Voice drift

Voice drift is the slow loss of a voice over the length of a draft. Chapter one sounds like the book you meant to write. By chapter twenty, the sentences are smoother, safer, and somehow anonymous, and a character who started sharp now talks like everyone else. Nothing broke in an obvious way. The voice just leaked out a little at a time.

It is one of the signature problems of long-form AI writing. A model writing chapter twenty does not truly remember how chapter one sounded, so without an anchor it reaches for its default register: fluent, even, and generic. The result reads as competent and forgettable, which is the texture of AI slop. The same thing happens to a character's dialogue, where a distinct way of speaking erodes into a neutral one.

Why drift is hard to see

Drift is gradual, which is what makes it dangerous. Any single chapter looks fine in isolation. The problem only appears when you read chapter one and chapter twenty back to back and hear how far the voice has traveled. By then you are editing a whole manuscript instead of correcting a paragraph.

How FireQuill catches it

FireQuill fights drift at the source and at the check. Every chapter is written against samples of the voice you set, so the model is never working from memory alone, and the character engine keeps each character's way of speaking in view. A voice check then compares new prose against your samples and flags the slide while it is still a paragraph, not a draft.

See how to set and hold a voice in how to make AI write in your voice.