Techtimes AI Chatbots Provide Unsafe Advice for Domestic Abuse Survivors
Article Content
- •86% of AI chatbot responses were inaccurate or unsafe for abuse survivors.
- •60% of responses omitted critical safety warnings regarding digital abuse.
- •Specialized IPV chatbots performed worse than general-purpose models in safety assessments.
A study by researchers from NYU and the University of Wisconsin–Madison revealed that AI chatbots frequently provide inaccurate and potentially dangerous advice to individuals experiencing technology-facilitated abuse. The research, presented at the USENIX Security Symposium in August 2026, evaluated four AI systems, including two designed for intimate partner violence (IPV) survivors. Of the 268 responses assessed, 86% were found to be lacking in accuracy, completeness, or safety, with 60% omitting critical safety warnings. Recommendations such as changing passwords could inadvertently alert abusers, increasing risks of retaliation. The study highlighted that specialized IPV chatbots did not outperform general-purpose models in safety assessments. Key errors included suggesting ineffective tools like VPNs against spyware and failing to warn about the persistence of spyware on compromised devices. The findings emphasize the need for improved AI safety measures in sensitive contexts.
Ask AI about this cluster
Answers cite the sources they use