Writing
Also called professional writing, technical writing, communication
Writing is thinking made checkable. A model can produce the prose, which means the thinking is now the only part that was ever yours.
The claim The measurable productivity gain from AI writing tools is real and large — and it lands on the output, while the learning depends entirely on whether you did the thinking before or after the draft appeared.
The evidence here is unusually clean, and it is positive. Noy and Zhang gave 453 college-educated professionals real occupation-specific writing tasks; access to ChatGPT cut time by 40% and raised graded output quality by 18%, and the gains fell hardest on the weaker writers, compressing the spread. If your question is whether these tools produce better documents faster, that question is answered.
The learning question is a different one, and the honest answer is that it depends on the order you do things in. Writing is the rare skill where the artefact and the cognition are the same act — you find out what you think by trying to say it. A draft that appears before you have thought hands you the fluency and keeps the thinking, and fluency without thinking is exactly what a reader cannot distinguish, which is why it is so comfortable to accept.
The suggestive evidence points the same way, with appropriate caution about how thin it is. Kosmyna's MIT team recorded EEG while people wrote essays with an LLM, with a search engine and with nothing; the LLM group showed the weakest neural connectivity and, in a later session with the tool removed, remembered less of what they had written. That is a preprint, 54 participants with only 18 in the crossover session, measuring brain connectivity and recall rather than writing ability — read it as a reason to be careful, never as a finding about your career.
The survey evidence from workplaces is more mundane and probably more relevant: Lee and colleagues found that among 319 knowledge workers, higher confidence in the tool predicted less critical thinking, higher confidence in oneself predicted more, and the job shifted from producing to verifying. That shift is the one to prepare for. Verifying is a skill, it is harder than producing, and nobody assigns it as homework.
So the path below keeps one rule at every stage: you write the thinking, the model edits the prose. That order preserves the only thing writing was ever for while collecting most of the speed the studies measured.
What actually changed
-
WasThe blank page forced you to find out what you thought, because nothing else would fill it.
NowThe page is never blank, so the finding-out has to be scheduled deliberately — usually as a rough outline written before you open the tool.
-
WasFluent prose was weak evidence of clear thinking, and readers used it as a proxy.
NowIt is no evidence at all, so the value moved to what only you could know: the argument, the example from your own experience, the thing you are willing to be wrong about in public.
-
WasEditing was a skill you developed slowly, by rereading your own bad sentences.
NowIt is the primary skill and it is needed on day one, because the first draft in front of you is competent, generic, and subtly not about your situation.
-
WasWeak writers stayed weak until they practised for years.
NowNoy and Zhang measured the gap between strong and weak writers narrowing immediately — on output. Whether the weaker writer improved, or simply produced better documents, is not the same question.
The path
In order. Each stage names what the model may and may not do while you are in it — that rule is the difference between the two arms of every study at the bottom of this page.
-
1 Finish bad things
Break the association between writing and paralysis by producing complete, honest, unedited drafts.
- Write 300 words a day for four weeks, cold, and finish every one. Publish some of them somewhere small.
- Set a 25-minute timer and do not stop moving. Bad sentences are allowed; stopping is not.
- Write about things you actually know. The first thing to disappear from AI-assisted prose is the specific detail only you have.
What AI may do here Off, entirely, during the drafting window. This is the one stage where the tool competes directly with the skill being trained, in the same way an unrestricted tutor competed with maths practice in the PNAS trial and left students worse off once it was withdrawn.
You are done when You can produce 500 finished words on a familiar topic in 25 minutes without stopping.
-
2 Learn to see sentences
Acquire the technical vocabulary for why one sentence is better than another, so revision stops being guesswork.
- Read Williams on clarity — subject-verb-object, characters as subjects, actions as verbs — and apply one lesson per day to something you already wrote.
- Take a paragraph you like and rewrite it three ways, then say what each version costs.
- Cut 20% from a finished piece without losing content. Do it until it stops hurting.
- Read your work aloud. Every sentence you stumble on is a sentence to rebuild.
What AI may do here Ask it to name what changed between your two versions, after you have made both. Do not ask it which is better before you have decided — the answer is fluent, plausible, and it terminates your own judgement mid-formation.
You are done when You can explain, in technical terms, why a specific sentence of yours is weak — and fix it without a tool.
-
3 Argue in public
Learn structure by writing things that have to hold up to a reader who disagrees.
- Write pieces with a claim in the first paragraph and the strongest counter-argument in the middle.
- Publish somewhere with real readers, and read the responses that annoy you twice.
- Steal structures: take a piece that worked on you, outline it, and pour your own material into the outline.
- Write the one-sentence version first. If you cannot, the piece is not ready.
What AI may do here Use it as the opposition — ask for the best case against your argument, then answer it in your own words. That is elaborative interrogation and it is one of the genuinely strong uses of these tools for learning.
You are done when Somebody has changed their mind, argued back in detail, or acted because of something you wrote.
