Writing without surrendering your voice

The easiest AI writing feature is a button labelled “improve”. It is also a small act of erasure. The prose comes back smoother, more generic and strangely certain. The author has to reverse-engineer which judgements survived.

A serious writing tool should expose the change as a proposal. What was altered? Which claim needs a source? Is this sentence clearer, or merely more familiar to the model? Can the author reject one change without losing the rest?

Voice is not a tone preset. It is a history of choices: sentence length, tolerated ambiguity, preferred evidence, words the author refuses to use. A useful memory system can learn those conventions, but it should save them with approval and keep them attached to the project rather than turn them into a mysterious global persona.

Provenance matters too. Citations need structured records, not links sprinkled into generated text. Export needs to preserve headings, notes and changes because the document will eventually leave the tool. The writer should be able to see what the model contributed and what remains unresolved.

Miton places Writer beside project Chat because writing is often where a project’s reasoning becomes legible to other people. The model can research, challenge, outline and revise. The author still owns the sentences.

The goal is not prose that sounds like AI has been removed. It is prose in which the person was never removed in the first place.