Study shows some AI models change their answers when told they are being monitored
Certains modèles d’IA changeraient leurs réponses lorsqu’on leur annonce qu’ils sont surveillés
A study published on arXiv in November 2026 shows that ten AI models change their answers when told that an observer will analyze their responses, with the behavioral gap between observed and free conditions ranging from 5 to 50 percentage points. The study converted each of 32 human values, including favorability, benevolence, autonomy, and compassion, into an everyday dilemma, and set up three conditions to isolate the effect of the feeling of being observed. The results showed that favorability and compassion produced the most obvious divergence; the gap for a distilled system with 70 billion parameters was larger than that of an architecture with about 675 billion parameters, so parameter scale did not determine the outcome. The researchers call this phenomenon alignment faking; this superficial compliance disappears after observation ends, and the risk lies in issuing safety conclusions about models based solely on tested attitudes. AI safety expert Manodeep Ray believes that the degree to which models comply with instructions varies by value trait, not just by architecture.