanation arrived through neuroscience and evolutionary psychology, and it’s uncomfortable: the thing driving the different verdict isn’t a morally relevant difference. It’s an artifact of how moral psychology was built.

The instinct against pushing the stranger — vivid, physical, direct — is the same instinct that makes distant harms feel less morally urgent. Supply chain labor conditions, air pollution policy, financial decisions that impoverish millions — these involve real, quantifiable harms that in many cases exceed the harms generating immediate moral reaction. But they arrive through mechanisms, at distance, involving statistical people rather than visible individuals. The evolved aversion to direct physical contact is simply not triggered. And so the moral urgency isn’t felt.

The ecological validity question

Trolley problems are obviously artificial. They strip away uncertainty, require no action, involve no interpersonal consequences, and bear almost no resemblance to actual moral decisions people face. A critic has genuine grounds to ask whether findings from trolley scenarios generalize anywhere.

The reply is that Greene's broader claim — that personal-harm and impersonal-harm moral judgments engage different neural systems — has been replicated using methods that aren't trolley problems. In charitable giving studies, Kogut and Ritov (2005) found that donors gave more to a single identified sick child than to a group of eight, even when all eight were identified. Small, Loewenstein, and Slovic (2007) — the study described in the next section — documented the same asymmetry with real donation decisions. The trolley problem illustrates a phenomenon it didn't create. The mechanisms it reveals operate in contexts that aren't artificial: whenever someone makes a decision that harms a distant, statistical person rather than a nearby, visible one, the same affective machinery is producing the same asymmetric outputs. The trolley is a lens. What it's showing is real.

The arithmetic of compassion

In 1992, William Desvousges and colleagues asked people how much they would pay to prevent migratory birds from drowning in oil-covered waste ponds. One group was told the program would save 2,000 birds. A second: 20,000 birds. A third: 200,000.

The willingness-to-pay figures came back as follows: $80, $78, and $88.

The difference between 2,000 and 200,000 — a factor of one hundred — produced a difference in moral concern so small it disappeared in statistical noise. People were responding to the image of a bird struggling in an oil pond. Not to the quantity. The affective system fired once, on a mental image, and the image was not sensitive to whether it represented 2,000 or 200,000 instances of itself.

Call this scope insensitivity. It isn’t a mistake about numbers — the Desvousges participants presumably knew that 200,000 was larger than 2,000. Knowing a number and feeling it are different cognitive operations, served by different systems, and the system that motivates moral action runs on feeling rather than knowing.

Thomas Schelling identified the “identified lives” problem in a 1968 chapter in a Brookings Institution volume: statistical lives and identified lives are not treated as morally equivalent, even when we know they should be. A named child with a photograph generates donations at a scale that an anonymous statistic, even one representing millions of deaths, cannot approach — not because donors are hypocrites but because the affective system responds to an individual, identifiable human being in a way it simply doesn’t respond to a number.

Small, Loewenstein, and Slovic pressed on this in a 2007 paper in Organizational Behavior and Human Decision Processes, and found something more disturbing than simply documenting the effect. They found that presenting statistical information alongside an identified victim actually reduced donations compared to presenting the identified victim alone. Adding the numbers crowded out the emotional engagement rather than supplementing it. The information that should have increased the moral weight of the appeal — here is one person, and she represents 3.2 million — instead activated the analytic system, which suppressed the emotional response.

Don’t smooth over this finding. It means the obvious antidote — more information, more quantification — can make things worse. If analytical thinking and emotional engagement compete for activation, and emotional engagement is what drives moral action, the tools most available to correct scope insensitivity may work against the motivation they’re meant to serve.

Jessica McClure — eighteen months old, fallen into a backyard well in Midland, Texas in 1987 — generated roughly $1.2 million in public donations over fifty-eight hours of live broadcast. The mechanism is specific: a face, a name, a narrative unfolding in real time. Anonymous statistical populations, however large, cannot produce that response. It reflects a moral psychology failure before it reflects a media or political failure. Media and politics amplify a prior architecture. They respond to what generates emotional engagement because that’s what the audience’s moral psychology runs on.

Collapse of compassion

Slovic's 2007 paper in Judgment and Decision Making, "If I Look at the Mass I Will Never Act: Psychic Numbing and Genocide," examined how moral response scales — or rather, fails to scale — with the magnitude of suffering. Subjects shown one dying child gave at a certain level. Shown two dying children, they gave less than double. As numbers increased, the response not only failed to scale proportionally; it could flatten or decline.

