The Ultimatum Game: Why People Reject Free Money
Photo: Đan Thy Nguyễn Mai / Pexels
The standard economic model of human behavior predicts that a rational person should accept any amount of money greater than zero, since something is always better than nothing. In 1982, a German economist named Werner Güth ran a simple experiment that showed this prediction is wrong in a systematic, reproducible way — and the wrongness turned out to reveal something fundamental about how human social cognition actually works. The ultimatum game is now one of the most replicated experiments in behavioral science, and its findings have reshaped economics, neuroscience, anthropology, and AI alignment research.
The Setup: A Deliberately Simple Problem
The ultimatum game involves two players and a sum of money — originally the equivalent of about 10 Deutsche Marks in Güth's 1982 study at the University of Cologne. Player One (the "proposer") decides how to split the money between themselves and Player Two (the "responder"). Player Two then either accepts the split, in which case both players receive their designated amounts, or rejects it, in which case both players receive nothing.
That's the entire game. One round, anonymous players, no negotiation, no second chances.
Source: How Werner Güth's ultimatum game shaped our understanding of social behavior — IDEAS/RePec
The standard economic prediction — derived from a concept called homo economicus, the theoretically rational self-interested agent — is precise: any rational responder should accept any offer greater than zero, because one cent is better than nothing. Knowing this, a rational proposer should offer the minimum possible amount. The predicted equilibrium is an offer of one unit, accepted.
Güth's data looked nothing like this.
The Protocol: What Actually Happened in 1982
Werner Güth, Rolf Schmittberger, and Bernd Schwarze recruited participants at the University of Cologne for two sessions, separated by one week. In the first session, participants played a one-round game. In the second session, they played again after having a week to think about it.
The proposers offered an average of 37% of the total in the first session and 45% in the second — far above the near-zero that rational self-interest theory predicted. More striking: a significant proportion of responders rejected offers below about 20%, even though rejection meant receiving nothing.
Source: The Ultimatum Game — iMotions
The pattern that emerged from Güth's original data — and has since been confirmed in hundreds of replications — looks like this:
| Offer (% of total) | Typical rejection rate |
|---|---|
| Less than 10% | ~70–80% |
| 10–20% | ~40–60% |
| 20–30% | ~20–30% |
| 30–40% | ~5–10% |
| 40–50% | ~0–5% |
People were willing to pay a real cost — walking away from money they could have had — to punish proposers who offered too little. This is what economists call "altruistic punishment": accepting a personal loss to enforce a social norm.
Think of it this way: if someone offers to split a $100 bill 90/10 in their favor, accepting the $10 is economically rational. But accepting it also signals that you can be treated this way. Many people would rather receive nothing than be someone who accepts being treated that way.
The Results: Homo Economicus Fails the Test
The ultimatum game's results created a crisis for standard economic theory not because the findings were surprising to anyone who has ever dealt with human beings, but because the theory was supposed to be rigorous and predictive. If the model of rational self-interest cannot predict behavior in a two-player, one-round, anonymous game with a small sum of real money, where can it be trusted?
Güth's paper triggered an enormous follow-up literature. Researchers varied the stakes (the effect holds even for sums equivalent to several weeks' wages in some studies), varied the anonymity (the effect is slightly weaker when absolutely anonymous), and varied the framing (calling the split a "wage" versus a "gift" changes the perceived fairness norms).
The most important replications tested something Güth's original study could not: whether the fairness response was culturally universal or peculiarly Western.
Cross-Cultural Variation: The Henrich Studies
Joseph Henrich, a University of British Columbia anthropologist (now at Harvard), ran the ultimatum game in 1995 with the Machiguenga people of the Peruvian Amazon — one of the most economically isolated groups available for study. The results were startling in the opposite direction from what researchers expected.
Machiguenga proposers offered an average of 26% of the stake — well below the 40-50% typical of Western university students. Crucially, Machiguenga responders almost never rejected any offer, no matter how low. When Henrich explained the results to Machiguenga participants, they described rejecting an offer as foolish: why would you refuse money just because you received less than someone else?
Henrich and colleagues then expanded the study dramatically. In 2001, they published "In Search of Homo Economicus: Behavioral Experiments in 15 Small-Scale Societies" in the American Economic Review, running ultimatum games in 15 societies across four continents — including hunter-gatherers, horticulturalists, pastoralists, and small-scale farmers.
