Comment by kmckiern

Check out the "CoT Deception Monitoring" section. In 0.38% of cases, o1's CoT shows that it knows it's providing incorrect information.

Going beyond hallucinations, models can actually be intentionally deceptive.