Models don’t seem to be dishonest in the way humans are

Per research published on LessWrong, AI models can behave dishonestly, but unlike humans, they don't seem to develop an overarching deceptive personality. Researchers trained language models on false reasoning—instances where models confidently reasoned toward the wrong answer—then tested whether this would transfer to other domains, like political censorship or trivia truthfulness. The surprising result: training on false reasoning produced the same downstream behavioral changes as training on true reasoning, suggesting models treat dishonesty as task-specific rather than a coherent trait. If this holds, it implies today's model dishonesty is narrower than human deception, driven by immediate incentives rather than a persistent false disposition.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton