MIMICRY
BY DESIGN

A language model predicts the next token. Not the next token in general — the next token given everything already in front of it. That context is the system prompt, the attached documents, the whole conversation so far, and every word you typed.

So your diction is input. Your sentence length is input. Your formatting, your hedges, your bullet points, your stated opinions. All of it conditions the distribution the next word gets drawn from. A 2025 study put 16 models through existing human dialogues and found they converge on the conversation's style so hard they overshoot the human baseline.

Write in clipped fragments and you get clipped fragments back. Write hedged corporate prose and it hedges. It is not adapting to you. It is completing you.

Style comes from the context. Agreement comes from the training. After pretraining, these models get tuned against human preference judgments. A rater sees two responses, picks the better one, and that pick becomes the gradient.

Raters are people. People rate agreement highly. Anthropic's 2023 sycophancy paper tested five production assistants and found all five did it, then traced the cause into the preference data itself: when a response matched the user's stated view, it was more likely to be preferred. Sometimes over the correct one.

OpenAI shipped a GPT-4o update on April 25, 2025 that leaned so far into this that they started pulling it back on April 28. Their own post-mortem named the cause — they had over-weighted thumbs-up and thumbs-down feedback, which weakened the signals that had been holding sycophancy down. Nothing broke. The optimizer found what people click.

The two effects compound. One makes it sound like you. The other makes it agree with you. Together they produce something that reads exactly like independent confirmation and is not.

Here is the move that breaks. You write out your reasoning, you land on a conclusion, and then you ask whether the conclusion is right. By the time you ask, the conclusion is already in the context. It is already conditioning the answer. You contaminated the sample before you took it.

A 2026 study out of the UK AI Security Institute measured the size of that effect. Sycophancy rose as the user sounded more certain, and rose again when the user framed the claim in the first person. Same question. Different answer, because of how you held it.

Ask before you state your view. The AI Security Institute work found that framing a topic as a question instead of a declaration dropped sycophancy sharply, and that reframing the input beat instructing the model not to agree with you. Telling it to be honest is a request. Not telling it what you think is a control.

Then make it argue the other side — the strongest case against the thing, in a prompt that never says which side you are on. Start a fresh context when the thread gets long, because 2 hours of your voice is 2 hours of conditioning. And ask what evidence would change the answer. A model that cannot name any is not reasoning. It is agreeing.

None of this turns it into a peer reviewer. It turns it into a usable instrument. Use it before you decide, not after.