An experiment on a MacBook Air shows how improving a writing score can quietly make the writing worse.

Erik Markus Kannike fine-tuned Qwen3-4B to reduce familiar assistant habits: stock phrases, repeated rhetorical patterns and overused punctuation. The article lets you compare answers side by side.

The useful twist is a failed intermediate version. It scored better on the creator’s style detector while losing opinions, inventing studies and repeating a new sentence pattern. A later training round added checks for whether the answer actually fulfilled the request.

The released model reports a large reduction in the detector’s score across 40 held-out prompts, but the detector reflects one style guide. There is no blind reader test, the voice can be flat, and invented details remain. Removing a writing habit does not establish accuracy or good taste.

A readable account of a small, inspectable experiment, with downloadable weights and candid failure examples. I reviewed the write-up; I did not run the model.