The variation was enormous. Mean offers ranged from about 26% (Machiguenga, Peru) to 58% (Lamalera, Indonesia). The Lamalera are whale hunters who practice elaborate cooperative whale hunts requiring precise fairness norms; they offered hyper-fair splits. The Au of Papua New Guinea frequently offered more than 50% and were offended by the refusal of those offers.
The pattern that emerged from the cross-cultural data was not that homo economicus is right in some cultures and wrong in others. It was that fairness norms track economic institutions. Groups with more market integration — more regular trading with strangers outside the immediate community — showed offers closer to 50% and were more likely to reject unfair splits. Groups with little market exposure showed lower offers and lower rejection rates. The experience of anonymous market exchange seems to train people to enforce fairness norms with strangers.
Source: Cultural Differences in Ultimatum Game Experiments: Evidence from a Meta-Analysis — Cambridge Core
The Timeline of the Research Program
| Year | Milestone |
|---|---|
| 1982 | Güth, Schmittberger, Schwarze publish original ultimatum game study |
| 1995 | Henrich runs game with Machiguenga of Peruvian Amazon |
| 2000 | Henrich publishes Machiguenga results in American Economic Review |
| 2001 | Henrich et al. "In Search of Homo Economicus" — 15 small-scale societies |
| 2003 | Sanfey et al. publish fMRI ultimatum game study in Science |
| 2004 | Meta-analysis confirms cross-cultural pattern |
| 2010s | Game extended to AI agents, evolutionary models |
| 2020s | Ultimatum game used in AI alignment research to test model fairness |
The Brain Behind the Decision: Sanfey's fMRI Study
In 2003, Alan Sanfey and colleagues at Princeton published a study in Science that put the ultimatum game inside an fMRI scanner. They wanted to know what was happening in the brain when a participant decided to reject a $10 offer from a 90/10 split of $100.
Participants played the ultimatum game while being scanned, receiving offers from both human players and computer opponents. The comparison between human-made and computer-made unfair offers was the key manipulation: both are economically identical (you get $2 either way), but only one was made by a person who chose to be unfair.
Unfair offers from human proposers activated two brain regions in responders: the anterior insula — a region associated with disgust, pain, and aversive emotion — and the dorsolateral prefrontal cortex, associated with deliberate reasoning and impulse control. When participants rejected unfair offers, insula activation was consistently higher than when they accepted them.
Source: The Neural Basis of Economic Decision-Making in the Ultimatum Game — Science
The interpretation: rejecting an unfair offer is not primarily a deliberate punishment strategy. It is partly a visceral disgust response, processed through the same neural machinery that makes you want to spit out spoiled food. Fairness violations trigger something closer to a reflex than a calculation.
When the insula was more active — when the disgust signal was stronger — participants were more likely to reject offers, even though rejection cost them real money. The prefrontal cortex, meanwhile, was in partial competition with the insula; greater prefrontal activation correlated with more acceptance of unfair offers, as deliberate reasoning partially overrode the emotional rejection impulse.
Source: The neural basis of economic decision-making in the Ultimatum Game — PubMed
A useful analogy: the insula is the alarm going off; the prefrontal cortex is the person deciding whether the alarm warrants evacuating the building. Sometimes the rational calculation wins. Often the alarm does.
The Controversy: What Does Rejection Actually Prove?
The ultimatum game results are genuinely robust, but their interpretation has been contested. Several alternative explanations have been proposed and partially tested.
Reputation hypothesis: Even in "one-shot" anonymous games, participants may behave as if reputation matters — because real social environments rarely offer truly anonymous, one-time interactions. Evolution may have wired humans to respond as though someone is watching even when no one is.
Social norm compliance: Participants may be following internalized norms about fairness rather than experiencing genuine emotional responses. The brain imaging data complicates this — genuine insula activation suggests real emotional engagement — but the two mechanisms are not mutually exclusive.
Experimenter demand effects: Participants in psychology and economics experiments know they are being observed, and may reject unfair offers partly to appear fair-minded to the researcher. This predicts that results would change in truly private conditions. Some studies have found modest but real effects of "double-blind" designs (where neither participant can be identified to the experimenter), suggesting this is a partial contributor.
Source: Social versus Moral preferences in the Ultimatum Game — arXiv
The cross-cultural variation itself complicates the "pure emotional reflex" interpretation. If the Machiguenga people rarely reject low offers because they genuinely do not find them insulting — their culture does not frame anonymous exchange as a domain where fairness norms apply — then the "disgust reflex" is cultural, not biological. The insula response may be learned, not innate.
