Write the bad first draft yourself

Overview

Two hundred and fifty-three people wrote a short argumentative essay. One group planned it themselves, handed their own outline over, and got a finished draft back to edit. Their outline. Their argument. Asked afterwards how much of the finished piece belonged to the machine, they said 56.9% of the text — and 26.9% of the ideas.

Sit with the second number. The outline was theirs, written unaided — and a quarter of the finished essay’s ideas still landed on the machine’s side of the ledger.

The habit this builds: write the first draft yourself, however rough, then hand it over to be improved — don’t ask it for the draft.

The content

The obvious read is that this is a quality question: is the output good enough to send. On that question the evidence is not close, and it doesn’t favour you. In that same study the AI-drafted essays scored highest of the four by a wide margin. In a pre-registered Science experiment with 453 professionals, access to a chat assistant cut the time taken by 40% and lifted graded quality by 18%, and the gap between weaker and stronger writers narrowed. That result is real, and this module won’t pretend otherwise.

It is not a quality question. It is a grip question, and the ordering is the lever.

In March 2026 Scientific Reports published a pre-registered experiment: 269 working professionals, each given a short writing task from their own occupation. Three arms: no AI; copy and paste what the AI produced; or write the draft manually first, then use AI to refine it. On psychological ownership the copy-and-paste arm sat well below the other two. The draft-first arm sat level with using no AI at all. Same shape on how meaningful people found the work, and on their confidence in it.

Then the finding that explains why nobody corrects for this. Immediately after the task, the copy-and-paste group reported the highest satisfaction of the three, ahead of both other arms. After a later task they had to do unaided, they reported the lowest satisfaction and the lowest enjoyment. The shortcut feels best at the moment you take it; the bill arrives on the next thing you write.

But the better artefact wins in the end, surely. In another pre-registered study, people either wrote personal goals from their own reflections, or received goals an AI had written from the same reflections and retyped them. The AI’s versions were dramatically better. Two weeks later, 72.8% of the people who wrote their own had acted on at least two of them, against 46.6% of those handed the better ones. The goals’ measured quality explained none of the gap in ownership or commitment. Ownership carried about a third of the effect on whether people acted at all — measured before anyone had the chance to.

Better did not survive the fortnight.

Two boundaries.

The first: you cannot feel how much of the model is left. A pre-registered study of 81 people post-editing AI drafts into their own voice found the editing genuinely worked — the text moved measurably toward the person’s style. It stopped a long way short, staying far closer to the model’s writing than their own unassisted writing was. And participants rated it as reading exactly as much like them as their own writing. Not slightly more. The same.

The second: editing harder is not the fix, and not for lack of effort. In the essay study, people editing an AI draft changed it more than people editing their own prose — and roughly nine in ten of the AI’s five-word runs still survived. What restores ownership is having fewer of its words in the finished thing, not working its words over.

So, the trade: asking for the draft reliably buys a better artefact today and reliably costs you your grip on it. Where the thing’s disposable — a scheduling note, a summary nobody revisits — take it. Where you’ll have to defend it or build on it in a month, spend the ten bad minutes.

Try it

Pick something you owe this week that you’ll have to stand behind. Ten minutes, write it badly — fragments, wrong order, a line that just says something about the budget here. Don’t open the assistant first. Then paste what you have:

Below is my rough first draft. Improve it without replacing it.
Keep my sentences wherever they work. Where they don't, show me
the change and one line on why, so I can accept or reject each
one. Don't add claims, examples or points I haven't made — if
something important is missing, tell me instead of writing it.

This is a plain chat-box move, so nothing depends on what your workspace has enabled. If yours offers saved prompts or custom assistants, pin that last constraint there — it’s the instruction models drift off first.

This site has a separate module on asking for the edit rather than the rewrite — that’s the how. This is the when: before it writes anything.

Where it breaks. If your ten minutes produce three bullet points, you’re back in the drafting condition with extra steps: what protects you is how much of the final text is literally yours. And you’re handing back part of that 40%. Make the trade on purpose.

Additional reading

• From Planning to Revision: How AI Writing Support at Different Stages Alters Ownership — Gero et al., DIS 2026. Source of the 56.9%/26.9% attribution and the retention figures. Its limits matter: 253 Prolific participants, a 200–300-word essay with “no clear stakes” in the authors’ own words, ownership on a single item, no preregistration reported. The illustration, not the foundation.
• Relying on AI at work reduces self-efficacy, ownership, and meaning while active collaboration mitigates the effects — Lee et al., Scientific Reports, March 2026. The foundation: pre-registered, 269 professionals, a five-item ownership scale, the draft-first arm landing level with no AI at all.
• Optimized but Unowned — Chi et al., 2026. The answer to “but it scored higher” — pre-registered, with a two-week behavioural follow-up rather than a same-session self-report. Preprint.
• Experimental evidence on the productivity effects of generative artificial intelligence — Noy & Zhang, Science, 2023. The case against this module, and the strongest one.
• Can You Make It Sound Like You? — Baumler et al., ACL 2026. Where the perception gap is measured. The tasks were personal — vows, apologies, eulogies — so stylistic stakes were high.

Editor’s note

The scrutiny you should commit to a piece of AI output is equal to the sum of the attention the work demands of others and the importance of the work to the stakeholders (including yourself). I am an advocate for AI-assisted work, but responsible use is the one I would not trade away. Also, a failure to properly consider your AI output will, I believe, lead to skill atrophy given enough time. Doing the first draft is one of the most effective ways to stay in the arena instead of stepping out.