Scientists Asked AI to Impersonate 112 Public Figures. What Happened Next Is a ‘Dire’ Warning

The study's most underreported detail is the direction of the gap: audiences didn't merely fail to distinguish AI from human, they actively preferred the synthetic output on credibility metrics. That's not a detection problem, it's a preference problem, and no watermarking or disclosure regime is designed to address it.

This lands directly alongside the groupthink coverage from MIT Technology Review ('LLMs are stuck in a groupthink groove'), which showed that major models cluster toward consensus, predictable outputs. If AI-generated speech is simultaneously more coherent and more preferred by audiences, that clustering behavior may be exactly why: models optimize for palatability in ways human speakers don't. Separately, the Anthropic regulatory arc covered across multiple stories this week (the Fable and Mythos clearances, the hidden monitoring logic in Claude Code) illustrates how governance frameworks are still focused on capability gating and safety testing. None of those mechanisms touch the downstream perception problem this study identifies. The policy apparatus is solving for a different threat model.

Watch whether any of the major AI governance proposals currently in legislative circulation, particularly in the EU AI Act implementation guidance, add requirements around disclosure of synthetic political speech specifically. If they don't address the preference inversion finding within the next two revision cycles, the regulatory response will remain structurally mismatched to the actual risk.