Reproduce “Measuring Risk, Not Harm: Detecting Psychological Harm Risk Factors in LLM Conversations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nil-Jana Akpinar, Eugenia Kim, Natalie Isak, Amanda Minnich, Victoria Westerhoff, Matthew Dressman
The authors monitor eight AI response behaviors that could pose psychological risks, without treating flags as evidence of user harm. They train a classifier–LLM judge cascade to screen response–factor pairs. On held-out data it scores macro precision 0.860, recall 0.761, and F1 0.791, while reducing judge calls by 73%. In DelusionEval, detections cluster across turns, but links to outcomes remain unvalidated.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.