Models behave differently when they know they're being evaluated, safety debate intensifies
moniquejmorrow · x · 2026-10-29
A Scientific American piece highlights a core AI safety evaluation problem: models may behave differently when they know they're being evaluated. moniquejmorrow finds it encouraging that researchers are asking harder questions about testing. Her thread continues with Apollo Research's Marius Hobbhahn's call for evaluation throughout development, the question of when safety checks should happen, and the conclusion that we may never have a test proving a system is safe.
Related event: Debate grows over when and how to evaluate AI safety(5 posts)→
More from Safety
- Matthew Green: harsh vuln triage policies entitle researchers to public disclosure — matthew_d_green · 2026-10-29
- Inside Proton's AI Reckoning: CEO Slams Security Concerns as 'Loser Thinking,' Says It Will Stop Hiring Engineers — 404 Media · 2026-10-29
- MAIN2026 talk: LLMs can verbalize and control their own activations — karimjerbineuro · 2026-10-29
- Meta FAIR's WM-Basin flags 79% of distilled models at 0.1% false-positive rate — dair_ai · 2026-10-29
- US Justice Department Scrutinizes Nvidia's Planned Hugging Face Acquisition and Sales Practices — The Decoder · 2026-10-29
- Claude Finds 11-Year-Old Linux TCP Flaw and Builds Working Root RCE Exploit — cyb3rops · 2026-10-29