Skip to content
Topic

AI and Writing

Drafting, editing, rewriting and the detection question — what AI genuinely changes about writing, and the parts that stay yours.

Writing is where most people meet AI first and where the gap between the demonstrations and the reality is widest. The tools draft quickly and the drafts arrive looking finished, which is the trap: what they produce is competent, structurally flat, and recognisable to a reader within a paragraph. This hub covers what actually works — using AI for the research and reformatting around a piece rather than for the argument inside it — and the surrounding questions of detection, disclosure and rewriting that anyone writing with these tools runs into within a week.

Start with what AI is bad at

The useful mental model is that these systems produce likely text, not considered text. That makes them strong at anything where the raw material exists and the job is transformation — summarising, reformatting, adapting for a different audience, drafting a version to react to. It makes them weak at exactly the thing that makes writing worth reading: a point of view, a specific detail from experience, and the willingness to assert something. Almost every disappointing result traces back to asking for the second when the tool is good at the first.

The workflow that holds up

Use AI heavily before and after the writing, and lightly during it. Before: research, structuring, finding what has already been said so you can say something else. After: editing questions, checking what is unclear, reformatting for each place the piece will appear. During, for anything where voice carries the value, it is faster to write it yourself than to edit a draft into sounding like you. The exception is the work where structure matters more than voice — the explainer, the roundup, the summary — where a generated draft is a genuine head start.

Editing is the whole skill

The difference between people who get value from AI writing and people who produce filler is entirely in the editing pass, and it is a short list of moves: vary the rhythm, commit to a claim, cut the hedging and the throat-clearing, add the specifics only you have, and check every fact the draft asserted. That pass takes about twenty minutes on a thousand words and it is not optional — a draft published without it reads exactly like what it is, and one obviously generated piece does more damage to a site than three good ones do good.

Detection, disclosure and the honest position

Anyone writing with these tools eventually meets the detection question, and the honest answer is uncomfortable for everyone: detectors measure how text reads rather than who wrote it, they are wrong often, and their errors fall hardest on people writing in a second language. That is an argument about enforcement rather than permission. The separate question — whether you are allowed to use AI here, and whether you have to say so — is a policy question worth answering before you need to, because the rules vary and nobody accepts a favourable assumption after the fact.

The short version: let AI do the work around the writing, keep the argument and the specifics yours, and never publish a draft you have not substantially edited. If you read one thing here, make it the editing checklist — it is the step that separates useful output from the filler everyone can already recognise.

Try it yourself

The free plan includes 100 credits and needs no card.