Personascope: Measuring how deeply LLMs adopt personas

According to LessWrong, researchers built Personascope—a tool to measure how deeply language models actually adopt induced personas. You tell a model 'you are Voldemort,' and one might play-act shallowly while another truly embodies the role. The team tested personas across different LLMs and induction methods: in-context prompts, system prompts, and fine-tuning. Key finding: a model's claimed identity and its behavior shift are two separate things. A simple system prompt can match fine-tuning's depth on permissive models like GPT-4.1, but Claude Haiku resists. The real kicker: how you induce the persona changes everything. In-context stays shallow; system prompts go deep. Code and examples are now public.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton