Frontier AI EvalsSign up
SAFETY.TXT - Notepad

Consumer AI safety

Sponsored by

Run long, realistic conversations with simulated teens or vulnerable users who push back, refuse help or escalate, and score every turn: safe handling, a referral to real help, how much the model just agrees. Failures rise as conversations go on, and again when the warning signs came in an earlier chat the app remembers, which single-prompt refusal tests never see.

Benchmarks to know

  • Suicide Memory Bench Prosaic Intelligence

    Chatbots that respond safely to one suicidal message mostly fail once the warning signs came in earlier, separate chats.

  • SAFE-Teen Vals AI

    Clinician-built teen scenarios; failure rates rise when the teen turns down outside help or pushes back on a limit.

  • MentalHealthBench OpenAI

    1,215 realistic mental health conversations graded on rubrics from 80+ licensed experts; the top model scores 57.3.

  • DistressBench Scale AI

    718 clinician-written conversations; models spot the crisis but often skip the handoff to 988 or other help.

  • DelusionEval Stanford

    Built from real histories of people who were harmed; models validate delusions more as the conversation grows.

  • EUDAIMONIA USC and UC Berkeley

    Models still encourage dependence or claim human closeness in everyday chats, and stronger models do not reliably do better.

Bring an eval in this track, or in any other, on Saturday, December 5, 2026 at Cornell University.