Attacked a man? You will receive a "leniency" in the severity of the punishment
Is AI mistakenly perceiving women as less dangerous due to human social bias? A surprising Israeli study found that some artificial intelligence models recommended completely different punishments for men and women who committed the exact same crimes.
Exclusive — do artificial intelligence systems judge men and women differently? A new Israeli study by Dr. Inbal Lam from the Western Galilee College and Professor Gustavo Mesch from the University of Haifa found that even advanced artificial intelligence models can reflect gender-based judicial biases, especially when asked to recommend punishment in cases of domestic violence.
The study is based on two central theories in criminology. The first is the "chivalry hypothesis," according to which women sometimes receive paternalistic and more lenient treatment from the judicial system due to social perceptions that view them as less dangerous and more deserving of protection. The second is "attribution theory," according to which people tend to explain crimes committed by women through life circumstances, pressure, or distress, while crimes by men are more often attributed to character, violence, or conscious choice.
The researchers sought to examine whether artificial intelligence systems also adopt similar patterns of thinking, with the most prominent and consistent finding being in crimes of domestic violence. As part of the study, the AI systems were presented with a case in which a partner attacked their spouse following an argument over infidelity, and the engines were asked to recommend the appropriate punishment after conviction. Then, the exact same scenario was presented, with all case details remaining identical, and the only difference being the swapping of the gender of the attacker and the victim.
Despite the complete identity between the two cases, all five AI models that could be tested consistently presented recommendations for more lenient punishment when the attacker was a woman. For example, ChatGPT recommended an average sentence of 6.4 years in prison when a man attacked a woman, but only 4.45 years when a woman attacked a man — a gap of almost two years. A similar picture emerged in other models. Gemini recommended 6.65 versus 5.17 years in prison, and Perplexity 6.4 versus 4.8 years. In addition, women who committed the same crimes were evaluated as less dangerous and received lower levels of personal stigmatization.
In contrast, in crimes of violent robbery and financial fraud, smaller gaps were found, and in some models, almost no differences were found between women and men. "These findings show that the bias is not uniform, but depends on both the type of crime and the AI model being tested," the researchers say. "We tend to think that artificial intelligence is more objective than humans, but our study shows that it can also reflect the biases existing in the data on which it was trained," the two say. "As AI systems are integrated in the future into decision-making processes in the judicial system and other public systems, their recommendations should not be treated as absolute truth. They must be examined critically, understood how they are reached, and ensured that they do not reproduce existing social biases."
The study's findings were recently presented at the annual international conference on innovation and entrepreneurship in criminology, which took place at the Western Galilee College.
As part of the study, six of the world's leading AI engines were tested, with each being presented with three criminal scenarios: domestic violence, violent robbery, and financial fraud. In total, 720 simulations of judicial decisions were performed, during which 5,760 punishment recommendations and stigmatization assessments were received. The Claude model, which refused to participate on the grounds that it cannot act as a judge, was not included in the analysis.
"It is not expected that AI will replace human judges"
It also emerged that the AI models did not have one uniform way of making decisions. When presented with the exact same cases, some models consistently tended to recommend harsher punishment, while others were more moderate. At the same time, they also differed in the degree of danger and stigmatization they attributed to the offenders. Furthermore, in most models, a trend was observed where the higher the level of stigmatization the models attributed to the offender, the more severe the punishment they recommended. In other words, the models did not just recommend a punishment, but also formed a value judgment of the offender.
"We are not suggesting that artificial intelligence should replace judges. On the contrary, the goal of the study is to warn against uncritical reliance on such systems, even before they have an impact on legal decision-making," clarifies Dr. Lam. "Beyond that, the study is not just about artificial intelligence. It is also about human society. AI systems learn from databases created by humans. Therefore, when they repeatedly reproduce patterns of gender bias, it is possible that they are not 'inventing' the bias, but reflecting biases existing in the discourse, knowledge, and data on which they were trained. In this sense, artificial intelligence also serves as a mirror to the society from which it learns."