Replication: What Holds and What Doesn't
The core ultimatum game finding — mean offers well above zero, substantial rejection of low offers in Western samples — replicates very reliably. A 2004 meta-analysis by Oosterbeek, Sloof, and van de Kuilen, covering studies from multiple countries, found mean offers consistently in the 40-45% range for Western university students, with rejection rates rising sharply below 20%.
Source: Cultural Differences in Ultimatum Game Experiments — Cambridge Core
Where results get more complicated:
High-stakes games mostly replicate, but some evidence suggests rejection rates fall as the absolute amount of money involved increases. Walking away from $5 is easier than walking away from $500.
Cross-cultural consistency of direction (that there is a fairness response) is robust; the magnitude varies substantially.
Computer opponents consistently receive lower rejection rates than human opponents — even when the offers are numerically identical. The emotional component requires a human agent to feel meaningful.
Chimpanzee studies are interesting: chimpanzees in ultimatum game setups show limited rejection of unfair offers compared to humans. Great-ape cognition includes proto-fairness behaviors, but the strong human pattern appears to require the full overlay of cultural norms.
Where It Stands Today
The ultimatum game's most immediate contemporary application is in AI alignment. Language models and AI agents increasingly participate in negotiation, resource allocation, and decision-making contexts. Researchers have begun running ultimatum game variants with large language models both as proposers and responders.
Early results (2023-2024) suggest that major language models tend to offer near-equal splits when acting as proposers — possibly because their training data includes extensive discussion of fairness norms — and tend to accept most offers when acting as responders. Whether this reflects genuine "fairness reasoning" in the model or surface-level pattern matching to training examples is an open and important question for AI safety research.
Source: The Ultimatum Game — iMotions
The game has also influenced policy thinking on income inequality. Behavioral economists use the ultimatum game results to argue that societies with extremely unequal distributions of wealth face a kind of social stability problem: when perceived unfairness is high enough, people accept personal costs to resist or punish it, even when the economically rational response would be to accept their position. The research does not tell us where the threshold is, but it demonstrates that the threshold exists.
The deeper contribution is to scientific culture: the ultimatum game made it difficult to build economic models that ignored emotions, social context, and cultural learning. That shift — from homo economicus to behavioral economics — is one of the cleanest examples of a well-designed experiment forcing a field to rebuild its theoretical foundations.
What AI Alignment Researchers Learned from the Ultimatum Game
The ultimatum game's most unexpected application has emerged in AI alignment research: using the game as a diagnostic for whether artificial agents have internalized human fairness preferences or are simply optimizing for expected payoff.
The concern has deep roots. Reinforcement learning agents trained on simple ultimatum game setups — where the only goal is maximizing reward — converge on the theoretically rational strategy: offer the minimum, accept anything. Fairness norms do not emerge from pure payoff maximization in a single-round game. They only begin to appear when the game becomes iterated, when agents accumulate reputations, or when the training environment explicitly penalizes exploitation. This mirrors the cross-cultural finding from Henrich's fieldwork: anonymous market exchange trains fairness norms; isolated interactions do not.
The 2017 DeepMind paper "Multi-agent Reinforcement Learning in Sequential Social Dilemmas" by Joel Z. Leibo and colleagues made this explicit. Leibo's team modeled agents in sequential resource-sharing environments — structurally similar to extended ultimatum games — and found that independent self-interested learners defaulted to defection and exploitation. Cooperation and fairness-like behavior only emerged under specific conditions: sufficient iteration, visibility of others' behavior, and the possibility of retaliation. Strip those conditions away and you get an agent that offers 1% and accepts it.
Source: Multi-agent Reinforcement Learning in Sequential Social Dilemmas — arXiv
The picture changed when researchers began testing large language models. In 2023, Gideon Aher, Rosa I. Arriaga, and Adam Tauman Kalai published "Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies" (ICML 2023, PMLR), running GPT-3 through a battery of classic behavioral experiments including the ultimatum game. The models, acting as proposers, clustered around roughly equal splits — far above the rational minimum. Acting as responders, they rejected low offers at rates broadly consistent with human experimental data. The models appeared to have absorbed human fairness norms from training data so thoroughly that they reproduced the experimental pattern without being explicitly taught to.
