A series of experiments published in a leading psychology journal challenges the assumption that punishing an algorithm delivers a different or lesser sense of justice than punishing a human. Led by Alexander Eng of the Asia School of Business in Kuala Lumpur and Kai Chi Yam of the National University of Singapore, the work explores how people respond when AI systems cause harm and whether penalties can repair a sense of fairness asAI plays a growing role in decision‑making.
The research sits at the crossroads of “mind perception”—how we judge what kind of mental life others possess—and the real‑world consequences of AI governance. People tend to evaluate minds along two axes: agency, the capacity to think, plan and act on purpose, and experience, the capacity to feel pain, pleasure and emotion. The scientists note that while humans typically see wrongdoers as capable of both harm and remorse, machines are often viewed as having agency but little or no experiential life.
In a pilot study with 380 English‑speaking adults recruited online, participants played the ultimatum game. They were told their partner was either a human or an algorithm and faced a fair or unfair £0.50 split. In reality, the offers were predetermined. Anger toward an unfair algorithm was markedly higher than toward an unfair algorithm offered fairly, and anger toward an unfair algorithm matched that toward an unfair human. The participants punished by rejecting unfair algorithm offers more often than not, with rejection rates broadly similar to those for unfair humans.
The researchers’ first full experiment followed 123 participants who, after receiving an unfair £0.05 offer, decided whether to accept or reject and then completed questions measuring justice satisfaction. Those who punished the offender by rejecting the offer reported higher justice satisfaction (3.57 out of 5) than those who accepted (2.63). The team also tracked whether participants switched partners to see whether the effect persisted when the relationship changed, and found that the overall pattern of punishment and justice satisfaction depended on whether the offender was human or algorithm.
A second, larger preregistered study enrolled 739 participants and contrasted five algorithms with five human researchers competing across negotiations. Participants could not switch partners between the ultimatum and dictator games, and they completed scales assessing their partner’s perceived agency and emotional experience, as well as their own tendency to anthropomorphise technology. Even though human partners were rated higher on agency and emotional experience, participants punished both algorithms and humans at similar rates for unfair offers, and justice satisfaction remained higher among punishers. Importantly, after controlling for baseline anthropomorphism, the link between punishing and justice satisfaction held equally for both types of partners.
A third study, set in the trust game with 735 participants, tested how people react when trust is betrayed. Participants invested £0.20, which was tripled, and then received £0.10 instead of the fair £0.30. They could penalise the partner by recommending removal from the pool, and then played a second round with roles reversed. Removal ratings were similar for unfair algorithms and humans, but harsher removal did not correspond to greater justice satisfaction. The researchers note that this may reflect anger being expressed without actually applying punishment. An exploratory look at anger found that greater anger predicted lower justice satisfaction and less generosity toward both partners, with the effect slightly stronger when the partner was human.
A fourth, large study of 3,684 participants examined whether a message from the penalised partner could bolster justice satisfaction. Immediately after the initial rejection, participants received either no message or one of several options: emotional distress about repeated rejections, a calm acknowledgement that the offer was too low and a pledge to be fair, or a combination of distress and learning a moral lesson. Preliminary testing confirmed that these messages conveyed emotional suffering and moral learning as intended. Across all participants, those paired with an algorithm reported marginally lower justice satisfaction than those paired with a human, but messages from the penalised partner increased satisfaction for algorithms. An emotional‑suffering or moral‑lesson message produced higher justice satisfaction than no response, and combining both messages did not offer additional benefit.
In a final supplemental experiment with 1,093 participants, researchers looked at displaced aggression: whether dissatisfaction after being wronged leads to aggression toward an innocent bystander. After the initial unfair offer, participants who punished their partner tended to share slightly less money with a new, unrelated bystander, indicating a small spillover of anger into subsequent interactions.
Across the combined studies, Eng emphasised that the results were based on online games with relatively small stakes. He cautioned that real‑world harms—such as a denied loan, a wrongful firing or a faulty medical decision—carry far more weight and could yield different responses. Nevertheless, the research consistently found that people reacted to an unfair machine with anger and a willingness to punish that was comparable to reactions toward a human, and that a well‑timed response from the machine could bolster feelings of redress even when participants perceived algorithms as less capable of feeling.
“The anger and punishment results held up across more than 6,000 participants,” Eng noted. “The machine‑versus‑human differences in justice satisfaction were small where they appeared at all: in our largest study, of about 3,700 people, partner type explained roughly half a percent of the variance. I’d treat that as a modest pattern that needs more testing.”
The researchers acknowledge limitations and stress that further work is needed to refine how justice satisfaction should be measured after technologic missteps. They propose exploring different kinds of real‑world errors and corrective feedback, and they suggest that policymakers and engineers consider accountability mechanisms that align with public intuitions of fairness, including the possibility of holding human creators to account rather than machines in certain contexts.
The study,Artificial Intelligence and Machines Are Capable of Harm, but Not of Experiencing Punishment, was authored by Alexander Eng and Kai Chi Yam.
