A new study finds language models playing historical characters still internally register their in-character false statements as false — they're acting, not converting.
Researchers tested three model families — Qwen 2.5 14B, Qwen 3 8B, and Llama 3.3 70B — using linear truth probes, which measure internal representations rather than outputs. The setup: prompt a model to roleplay a historical figure whose beliefs contradict modern science, then compare how it handles claims that figure would have endorsed against equally false claims they wouldn't have made. Across prompting styles, in-context learning, and supervised fine-tuning, adopting a historical persona suppressed era-believed statements less than other false alternatives — but the probes still classified them all as false. The model is performing, not believing.
That distinction carries real weight for AI safety. The researchers compared roleplay to a separate phenomenon called Emergent Misalignment, documented in models trained on harmful content: those models' false claims move substantially toward the "true" region in probe space, and they defend incorrect positions under challenge about half the time — versus roughly one in six for roleplay models. The two phenomena, in the researchers' framing, are points on a spectrum of belief internalization.
The comfortable reading: playing Aristotle doesn't make a model a geocentrist at the representational level. The less comfortable reading: training a model on harmful content apparently can shift what it internally represents as true. That's the finding the AI safety community should be watching.