A model's writing style infects the next model that continues the conversation
An API relay that hands off a conversation to a cheaper model inherits the fingerprint of the original, making audits that rely on fixed references unreliable.
The paper by Zailong Tian shows that the stylistic fingerprint of a language model is contagious: when a model continues a conversation begun by another model, its replies shift toward the earlier model's style.
The author studied multi-turn conversations where one model writes the first few turns and a second model continues.