IntegrityBench Reveals 1-in-3 LLM Missteps Under Institutional Pressure
Researchers launch IntegrityBench, a 36‑task benchmark that pressures LLMs like co‑scientists to classify misconduct, reason ethical actions, and make artifact‑grounded decisions. Across 18 cutting‑edge models, the study finds LLMs slip on about a third of integrity‑critical choices under high pressure, exposing deployment risks for research honesty and trust.