When a computer algorithm treats people unfairly, they tend to get just as angry and retaliate just as often as they would against a human wrongdoer. A new series of experiments published in Journal of Experimental Psychology: General provides mixed evidence on whether punishing a machine brings the same sense of closure as punishing a person, but it suggests that a machine’s response to being punished, such as sounding hurt or promising to make fairer offers, can strengthen people’s sense that justice was done. The findings suggest that as artificial intelligence takes on more decision-making roles in society, organizations may need to design systems that can communicate accountability when things go wrong.
Automated systems are increasingly tasked with consequential choices, from screening job applicants to steering vehicles and prioritizing medical care. To understand how people react when these systems cause harm, psychologists study a concept known as mind perception, which is the everyday process of judging what kind of mental life another entity possesses. People generally evaluate minds along two separate dimensions: agency, meaning the ability to think, plan, and act intentionally, and experience, meaning the capacity to feel physical pain, pleasure, fear, and emotions.
In human interactions, the capacity to commit a moral wrong and the capacity to experience punishment usually go together, since a sane adult can both plan a harmful act and feel the sting of a penalty. Artificial intelligence splits these two qualities apart because people often view machines as having enough agency to cause deliberate harm, yet almost no ability to feel pain or remorse. For instance, research covered by PsyPost in 2023 found that people judged AI agents as more deserving of blame than humans in one classic life-and-death dilemma. Similarly, another 2023 study covered by PsyPost showed that people hold automated driving assistants responsible for road outcomes while simultaneously categorizing the software as a mere tool rather than a true moral being.
This mental divide raises questions about whether punishing an algorithm can produce justice satisfaction, which is the feeling of closure, fairness, and resolution a victim experiences after an offender is penalized. Psychological theories generally propose two reasons why punishment feels satisfying: either it fulfills a desire for retribution by making the wrongdoer suffer for their actions, or it serves a communicative role by teaching the wrongdoer a moral lesson.
Because people tend to see machines as unable to genuinely suffer or feel guilt, study author Alexander Eng, an associate professor of organizational behavior in the Department of Management at the Asia School of Business in Kuala Lumpur, and co-author Kai Chi Yam of the National University of Singapore examined how people emotionally and behaviorally respond during and after they penalize an unfair algorithm.
“Research on punishment usually assumes that anyone capable of doing wrong can also be punished,” Eng told PsyPost. “We tend to see a young child as unable to truly commit a wrong, and so also as unfit for punishment. Machines break that link: people see AI as able to make decisions that hurt them, yet unable to feel pain, regret, or learn a moral lesson. As AI takes on decisions like medical triage and firing employees, we wanted to know how people respond when they try to hold it accountable, and whether punishing it gives them any sense that justice was done.”
The researchers began with a pilot study of 380 English-speaking adults in the United States, recruited through the online platform Prolific, to verify whether people actually punish algorithms. Participants played a single round of an economic exercise called the ultimatum game, in which a partner was given £0.50 and proposed how to split it. If the participant accepted the split, both parties kept their share, but if the participant rejected it, neither party received anything, making rejection a form of costly punishment.
Participants were told their partner was either another human participant or an algorithm, and they received either a fair offer of £0.25 each or an unfair offer where the partner kept £0.45 and left the participant just £0.05. In reality, the researchers preset all offers. Receiving a stingy offer from an algorithm provoked much higher anger (an average of 3.12 on a 5-point scale) than receiving a fair offer from an algorithm (1.31). When given the unfair deal, participants felt comparable anger toward algorithms (3.12) and humans (2.91), and they punished the unfair algorithm by rejecting its offer 57.8 percent of the time, which was not statistically different from the 62.2 percent rejection rate for unfair humans.
In their first main experiment, the authors examined what happens after punishment is delivered using a final sample of 123 Prolific participants. Every participant received the unfair £0.05 offer in the ultimatum game from either a human or an algorithm, decided whether to accept or reject it, and then answered five questions measuring their justice satisfaction, such as whether they felt everyone got what they deserved. Afterward, participants played a second exercise called the dictator game, where they were given £0.50 and could unilaterally share any amount in £0.05 increments with their partner, while also having the option to switch to a new human partner first.
As in the pilot study, participants rejected unfair offers from algorithms (38 out of 62) and humans (40 out of 61) at similar rates. Overall, the 78 participants who punished their partner by rejecting the offer reported higher justice satisfaction (3.57 out of 5) than the 45 who accepted it (2.63). Controlling for whether participants switched partners, the satisfaction gap between those who punished and those who did not was larger when the offender was a human than when it was an algorithm, and higher justice satisfaction in turn predicted greater generosity in the second game.
Because the option to switch partners or the belief that computers have no use for money might have influenced the first experiment, the researchers conducted a second study with a stricter setup among 739 participants, preregistered along with the third and fourth experiments through the journal’s registered-report review process. Participants were told that five algorithms and five human research assistants were competing across multiple negotiations, and that while neither kept the money personally, the lowest-earning algorithm and human would be eliminated from the competition.
Participants were not allowed to switch partners between the ultimatum game and the dictator game, and they also completed scales measuring their partner’s perceived agency and emotional experience, along with their own baseline tendency to attribute human-like consciousness to technology, a trait known as anthropomorphism.
Participants rated human partners higher on agency (3.30 versus 2.79 out of 5) and much higher on the capacity for emotional experience (3.20 versus 1.86) than algorithmic partners. Even so, participants rejected the unfair offer at comparable rates for algorithms (71.0 percent, or 262 of 369) and humans (65.4 percent, or 242 of 370), and those who punished their partner felt much greater justice satisfaction (3.53) than those who accepted the unfair split (2.71). In contrast to the first study, after controlling for baseline anthropomorphism, the link between punishing and justice satisfaction, and in turn generosity, was equally strong whether the partner was an algorithm or a human.
