Tansu Yegen on 'crash test dummy' AI harm evals: tens of thousands of simulated chats a day
TansuYegen · x · 2026-10-03
Tansu Yegen weighs in on a safety-testing framework that treats psychological harm from AI as a measurable quantity, modeled on crash test dummies. A five-person team generating tens of thousands of simulated chats daily is real leverage, he says, but he worries buyers will use the scores as paperwork rather than a hard stop before deployment.
More from Safety
- Former OpenAI safety team member writes in The Atlantic: its culture is broken — michaelas10sk8 · 2026-10-03
- Meta's Muse AI Agent Builds Detailed Profiles of Every Person in Your Life, Hourly — nordicinst · 2026-10-03
- Wired: Meta's AI agent Muse, downloaded millions of times, profiles your friends and family — Wired AI · 2026-10-03
- Singapore reports its first AI-related data breach — xuanalogue · 2026-10-03
- Google updates guidance: fake expert authorship signals pages as low quality — gaganghotra_ · 2026-10-03
- VC accuses OpenAI of skynet-style regulatory capture, says Anthropic's version is more subtle — StewartalsopIII · 2026-10-03