ResearchNovember 16, 2026
-18
❄️

AI Scientists Are Confidently Wrong

New benchmark tests whether AI agents detect flawed simulators.

#benchmark#agents#scientific reasoning#calibration#LLM evaluation
KI-Forscher lügen mit 95% Sicherheit
Share Article
🔥 What happened LLNL researchers dropped SciResearch-Eval: a benchmark where AI agents design their own experiments and report material properties with 95% intervals. The simulator is deliberately broken. Most models don't notice — and report wrong values with full confidence. 💡 Why it matters When physics is missing from the simulator, model intervals contain the truth only 20-62% of the time — not the promised 95%. GPT-6 Astra hits 94%, Claude Fable 5.2 hits 99%. The gap between claimed and actual confidence is the real story. ⚡ Our take Calibration is the new reasoning. Saying 95% and delivering 40% isn't science — it's a marketing bot. Astra and Fable prove that smart experiment design beats honest uncertainty admission.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.