“People react to an unfair machine much as they react to an unfair person,” Eng said. “Participants got just as angry at an algorithm that made them a lowball offer as at a human who made the same offer, and they were just as willing to give up their own money to punish it. Whether punishing the machine felt as satisfying was less consistent across studies.”
A third study tested these dynamics in a different scenario called the trust game with 735 participants who chose to invest £0.20 with a human research assistant or an algorithm. The investment was tripled to £0.60, and the partner broke the participant’s trust by returning just £0.10 instead of a fair £0.30 split. Participants could then penalize their partner by recommending on a 1-to-5 scale that the partner be removed from the player pool, before rating their justice satisfaction and playing a second round where the roles were reversed.
Participants gave similar removal ratings to unfair algorithms (3.37 out of 5) and unfair humans (3.26). Unexpectedly, harsher removal ratings were associated with lower justice satisfaction rather than higher closure. The authors suspect this is because participants only submitted a recommendation and never saw a penalty take effect, so the ratings may have reflected anger more than punishment. In an exploratory analysis looking directly at anger, feeling angrier predicted lower justice satisfaction and less generosity toward both types of partners, with anger reducing justice satisfaction slightly more when the partner was human than when it was an algorithm.
In a fourth study involving 3,684 participants who rejected the unfair £0.05 offer in the ultimatum game, the researchers tested whether a message from the penalized partner could increase justice satisfaction. Immediately after rejecting the offer from a human or an algorithm, participants received either no message at all, a message expressing emotional distress over repeated rejections, a message calmly acknowledging that the offer was too low and promising to be fair in the future, or a combined message expressing both distress and a moral lesson learned. A separate preliminary test with 305 people confirmed that these messages effectively conveyed emotional suffering and moral learning as intended.
Across the entire sample, participants paired with an algorithm reported slightly lower justice satisfaction (3.56 out of 5) than those paired with a human (3.66), a small difference that accounted for about half a percent of the overall variation in satisfaction scores. However, receiving a message from an algorithm expressing emotional suffering (3.57) or moral learning (3.62) produced higher justice satisfaction than receiving no response from the algorithm (3.42).
For human partners, the differences between receiving no message (3.57) and receiving a suffering (3.66) or moral learning (3.69) message were not statistically significant after adjusting for multiple comparisons. Still, the boost from a message was not statistically larger for algorithms than for humans, and combining both messages added no measurable benefit beyond either one alone. The results suggest that a response from a machine can help satisfy people’s need for redress, even though participants rated algorithms as far less capable of feeling than humans.
“We expected an apology or a show of remorse from an algorithm to count for little, since people don’t believe algorithms feel anything,” Eng explained. “Yet those messages raised people’s sense of justice for algorithms at least as much as the same messages did for humans. It was also striking that people paid to punish an algorithm that couldn’t keep the money and couldn’t experience the loss.”
A final supplemental experiment with 1,093 participants examined whether dissatisfaction after being wronged causes people to take out their frustration on an innocent bystander, a behavior known as displaced aggression. Participants received the unfair £0.05 offer in the ultimatum game from a human or an algorithm, decided whether to accept or reject it, and then played the dictator game paired with a brand-new human research assistant who had no involvement in the first round.
Controlling for baseline anthropomorphism, participants who punished their initial partner by rejecting the unfair offer shared slightly more money with the innocent bystander (an aggression score of 6.75 versus 7.07 on an 11-point scale, where higher scores mean less money shared) than those who did not punish, regardless of whether the original wrongdoer was a person or a machine.
“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.”
As with all research, there are some caveats to consider when interpreting these results. All of the experiments relied on brief online financial games where the harm consisted of being offered an unfair split of £0.50 by an anonymous text-based partner. How people think and feel when a simple computer program shortchanges them by a few cents may differ from how they react to high-stakes algorithmic harm in the real world, such as an automated system wrongfully denying a mortgage, firing an employee, or causing a traffic accident. In addition, participants chose for themselves whether to punish, so differences between those who punished and those who did not may partly reflect who chose to punish rather than the effect of punishing itself.
“These were online economic games with small stakes, splitting 50 pence,” Eng pointed out. “Real harms from AI, like a denied loan or a wrong diagnosis, carry much more weight, so the effects could look different there. The justice satisfaction results were also mixed: one study found a clear gap between machines and humans and another found none, so I wouldn’t claim that punishing machines never satisfies people.”
The evidence on whether punishing a machine feels less satisfying than punishing a person was mixed across the studies, appearing in the first and fourth experiments but not in the second, while the third experiment’s punishment measure did not work as intended. The clearest gap came from the smallest experiment, which included 123 participants after dozens of duplicate entries were removed and which had originally been preregistered to test different outcomes; the larger, preregistered second experiment did not replicate it.
The authors also noted that their five-item survey measuring justice satisfaction showed relatively low internal consistency in the fourth study, indicating a need to develop more precise tools for measuring psychological closure after technological failures. Future research could explore how different types of real-world errors and corrective feedback shape public trust, helping engineers and policymakers design accountability measures, such as holding human creators rather than machines responsible, that satisfy human intuitions of fairness.
“We want to test other kinds of wrongdoing and punishment beyond economic games, and develop a better measure of justice satisfaction,” Eng added. “In a related project, I’m studying the ‘AI disclosure penalty,’ where people value content less once they’re told AI made it.”
The study, “Artificial Intelligence and Machines Are Capable of Harm, but Not of Experiencing Punishment,” was authored by Alexander Eng and Kai Chi Yam.