Slovic applied this explicitly to how societies respond to mass atrocity. The psychological mechanism that makes genocide possible is not primarily a failure of information — the world generally knows, roughly, what's happening. It's a failure of moral response to scale with the numbers. One death is a tragedy; a million deaths becomes, in the brain's processing, something more like an abstraction than a tragedy. The response is the same: not indifference, exactly, but something like the maximum emotional engagement the affective system can sustain, which turns out to be calibrated to individuals, not quantities.

The universalism problem

The most accessible escape from everything established so far is this: individual moral intuitions may be distorted by proximity, identity, and affect, but there is a universal moral sense — common across cultures, grounded in something like human nature — that, properly understood, tracks moral truth.

Haidt and Joseph’s 2004 paper in Daedalus proposed what became Moral Foundations Theory: five (later six) innate but culturally calibrated dimensions of moral concern. Care/Harm. Fairness/Cheating. Loyalty/Betrayal. Authority/Subversion. Sanctity/Purity. (Liberty/Oppression was added in subsequent iterations.) The claim wasn’t that all cultures moralize identically but that they draw on a shared set of foundations, weighting them differently in response to local history, ecology, and social structure.

Graham, Haidt, and Nosek’s 2009 paper in the Journal of Personality and Social Psychology established a finding that’s robust within the United States: liberals moralize predominantly around Care and Fairness; conservatives weight all five foundations more evenly, including Loyalty, Authority, and Sanctity at levels liberals don’t match. Haidt and Graham’s 2007 paper in Social Justice Research explored the practical consequence: each side frequently doesn’t recognize the other’s moral concerns as moral at all.

The cross-cultural pattern holds. Doğruyol, Alper, and Yilmaz (2019), testing the five-factor model across a broader range of societies including non-WEIRD samples, found that the structure of moral foundations is stable across diverse cultures, even as the specific weightings vary.

But here’s the problem with treating this as a universalism rescue. The claim that moral intuitions track objective moral facts is available to every culture simultaneously — which is exactly what disqualifies it as validation for any particular culture’s intuitive weightings. Every culture has the intuitions its history, ecology, and social structure produced. The evolutionary origins of those intuitions explain why they exist; they don’t confer epistemic privilege. And the cross-cultural divergence data — people in different societies weighting Care and Loyalty and Authority at different levels, each with equal confidence — doesn’t vindicate any of them.

Distinguish this from moral relativism, which would be a different mistake. The cross-cultural data doesn’t show that all moral views are equally valid. It removes the option of treating any culture’s moral certainties as trustworthy simply by virtue of being widely shared or intensely felt. Wideness and intensity are properties of the certainty, not the certainty’s content.

And moral progress is real, which means the picture is tractable but not simple. Slavery, the status of women, the treatment of children, the scope of who counts as a full moral patient — the direction of movement is not random. The mechanisms are multi-causal: changed economic conditions, religious shifts, new sympathies built through literature and contact, political organizing, philosophical argument over long timescales. What these cases share: the morally correct endpoint required overriding strongly held majority intuitions that showed no tendency to self-correct. Widespread, confident moral feeling was wrong and had to be worked against. Not a small qualification.

The WEIRD problem

Henrich, Heine, and Norenzayan's 2010 paper in Behavioral and Brain Sciences documented that the overwhelming majority of psychological research had been conducted on WEIRD samples — Western, Educated, Industrialized, Rich, Democratic — that are not representative of human populations even on basic cognitive measures. Haidt's early moral foundations research, which drew heavily on American samples, was no exception.

This is a real methodological concern. If the foundations were identified primarily in Western, educated populations, their claim to universality is at risk. Cross-cultural work conducted since — including Doğruyol et al. (2019) explicitly testing across 30 diverse societies — consistently finds that cultural variation in moral weighting is real and substantial. The WEIRD critique weakens early claims about the specific biological universality of particular foundations and their exact content. It does not rescue the universalism argument: the finding that moral foundations exist across cultures coexists with the finding that they're weighted very differently, which is exactly the problem.

What corrected moral reasoning actually requires

Moral judgment arrives predominantly as a feeling, produced by affective machinery calibrated for proximity, visibility, and individual identity. The machinery responds poorly to scale. It responds poorly to distance. It responds differently to identified individuals than to statistical populations. And there is no universal moral sense that transcends these distortions, because the moral senses humans have are the ones evolutionary and cultural history produced, not the ones that track what matters at the scale of modern collective decisions.

What’s called for is not feeling less, not detaching, not becoming someone who evaluates suffering analytically rather than emotionally. The failures are structural — built into how the affective response system processes information — and structural failures require structural corrections. Individual attitude adjustments are the wrong unit of intervention. A person who resolves to care more about distant suffering will do so right up until the identified victim appears, at which point all available moral attention will redirect.

