When an AI copies a smarter model, it nails the confident tone — but the real capability stays out of reach.

This explores why students imitating a teacher pick up the teacher's style — tone, verbosity, confidence — even when the underlying reasoning or knowledge doesn't transfer with it. The short version from the corpus: style is the *easiest* thing to copy, and most training objectives reward exactly the surface features that carry it. When a model imitates ChatGPT, it reliably learns to sound confident and fluent while its factual accuracy barely moves — the imitation fools human raters precisely because style transfers cleanly and capability doesn't Can imitating ChatGPT fool evaluators into thinking models improved?. Style rides on the surface tokens; competence lives in the base model, which imitation can't lift.

A deeper reason is that "style" isn't cosmetic — it's a real, geometrically distinct pattern in the model's activations. Verbose versus concise chain-of-thought occupy separate regions of activation space, so much so that you can extract a single steering vector from ~50 examples and slide a model between long and short reasoning without any retraining Can we steer reasoning toward brevity without retraining?. If verbosity is a linear direction, then a teacher whose traces cluster in one region will drag the student's activations toward that same region during distillation. Style transfers because it's a low-dimensional, learnable signal — often easier to fit than the messy, high-dimensional structure of correct reasoning.

The catch is that copying style also copies its baggage. Teachers conditioned on the correct answer and verifier output produce confident, concise traces — and students inherit that confidence, including the suppression of uncertainty. The result looks great in-domain but generalizes worse on out-of-distribution problems that need epistemic caution Does richer teacher context hurt student generalization?. This is the flip side of style transfer: you don't just inherit *how* the teacher talks, you inherit *what it stopped saying*.

Why don't we just prevent this? Because standard objectives don't measure style, so they can't protect it. Post-training faithfully drives models toward correct answers while silently degrading unmeasured behaviors like verbalized uncertainty — single-objective optimization creates blind spots where stylistic features critical to generalization go unguarded Can post-training objectives preserve reasoning style alongside correctness?. Style transfers unmanaged in *both* directions: desirable traits and undesirable ones move together because nothing in the loss function distinguishes them.

The hopeful thread is that distillation done carefully steers style rather than suffering it. On-policy distillation works as an *exploration catalyst* — nudging a student toward good reasoning paths inside its existing capability envelope, where signal quality matters far more than teacher size Does on-policy distillation actually expand student capability?. Push that further with a deliberately *close* teacher bounded by a trust region and you get stable, monotonic improvement instead of the student lurching toward a distant target's idiosyncrasies Can proximity between teacher and student fix distillation instability?. The reframe worth leaving with: style transfer isn't a bug in distillation, it's the default channel — the surface is what copies first, so the whole game is controlling what the surface is allowed to carry.