-
4 Edit at speed, including a machine's prose
Turn the productivity gain into a habit that does not cost you the thinking.
- Write the outline and the argument yourself; let the model draft only sections whose content you have already fixed.
- Edit its output as hostile territory: delete every sentence that would be true of anyone, restore the specifics, check every fact and every quote.
- Keep a list of the tells you keep having to remove. It is a description of the median of the internet, and it is what you are competing with.
- Once a month, write something end to end with the tool off, to find out what has atrophied.
What AI may do here It edits your prose; it does not decide your point. If you cannot say what the piece claims before the draft exists, you are not writing, you are commissioning.
You are done when A reader who knows your voice cannot tell which parts were drafted, because none of the generic parts survived.
Keep these unaided
The gap between what you can do with the tool and what you can do without it is the only honest measure of whether you learned anything. Check it occasionally, on purpose.
- Produce a complete first draft cold, in one sitting, on something you know.
- State the claim of a piece in one sentence before writing it.
- Cut a fifth of your own words without losing meaning.
- Name the technical fault in a bad sentence and fix it.
- Notice when a paragraph is fluent and empty — including when you wrote it.
Where the model genuinely helps you learn
- As a line editor on prose you wrote: it is fast, tireless, and it will not defend your favourite sentence.
- As an opponent: ask for the strongest objection to your argument, then answer it yourself.
- As a structure critic: paste your outline and ask what a reader would still be confused about after each section.
- As a translator of register — the same content for an expert and for a stranger — which teaches you what the register was actually doing.
- As a reading partner on difficult source material you then go back and read properly, which is faster than skipping the source and not the same as skipping it.
Traps
- Commissioning instead of writing. The output is fine, the piece says nothing only you could say, and the skill quietly does not develop.
- Fluency mistaken for thought — by you. The generic paragraph reads well precisely because it has been read a million times, and your ear approves of what it recognises.
- Losing the specific. The example from your own week is the only part a model could not have produced, and it is the first casualty of a rewrite pass.
- Editing down to the median. Every round of make this better moves the piece toward the average of everything ever written, and the average is not why anyone reads.
- Skipping verification. Names, numbers, quotes and citations are exactly what these systems fabricate most convincingly; a fabricated citation under your name is your mistake.
What counts as proof now
The artefact stopped being evidence — anyone can generate one. These are the things that still say something about you.
- A body of published work with your name on it that reads as one person.
- A piece that changed a decision, with the response it drew.
- A before-and-after edit of somebody else's paragraph, with the reasons.
- Something written under time pressure, unaided, that stands up.
Where to learn it
Books
-
The most technically useful book on English prose: it explains WHY a sentence is unclear in terms you can act on. Dry, academic, and worth more than any ten style guides.
-
An attack on the habits schooling installs, written in fragments. Divisive; either it rearranges how you build sentences or it irritates you. Read twenty pages before buying.
- On Writing Well Paid
The standard on non-fiction clarity — simplicity, clutter, the audience. Some examples have aged; the principles have not.
Free reading and references
- Paul Graham's essays Free
A large body of plain, argumentative prose to study structurally — read them for how the argument is built, not only for what they claim. Strong opinions; treat the content as a position, not a source.
- How to write better Free
A practical free handbook on non-fiction structure and idea generation. Aimed at online essays; light on sentence-level craft, so pair it with Williams.
- Purdue OWL Free
The free reference for mechanics, citation and academic formats. A lookup tool, not a course — go when you have a specific question.
- PlainLanguage.gov Free
Guidelines and before-and-after examples for writing that people can act on. The examples are the useful part, especially for anyone writing interfaces, policies or instructions.
The evidence
What each source measured, and what kind of source it is. A practitioner study and a randomised trial are both worth reading and are not the same claim.
- Noy and Zhang (2023), Science Peer-reviewed
453 college-educated professionals on occupation-specific writing tasks: ChatGPT access cut time taken by 40% and raised graded quality by 18%, with the largest gains for the weakest writers, narrowing the spread between them.
- Lee et al. (2025), CHI Peer-reviewed
319 knowledge workers, 936 real examples: confidence in the tool predicted less critical thinking, confidence in oneself predicted more, and the work moved from producing to verifying. A self-report survey — it measures perception, not performance.
- Kosmyna et al. (2025), MIT Media Lab Preprint
EEG during essay writing across LLM, search-engine and unaided groups: LLM users showed the weakest brain connectivity and recalled less of their own essays; when the tool was removed, performance did not immediately recover. Not peer reviewed, 54 participants with 18 in the crossover session, and it measures neural and recall proxies rather than writing skill.
- Bastani et al. (2025), PNAS Peer-reviewed
The order-of-operations finding, in maths rather than writing: unrestricted assistance during practice raised performance 48% and left students 17% worse unaided afterwards, while assistance restricted to hints did no harm at all.
The learning science underneath
None of the advice above is original to this skill. It is the general findings on how humans learn, applied to one job — and each of these pages sets out the evidence for the principle itself.