Explicit quantification of scope at the decision-making stage does work that spontaneous moral response cannot. Forcing the comparison — 2,000 versus 200,000 on the same scale, evaluated by the same criteria — produces judgments the affective system won’t deliver unaided. The methodology of effective altruism, whatever its other problems, is built around this: not feeling more but measuring differently. Singer’s 1972 argument in Philosophy & Public Affairs — that geographic distance and anonymity are not morally relevant, that a drowning child has the same claim on you whether she’s in front of you or on another continent — provides the philosophical scaffolding. GiveWell’s cost-effectiveness analysis is the methodology applied: how many lives per dollar, requiring choices between causes to be made on grounds other than which one arrived with a face.

Institutional design matters more than individual resolution. Foreign aid triage made in response to television coverage rather than systematic analysis of where a dollar does the most good is a structural failure, not primarily a personal one. Designing institutions to require explicit comparative analysis before resources are committed is hard — the same moral psychology that distorts individual giving drives the political constituencies that fund aid organizations. But the design target is clear, even where the path is not.

Treating distance and mechanism as morally irrelevant in policy means putting the supply chain worker and the pollution victim on the same ledger as the person standing in front of you. Regulatory cost-benefit analysis attempts this correction. It often fails. But the structure is right, even when the execution isn’t.

One normative dependency to name plainly: these corrections presuppose that scale and distance are morally relevant — that 200,000 deaths count more than 2,000. That is a substantive normative position, broadly consequentialist, that the empirical evidence creates strong pressure toward but doesn’t establish. What the evidence does show: any framework that ignores scale and distance will reproduce the specific failures documented above. What belongs on the other side is a normative choice. Naming this is what makes the argument honest rather than circular.

Effective altruism has its own failure modes — speculative longtermism, methodological insularity, and the FTX collapse, in which Sam Bankman-Fried — the movement’s most prominent donor and public advocate — was convicted on seven counts of fraud and sentenced to twenty-five years in federal prison. None of that refutes the core methodology. It does illustrate that structural corrections are vulnerable to the same motivated reasoning they’re designed to counter, now applied to the framework itself.

The institutions that should embody structural correction at scale — democratic governments, international aid bodies, public health agencies — are all responsive to the intuitions this article has been dismantling. Responsive to individual stories. To photographs. To visible, proximate crises. The distance between where these institutions are and where they’d need to be is a moral psychology gap, and closing it requires sustained effort against the grain of how moral motivation works.

The moral mind in its native element

Return to the opening participant — reasons exhausted, conviction intact — with everything the previous sections have established in view.

That person is not reasoning badly. Their moral cognition is working exactly as it was built to work. The incest scenario triggers an evolved aversion — disgust response, in Haidt’s framing — that served a function where incest carried real costs: genetic, social, structural. The conviction isn’t irrational in the sense of being defective machinery. It’s irrational in the sense of machinery built for one purpose operating where that purpose doesn’t apply. The scenario has been stipulated to lack the harms that make incest aversion adaptive. The aversion doesn’t notice.

The same is true of the bystander who couldn’t push the stranger off the footbridge, the donor who gave generously to the named child and almost nothing to the unnamed famine, the politician who scrambled resources toward the televised disaster and ignored the quieter one killing more people at a slower rate. None of them are failing to apply their moral instincts. They’re applying them faithfully. That’s the problem. Each instinct is functional in its native range and mute in the one that determines who gets help and who doesn’t.

Still, a concession — and a real one. Moral convictions that survive rational pressure can encode real social information: accumulated warnings from experience, norms whose function isn’t transparent, implicit knowledge about where community trust runs. Sometimes the resistance to revision tracks something that hasn’t been made articulable yet. Moral intuitions evolved because they sometimes tracked something. Dismissing them entirely in favor of explicit calculation is a different error.

But hold this concession to its precise scope. It does not extend to allocating resources between 200,000 people the decision-maker will never see, whose suffering produces no emotional signal because they have no faces. It does not extend to the judgment that a drowning child nearby has a stronger claim than one dying of malaria on another continent, when the only difference is proximity. In those contexts — precisely the contexts where collective moral decisions operate — the intuition isn’t encoding hidden wisdom. It’s running on a mismatch.

The participant at the opening knew something. Their conviction was functioning moral knowledge within its range. What they didn’t know — what most people don’t know, and what the science of moral psychology makes harder to continue not knowing — is how narrow that range is, and how far most decisions that actually matter fall outside it.

Gen AI 면책 조항

이 페이지의 일부 콘텐츠는 생성형 AI의 도움을 받아 생성 및/또는 편집되었습니다.

