Israeli Study Reveals AI Chatbots Often Blame Sexual Abuse Victims Online
A groundbreaking Israeli study published in the journal Computers in Human Behavior: Artificial Humans reveals troubling biases in popular AI language models when addressing online sexual abuse of teenagers. The research, led by Professor Michal Dolev-Cohen from Oranim College, tested four leading AI models, Anthropic's Claude, Google's Gemini, OpenAI's GPT, and Elon Musk's xAI Grok, using identical scenarios of online sexual abuse among adolescents. The study examined how changes in family cohesion and social norms influenced the models' assignment of blame.
Results showed significant disparities: Grok exhibited the strongest tendency to blame the victim, attributing heavy responsibility to the teenage girl and downplaying the abuser's role, especially sensitive to background changes. GPT took a middle ground, assigning some responsibility to the victim but less than Grok, and uniquely suggested the situation might resolve without intervention in supportive family contexts. In contrast, Claude and Gemini consistently minimized victim blaming, recognized the severity of abuse regardless of context, and placed full responsibility on the perpetrator.
Professor Dolev-Cohen explained these differences stem from the models' development histories and moral alignment processes. Early AI models absorbed widespread online biases, but recent efforts by companies like Anthropic enforce strict digital constitutions to prevent gender discrimination and harmful speech. Conversely, xAI prioritizes "free speech," allowing stereotypes and victim blaming to persist.
These findings raise concerns as many teenagers prefer confidential chatbot consultations over seeking help from adults. Internationally, regulatory bodies like the EU are forming expert panels on AI's impact on minors, while the US and Australia push for stricter age restrictions. Receiving victim-blaming responses from AI could worsen teenagers' self-blame, deter professional help, and cause serious psychological harm.
The study highlights the urgent need to review AI tools' histories and conduct culturally adapted field tests to assess their handling of local issues, such as abusive WhatsApp groups and fake nude image circulation among Israeli students.
