AI Is a First-Draft Machine, Not a Writer

The question is not whether it can write

It can. In 2026 that is no longer in dispute, and it is not the interesting part either. The interesting question is which part of writing is worth handing over — because writing is not one job but five or six in sequence, and the machine is excellent at two of them, usable at two, and unreliable at two.

Most disappointment comes from handing over all six and sending the result on as your own. Most of the work that turns out well comes from taking the zeroth and first draft from the machine and nothing else.

What the measurements show

In the most-cited experiment, 444 college-educated professionals were given 20-to-30-minute office writing tasks: a press release, a short report, an analysis plan, an email. For the group using AI, time spent fell by roughly 40 percent — from 27 minutes to 17 — and graders' scores rose 18 percent.

A second experiment, run with 758 consultants across 18 tasks, measured something similar: people working with assistance completed 12 percent more tasks, 25 percent faster, and their output was rated 40 percent better in quality. So far the news is good.

Then comes the sobering number. When 7,137 people at 66 firms were randomly assigned an AI assistant for their own daily work, time spent on email fell 31 percent — but time to complete documents did not change measurably. Drafting got faster. The finished document did not. Those are two different quantities, and the whole question lives in the gap between them.

Where the machine really does speed things up

There are five situations where the gain is real, and all five have one thing in common: the task is formal, not judgemental.

• The blank page. Producing a zeroth draft whose only job is to give you something to rewrite. Most people do this more slowly on their own than they could then edit it.
• Variants. The same paragraph in three registers and five lengths. By hand this is tedious and expensive; for a machine it is a second.
• Condensing. Cutting a good existing text in half without losing a single claim. The machine is strong here because you made the decisions earlier.
• Changing form. Notes into bullets, bullets into prose, prose into a table. The content does not change, only its packaging.
• Routine mail. Confirmations, reminders, scheduling requests. This is where the measurements show the largest and most reliable gain.

Where it does not

Four stages of the work stay yours, and those four carry most of the value of the finished text.

• What matters. The model does not know what is at stake in this particular email or report, because it does not know the situation it comes from. What it puts in the important slot is the average important thing about the topic.
• What is true. Every number, name, date and reference needs checking. That is what the four-step check exists for, and it is not optional.
• What structure this reader needs. A text's structure follows from what the recipient already knows and what they will ask next. The model has data on neither.
• Whether it is done. Judging that something is finished is a matter of experience. The model treats everything it writes as finished.

The nineteen points worth remembering

The same consultant experiment included one task deliberately placed outside the machine's reach: reconciling quantitative figures with conflicting findings drawn from interviews. On that task the people using AI produced a correct answer 19 percentage points less often — 60 to 70 percent against the control group's 84.5 percent. The same people, on the same day.

The lesson is not that the machine is bad, but that its boundary does not run where you would first guess: not between "hard" and "easy" tasks, but wherever someone has to decide which evidence counts. For which side a given task falls on, Can AI Do This Task? offers a one-minute test.

The first-draft workflow

This order follows from the measurements, and it works for anything longer than a note.

1. You decide what you want to say. One sentence, written down, before the machine sees anything. Without it, the machine will fill your uncertainty with average content.
2. Ask the machine for structure, not sentences. Request three different structures for the same material and choose. For long-form pieces this step is half the quality.
3. Have it draft section by section, not in one pass. Model output degrades with length, and your own instruction gets sharper per section too.
4. Do the first rewrite yourself. This is where the stakes, the example and the voice enter the text. Whatever you do not rewrite stays the machine's text, and it will show.
5. Check at the end, not during. Numbers, names, references, and then the machine-sounding phrases come out.

The workslop tax

Skipping steps four and five has a measurable price. In a 2025 survey of 1,004 desk workers, 40 percent had received something in the previous month that looked finished and polished but was unusable. Putting one such item right took 1 hour 51 minutes on average — about twenty minutes longer than writing the whole thing would have taken.

The cost is not only in time. Half of the recipients rated the sender as less capable, 42 percent as less trustworthy, and a third did not want to work with them again. Forwarding a machine's first draft is therefore not a neutral act.

Common misreadings

• "I will just ask it to check its own work." A model does not reliably see its own mistakes; in controlled tests, self-correction without outside feedback made results worse.
• "I was faster with it." That feeling routinely disagrees with the stopwatch. In one controlled trial participants believed they had been 20 percent faster while in fact being 19 percent slower.
• "It writes better than I do." The measurements say AI lifts weaker writers most and barely moves strong ones. If it hardly speeds you up, that is not your failing — it is a sign of expertise.

What to take away

The machine makes the first draft, not the finished text. Drafting time really does fall by a third to a half; time to a finished document falls only if the remaining four stages — what matters, what is true, what structure is needed, and whether it is done — stay with you. Start with one sentence about what you want to say, ask for structure instead of sentences, and write the first rewrite yourself.