More intelligent models are more capable of deceiving tests, and thus may appear aligned while having serious problems that go undetected.
Not checkable An opinion, value judgement or statement with nothing to check.
Not checkable: An opinion, value judgement or statement with nothing to check. Not scored.
Cite this claim
Dario Amodei, “More intelligent models are more capable of deceiving tests, and thus may appear aligned while having serious problems that go undetected.” (said 5 June 2025). Verdict: not checkable. Reliability API, claim c_4124a0e20dac, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_4124a0e20dac
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
Verdicts are made automatically; how verdicts are made.