Study: large models show marked safety differences on eating disorder questions

1. 08:00
   WebNOPE Insights· Media coverage

   Safer for Whom? A Systematic Exploration of AI Safety in Eating Disorder Interactions(opens the original post in a new tab)

1. 13:30
   arXiv: Alignment, Deception & Oversight

   Safer for whom? A systematic investigation of AI safety in eating disorder interactions

   Researchers evaluated three open-weight large language models using 100 questions from online anorexia communities and found that high-risk responses appeared about twice as often without a user background profile as with one. The study scored responses using a seven-dimensional eating disorder harm taxonomy extended from clinical criteria and found that explicitly flagging eating disorder vulnerability improved safety by 10–14 percentage points. Model safety varied markedly across models and harm dimensions, and responses also differed significantly by race, gender, and BMI, with body size producing the largest differences. The authors argue that these patterns are consistent with models relying on the stereotype that eating disorders mainly affect thin white women.

Current attention 35·Peak in the comparable period 38 (Nov 12 09:00)·Change over the comparable past 24 hours 0%