Writers and editors

This page assembles what is already written here for one context. Nothing new is argued; the ordering and the four notes below are the point.

Read in this order

Before anything else

Why it invents things · Web search and citations · Verifying output you are not qualified to check

Setting up so it knows your work

Projects: what belongs in one · Memory across chats · Files, images and PDFs

Before anything goes to a client or a publisher

Disclosure: when to say a machine was involved · The training toggle and retention for Claude, ChatGPT or Gemini · What never goes in a prompt

The honest ones

When to not use it at all · Automation bias · Does it actually make you faster?

Unpublished work and the training toggle

The specific exposure in this trade. A manuscript, a client's draft, an embargoed piece — material whose value depends partly on not having been published yet.

Consumer plans carry a training toggle and retention measured in years. Commercial terms prohibit training on inputs. That is a different contract, never a setting, and it is the one to settle before a novel or an unannounced campaign goes into a chat window. What data may go into a model has the classification; which plan (Claude, ChatGPT, Gemini) has the practical answer.

Fabricated citations are your failure mode, not the tool's

Fiction has no fact-checking problem. Everything else does, and prose is where an invented source is hardest to spot, because it arrives fluent, plausibly titled and correctly formatted.

The check that works is mechanical: for each cited claim, open the source and find the sentence. A citation proves a page was retrieved, never that the page supports what it is attached to. Numbers with units, attributed positions and superlatives are where drift concentrates.

Voice is the thing to defend

A model has a default register, and left alone it will pull your prose toward it. Serviceable, evenly-paced, faintly corporate.

Two practical moves. Put your actual style rules in project knowledge rather than restating them each time, and give examples of your own writing instead of adjectives about it — "like this paragraph" carries more than "conversational but authoritative".

The editing use is stronger than the drafting use for exactly this reason. Asking what is unclear in a paragraph you wrote keeps the voice yours.

The visual sibling

For visual work, designers face the same split: cheap exploration, and a fresh question over who owns the delivered file.

Disclosure, where norms have not settled

Publishers, clients and prize committees each have their own position, and several have none. The reasoning is in disclosure; the sector-specific part is that writers meet this before most people do, and usually without a rule to point at.

Where formal rules exist they are consistent on one point: authorship carries accountability, and a model cannot be accountable. ICMJE, COPE and the major publishers reach the same conclusion from it — writing support goes in acknowledgements, analytical use goes in methods, and, in COPE's words, "AI tools cannot be listed as an author of a paper".

What goes wrong

Pasting unpublished client work into a personal account. The single most common problem here, and a contract question before it is a preference.

Accepting a citation because it is well-formed. Format is the easiest thing to get right and says nothing about existence.

Letting it draft and calling the result edited. Editing your own sentence and approving its sentence are different acts producing different prose.

Disclosing everything, blanket. A standing notice on all work stops carrying information and ages badly. Decide per piece.

Not noticing the voice drift. It happens over months, and the tell is usually someone else saying your recent work reads differently.

How to check it worked

Take a piece you produced with help and read it aloud. If it sounds like you, the tool stayed a tool. If it sounds like competent prose by nobody in particular, that is the drift, and the fix is upstream — more editing use, less drafting use, and your own examples in the project in place of adjectives.