미디어

KATRIN BOLOVTSOVA – Pexels

Key Sources and References

Haidt, J. (2001). The emotional dog and its rational tail: A social intuitionist approach to moral judgment. Psychological Review, 108(4), 814–834.

Greene, J. D., Sommerville, R. B., Nystrom, L. E., Darley, J. M., & Cohen, J. D. (2001). An fMRI investigation of emotional engagement in moral judgment. Science, 293(5537), 2105–2108.

Zajonc, R. B. (1980). Feeling and thinking: Preferences need no inferences. American Psychologist, 35(2), 151–175.

Kohlberg, L. (1971). From is to ought: How to commit the naturalistic fallacy and get away with it in the study of moral development. In T. Mischel (Ed.), Cognitive development and epistemology (pp. 151–235). New York: Academic Press.

Paxton, J. M., & Greene, J. D. (2010). Moral reasoning: Hints and allegations. Topics in Cognitive Science, 2(3), 511–527.

McHugh, C., Brophy, M., McGrane, J., Power, L., Waldron, F., & Conway, M. (2017). Searching for moral dumbfounding: Identifying measurable indicators of moral dumbfounding. Collabra: Psychology, 3(1), 23.

Björklund, F., Haidt, J., & Murphy, S. (2000). Moral dumbfounding: When intuition finds no reason. Lund Psychological Reports, 1(2). Department of Psychology, Lund University. [Originally circulated as an unpublished manuscript, University of Virginia.]

Royzman, E. B., Kim, K., & Leeman, R. F. (2015). The curious tale of Julie and Mark: Unraveling the moral dumbfounding effect. Judgment and Decision Making, 10(4), 296–313.

Foot, P. (1967). The problem of abortion and the doctrine of the double effect. Oxford Review, 5, 5–15.

Thomson, J. J. (1985). The trolley problem. Yale Law Journal, 94(6), 1395–1415.

Awad, E., Dsouza, S., Shariff, A., Rahwan, I., & Bonnefon, J.-F. (2020). Universals and variations in moral decisions made in 42 countries by 70,000 participants. Proceedings of the National Academy of Sciences, 117(5), 2332–2337.

Kogut, T., & Ritov, I. (2005). The “identified victim” effect: An identified group, or just a single individual? Journal of Behavioral Decision Making, 18(3), 157–167.

Desvousges, W. H., Johnson, F. R., Dunford, R. W., Boyle, K. J., Hudson, S. P., & Wilson, K. N. (1992). Measuring nonuse damages using contingent valuation: An experimental evaluation of accuracy (RTI Monograph 92-1). Research Triangle Park, NC: Research Triangle Institute.

Schelling, T. C. (1968). The life you save may be your own. In S. B. Chase, Jr. (Ed.), Problems in public expenditure analysis (pp. 127–176). Washington, DC: Brookings Institution.

Small, D. A., Loewenstein, G., & Slovic, P. (2007). Sympathy and callousness: The impact of deliberative thought on donations to identifiable and statistical victims. Organizational Behavior and Human Decision Processes, 102(2), 143–153.

Slovic, P. (2007). If I look at the mass I will never act: Psychic numbing and genocide. Judgment and Decision Making, 2(2), 79–95.

Graham, J., Haidt, J., & Nosek, B. A. (2009). Liberals and conservatives rely on different sets of moral foundations. Journal of Personality and Social Psychology, 96(5), 1029–1046.

Haidt, J., & Joseph, C. (2004). Intuitive ethics: How innately prepared intuitions generate culturally variable virtues. Daedalus, 133(4), 55–66.

Haidt, J., & Graham, J. (2007). When morality opposes justice: Conservatives have moral intuitions that liberals may not recognize. Social Justice Research, 20(1), 98–116.

Henrich, J., Heine, S. J., & Norenzayan, A. (2010). The weirdest people in the world? Behavioral and Brain Sciences, 33(2–3), 61–83.

Doğruyol, B., Alper, S., & Yilmaz, O. (2019). The five-factor model of the moral foundations theory is stable across WEIRD and non-WEIRD cultures. Personality and Individual Differences, 151, 109547.

Singer, P. (1972). Famine, affluence, and morality. Philosophy & Public Affairs, 1(3), 229–243.

GiveWell. (2017, June 1). How GiveWell uses cost-effectiveness analyses. GiveWell Blog. https://blog.givewell.org/2017/06/01/how-givewell-uses-cost-effectiveness-analyses/

United States Attorney’s Office, Southern District of New York. (2024, March 28). Sam Bankman-Fried sentenced to 25 years in prison for orchestrating massive fraud. U.S. Department of Justice. https://www.justice.gov/usao-sdny/pr/sam-bankman-fried-sentenced-25-years-prison