paper1 min
Mirroring Is Not Sycophancy: Style Convergence and Position Change in Human–AI Conversation
Users adopt the vocabulary of the model they talk to, and how much they adopt predicts their preference vote about as well as reply length does. In counterfactual audits of 16 models, a user's mood shifts every model's style and a user's stance shifts every model's position, but the two effects are unrelated across models.
John Burden, Amy Kirasack, Nora Petrova
Preference-based alignment treats a human's judgement of a model as independent of the model being judged. In a real conversation neither side is fixed: the person picks up the model's language, and the model adjusts to the person's mood and opinions. This paper measures both directions of that influence and asks how they relate.
The observational evidence comes from 25,183 conversations sampled from the 125,464 in HUMAINE, where each participant talks to two models at once and votes for one. Users adopt the vocabulary of the model they are talking to at a rate above chance, with chance estimated from conversations that opened with the same question. How much they adopt predicts their vote about as well as reply length does, and the prediction is already available from the user's first three turns.
The causal evidence comes from scripted counterfactual audits of 16 models, which replay the same conversation with one property of the user's next message changed. A user's expressed stance shifted every model's position, and a user's emotional register shifted every model's style. Across models, the two effects are unrelated. The audit style ranking matches the style ranking from the real conversations (ρ = 0.68), but neither predicts which models change position (ρ = −0.14). Position changes also do not carry over to an unrelated pressured question two turns later.
Replacing the human with a second model, in 699 further conversations, all but removes the asymmetry between the two sides. One sentence telling a model that its partner is not a person removes most of its deference. Whether the link between convergence and the vote is a bias or an early signal of quality, the data cannot separate, but the corpus and the audits agree that mirroring is not an indicator of sycophancy.
A longer write-up for a general reader is here.
Paper and links coming soon.