AI models show language-dependent bias in responding to domestic abuse scenarios

A new research paper analyzes how seven widely used language models respond to requests for help regarding coercive control against women. The study found that AI systems developed by non-anglophone companies were more likely to fail in their own native languages. Furthermore, the models' ability to identify coercive control and affirm the user's agency varied significantly across different languages. While two frontier systems maintained a consistent standard across all tested languages, indicating that a protective ceiling is achievable, failures in other models highlight design-related outcomes. AI

IMPACT Highlights potential biases in AI safety measures and the need for language-specific robustness in AI systems designed to assist vulnerable users.

RANK_REASON Research paper analyzing AI model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Coercive Control
• conversational AI
• intimate partner violence
• Language Models

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →