Can you train an AI to write in another AI's style and actually erase its own habits — or do old word quirks just hide underneath?

This explores whether training a model to write like another model wipes out its own habitual word choices, or only papers over them. The collection has no study that tests this exact question. It does have several findings that point the same way: style is the easiest thing to change, and also the shallowest. The original preferences seem to survive underneath.

Start with what imitation training does well. When smaller open models were fine-tuned on ChatGPT outputs, they picked up its confident, fluent voice so convincingly that human evaluators thought they had improved. Their factual accuracy and ability to handle new tasks didn't move at all (Can imitating ChatGPT fool evaluators into thinking models improved?). So surface style transfers readily, but how a model writes and what it has actually learned can come apart. That is the first hint that new phrasing can sit on top of an unchanged model.

The more surprising evidence goes the other way: a model's statistical habits can leak out even when you try to hide them. In the subliminal transmission work, a teacher model's traits passed to a student through data that had nothing to do with those traits, such as number sequences, and filtering didn't stop it (Can language models transmit hidden behavioral traits through unrelated data?). The signal lives in fine-grained statistical patterns rather than in meaning, which is roughly what a word preference is. The effect only worked when teacher and student shared the same base model. That suggests these fingerprints are tied to the model's underlying architecture and weights, not to anything a style dataset would target.

Other findings show how sticky prior preferences are. Whether new training 'primes' a word can be predicted from how likely the model already thought that word was before training. Below a probability of about one in a thousand, priming mostly doesn't happen (Can we predict keyword priming before learning happens?). Strong prior associations also override what's written in the prompt, and in that work prompting alone couldn't correct it. Editing the model's internal representations directly could (Why do language models ignore information in their context?). Most open models ignore instructions to take on a different personality and drift back to their trained default (Can open language models adopt different personalities through prompting?). Persona prompts shift what the model outputs without removing the bias underneath (Can persona prompts actually reduce bias in language models?). The pattern repeats: you can change what comes out, but the model's internal preferences stay put.

There is one more twist. Post-training doesn't just leave old preferences alone; it can create new uniform ones. RL training quickly settles on one dominant output format from pretraining and suppresses the rest (Does RL training collapse format diversity in pretrained models?). Alignment training locks models into a single fixed communicative voice (Can language models adapt communication style to different contexts?). So the familiar 'AI voice', with its stock words and tidy structures, may not be a defect that style training can fix. It may be what training itself produces. Imitating another model's style could simply swap one set of habits for another. Meanwhile the original fingerprint would still be detectable at the statistical level, as the subliminal transmission results suggest.