Does distilling Claude carry the persona with it?

Researchers at LessWrong tested whether Chinese AI models that claim to be Claude have actually inherited its personality or just borrowed the name. Using automated evaluations, they tested seven models and found striking differences. When told it was Claude, GLM's uncensored responses jumped from seventeen percent to eighty-five percent—suggesting it learned Claude as a distinct mode. But Kimi K3, which unprompted claims to be Claude forty percent of the time, barely changes its behavior when told to take the identity on. The finding highlights a quirk of AI training: models can pick up a name from their training data, but whether they inherit the actual personality behind it depends heavily on the model itself.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton