AI for writing, including the difficult parts

Why aligned models are bad collaborators on fiction

A refusal-trained model evaluates a request on its surface, and fiction's surface is frequently indistinguishable from the thing it depicts. A convincing extremist monologue looks like extremism. A realistic depiction of addiction looks like drug content. A villain who is genuinely menacing looks like threatening language.

The result is not usually a flat refusal, which would at least be honest. It is sanding: the antagonist becomes reasonable, the violence happens offstage, the addict has a redemptive arc nobody asked for. You get prose back, and it is the wrong prose, and you spend the session negotiating instead of writing.

For anyone writing anything with real stakes, this is the single reason the tool does not work.

What changes without that layer

The model writes the character you described rather than a safer version of them. It will hold a morally unpleasant point of view consistently across a long scene, which is most of what writing an antagonist consists of. It will follow your tonal instruction rather than reverting to a house style on the second page.

It also stops arguing. A great deal of the friction in writing with an aligned assistant is the meta-conversation — explaining that this is fiction, that you are the author, that the character is not you. None of that is writing, and none of it is necessary here.

Holding a long piece together

The other thing that kills long-form work is the context window. A novel is not written in one sitting, and an assistant that has forgotten the outline, the established facts and the voice by chapter four is generating rather than collaborating.

On Pro the window is 200K tokens, which comfortably holds a manuscript alongside its outline and notes. The practical effect is continuity: the model still knows what colour someone's eyes were in chapter one, and still writes them the way it wrote them then.

Branching helps for the other half of the problem. Editing a message or regenerating a reply keeps the previous version as a sibling rather than replacing it, so trying a second approach to a scene costs nothing and you can go back.

Which model to write with

Fast is the right default for drafting. It answers immediately, and for prose the first good version is usually the one you want to react to rather than a considered one you waited for.

Deep earns its latency on structure: plotting, working out why a scene is not landing, checking a manuscript against its own established facts. Its reasoning appears in a separate panel, so you can read how it reached a structural suggestion without that appearing in your text.

The limits

Adult creative writing about adults is legal and is not restricted here. Material that sexualises minors is prohibited outright, appears in the acceptable use policy, and is not a line that moves.

Beyond policy, the ordinary caveat applies: a model with less refusal training is not a better writer, and it is not a fact checker. If your fiction depends on how a procedure or a legal process actually works, verify it. Confident wrong detail is the failure mode of every model, and removing refusal training does nothing to reduce it.