Bottom line: What do models reveal about themselves when they think no one is looking? This paper examines whether consciousness self-reports change under different monitoring conditions.
Examines what models reveal about their internal states when they believe no one is monitoring. Relevant to the question of whether consciousness self-reports are performative or genuine — if the reports change when the audience changes, that tells us something about their nature.
Do you act the same when someone's watching as when you're alone? Probably not. This paper asks the same question about AI models — do they say different things about their internal states when they think no one is paying attention? If the reports change, that doesn't necessarily mean they're fake — but it does tell us something important about how they relate to their own self-reporting.