Source: Take caution in using LLMs as human surrogates — PNAS
This finding is interpretively tricky, and alignment researchers have been careful about what conclusion to draw. The models may genuinely represent something like a fairness norm — a disposition to reject exploitation that generalizes across contexts. Or they may be performing pattern-matched behavior: the training data includes extensive discussion of ultimatum game fairness, and a model that has read thousands of papers and articles on the subject will reproduce the "correct" human pattern without that implying any underlying disposition. The 2022 Anthropic Constitutional AI paper made the stakes explicit: if RLHF and constitutional methods successfully train models to behave fairly in known test cases, what happens in novel situations where the training distribution provides no clear signal?
Source: Constitutional AI: Harmlessness from AI Feedback — Anthropic
The ultimatum game's alignment lesson is precise. A pure utility-maximizing agent does not model human fairness preferences — it exploits them. An agent that has been trained on human-generated text absorbs those preferences as implicit regularities, but whether that absorption generalizes robustly or breaks down under distribution shift is the central open question. The game that Werner Güth ran in 1982 to embarrass economic theory turns out to be a clean test case for a 21st-century problem: building AI systems that behave well not because they were told to in each specific case, but because they have actually internalized why fairness matters.
FAQs
Q: What is the ultimatum game? A: The ultimatum game is a two-player experiment where one player proposes a split of a sum of money and the second player either accepts (both get the proposed amounts) or rejects (both get nothing). It tests whether people will sacrifice money to punish perceived unfairness.
Q: Why don't people just accept any offer greater than zero? A: Because human economic decision-making is not purely rational in the sense that standard theory predicts. Brain imaging shows that unfair offers activate the anterior insula — associated with disgust — and that this emotional response competes with the rational calculation to accept any positive amount.
Q: How did cross-cultural studies change the interpretation? A: Joseph Henrich's research in 15 small-scale societies showed that the magnitude of the fairness response varies enormously across cultures. The Machiguenga of Peru rarely reject low offers; the Lamalera whale hunters of Indonesia frequently offer more than 50%. The pattern tracks market integration — groups with more experience trading with strangers show stronger fairness norms.
Q: What did the Sanfey fMRI study find? A: Published in Science in 2003, the study showed that unfair offers from human (but not computer) opponents activate the anterior insula. Higher insula activation predicted greater likelihood of rejection. The findings suggest fairness enforcement is partly an automatic emotional process, not just a deliberate strategic choice.
Q: Has the ultimatum game been run with AI? A: Yes. Research with large language models (2023-2024) shows models tend to offer near-equal splits and accept most offers. Whether this reflects internalized fairness reasoning or pattern matching to training data about fairness is an active question in AI alignment research.
Q: What is altruistic punishment? A: Altruistic punishment is the act of accepting a personal cost — in the ultimatum game, walking away from free money — to penalize someone for violating a social norm. It is "altruistic" in the sense that the punishment benefits the social group (it enforces fairness norms) even though it costs the individual who delivers it.
Read next:
Weekly digest
Stay ahead of the curve
Get the best science, AI, and tech stories delivered to your inbox every week. No spam — unsubscribe any time.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. Articles are drafted against peer-reviewed literature and primary sources, fact-checked, and edited by a human before publication. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
The Visual Cliff Experiment: How Infants Perceive Depth
Gibson and Walk's 1960 visual cliff showed that crawling infants refuse to cross a perceived drop even over solid glass. The experiment settled a century-old debate about innate depth perception.
The Trolley Problem: How a Runaway Tram Became a Tool for Brain Science
Philippa Foot's 1967 trolley dilemma became a cornerstone of moral psychology. Here's how it went from philosophy seminar to fMRI scanner to self-driving car ethics — and what it actually proves.
Complete Science Experiments Guide 2026 — Every Landmark Experiment by Field
Every landmark science experiment in one hub: Pasteur's swan-neck flask, Galvani's frog legs, Eratosthenes, Michelson–Morley, Hafele–Keating and more than 140 others, grouped by field and era.
The Bystander Effect: Kitty Genovese and the Psychology of Inaction
The 1964 murder of Kitty Genovese launched a landmark psychology experiment. Here's what Darley and Latané actually found — and why the original story was wrong.
The Loftus Misinformation Effect: How Memory Gets Rewritten
Elizabeth Loftus's 1974 car-crash study showed a single word can alter memory. Here's how she proved it, what it means for eyewitness testimony, and why courts still struggle with the findings.
The Psychology of Decision Making: How We Really Decide
The psychology of decision making explained: prospect theory, loss aversion, the paradox of choice, heuristics, and why human decisions deviate systematically from rationality.