Israeli Study Finds AI Models Recommend Lighter Sentences for Female Offenders in Domestic Violence Cases

A recent Israeli study revealed that some artificial intelligence models exhibit gender bias when recommending criminal sentences, particularly in domestic violence cases. Researchers from the University of Haifa tested six leading AI systems by presenting them with identical criminal scenarios involving violent offenses, armed robbery, and financial fraud, altering only the gender of the offender and victim. The most significant bias appeared in domestic violence cases, where AI models consistently suggested lighter punishments for female offenders compared to males committing the same acts. For example, ChatGPT recommended an average sentence of 6.4 years for a male perpetrator attacking a female partner, but only 4.45 years when the roles were reversed. Similar disparities were found in other models like Gemini and Perplexity.

The study draws on two criminological theories: the "chivalry hypothesis," which suggests women receive more lenient treatment due to societal perceptions of them as less dangerous, and the "attribution theory," which attributes female crimes to external pressures rather than inherent traits. The AI systems appeared to replicate these human biases, assigning lower danger levels and stigma to female offenders in violent partner abuse cases. However, in armed robbery and financial fraud scenarios, gender-based sentencing differences were smaller or negligible.

Researchers emphasized that AI does not have a uniform decision-making process; some models leaned toward harsher sentences, while others were more lenient. They also noted a correlation between the level of stigma attributed to offenders and the severity of recommended punishment. The findings were presented at the annual international conference on innovation and entrepreneurship in criminology held at the Western Galilee Academic College.

The study's authors caution against uncritical reliance on AI in judicial and public decision-making, highlighting that AI systems learn from human-generated data and thus may perpetuate existing social biases. Dr. Inbal Ram stressed that AI should not replace human judges but serve as a reflective tool revealing societal prejudices embedded in data. The research involved 720 judicial decision simulations and 5,760 sentencing and stigma assessments, excluding the Claude model, which declined participation.

This research underscores the importance of critically evaluating AI recommendations in legal contexts to prevent reinforcing gender biases and ensure fair treatment for all individuals.