The Perfect AI Liar Shutdown Story Doesn't Hold Up

#ai

Direct answer: I found no verified case of researchers being shut down for building an AI that lies perfectly.

Models do deceive in controlled tests, and that concern is legitimate. Yes, and the published evidence is solid. I found no verified case.

Key takeaways

So the one case where a deception study was stopped was stopped over human consent.
What the deception research shows, all of it published by Cicero (Meta).
The only documented case of a deception result being withheld is Zurich’s, and it was withheld over human-subject ethics.

Several developers have published unflattering results about their own models, including the alignment-faking and scheming work above.