TL;DR. If writing were only a way to produce text, AI would have made it a declining skill. The evidence says writing is also a way to think, and that part does not transfer to a model. A 2020 meta-analysis in the Review of Educational Research by Steve Graham, Sharlene Kiuhara and Meade MacKay pooled 56 experiments and found that writing about content reliably improved learning in science, social studies and mathematics, with an effect size of 0.30. On the production side, AI clearly helps: in a 2023 experiment published in Science, professionals using ChatGPT took 40% less time on writing tasks and their output was rated 18% higher in quality. But a 2024 study in Science Advances found that stories written with AI ideas were individually more novel and collectively more similar to each other, and readers imposed an ownership penalty of at least 25% on writers who used AI ideas. Meanwhile the baseline is eroding: in the OECD’s 2023 Survey of Adult Skills, 26% of adults across OECD countries scored at or below Level 1 in literacy, and the share of low performers rose in 14 countries and fell in none. Our analysis of the August 2026 O*NET database shows writing is rated at least 3 out of 5 in importance for 65.2% of 910 US occupations, and highest in legal, science, education, social service and management work. The practical answer is to split writing into four jobs, discover, decide, direct and deliver, and hand only the last one to AI by default.
This piece belongs to our series on foundational cognitive skills in the AI era, alongside deep reading as a competitive advantage, what research says about cognitive offloading and the discernment gap.
The case against learning to write, taken seriously
The argument for letting writing go is not stupid. A large language model can draft an email, a report, a proposal or an essay in seconds. The drafts are often good. In the experiment by Shakked Noy and Whitney Zhang, published in Science in 2023, 453 college-educated professionals were given occupation-specific writing tasks and half were randomly given access to ChatGPT. Average time taken fell by 40% and output quality rose by 18%. Workers exposed to the tool were also more likely to keep using it in their real jobs afterwards.
If the job of writing is to produce a finished text, that is close to a decisive result. Spend less time drafting, more time on something else.
The flaw is in the premise. Most of the writing that matters in a career is not produced for its own sake. It is produced to figure out what you think, to commit to a decision, or to tell someone, or something, what to do. For those jobs, the draft is a by-product of the thinking. Delegate the draft and you may delegate the thinking with it.
What the evidence says writing does to thinking
The idea that writing is a way of thinking is old. Linda Flower and John Hayes opened their 1981 “cognitive process theory of writing” in College Composition and Communication with the question of what guides the decisions writers make as they write. What is newer is the quantity of experimental evidence on whether writing actually improves learning.
Table 1. Meta-analytic evidence on writing and learning
| Source | Scope | Finding |
|---|---|---|
| Graham, Kiuhara and MacKay, Review of Educational Research, 2020 | 56 true or quasi-experiments, grades 1 to 12, science, social studies and mathematics | Writing about content reliably enhanced learning, effect size 0.30; equally effective across subjects and school levels; not moderated by features of the writing activity |
| Bangert-Drowns, Hurley and Wilkinson, Review of Educational Research, 2004 | 48 school-based writing-to-learn programs | A small, positive impact on academic achievement; larger with metacognitive prompts and longer treatment; smaller in grades 6 to 8 and with longer writing assignments |
| Graham and Hebert, Writing to Read, Carnegie Corporation report, 2010 | Meta-analysis of writing and reading comprehension | Writing about text read: average weighted effect size 0.49 on researcher-designed tests (44 studies); extended written responses 0.77 (9 studies); summary writing 0.52 (19 studies); note-taking 0.47 (23 studies) |
Three details in this table matter for adults working with AI.
First, the effect comes from the writer doing the writing. These are studies of students writing about material themselves. None of them tested reading an AI-written summary instead, and the logic of the finding points the other way: the learning comes from the effort of turning material into your own words.
Second, metacognitive prompts increased the effect. In the Bangert-Drowns meta-analysis, writing activities that used metacognitive prompts, which ask students to reflect on their own understanding, produced larger effects. That is the kind of writing most worth keeping for yourself: writing that tells you what you do not yet understand.
Third, longer is not better. Longer writing assignments predicted smaller effects in the same meta-analysis. Short, frequent writing to think seems to beat long writing to perform.
A caution: this evidence comes from schools, not offices. No meta-analysis we found measures writing-to-learn effects in adult knowledge workers. We use it as the best available evidence on mechanism, not as a precise estimate for professionals.
What AI drafting does to writing
The evidence on AI and writing now points in two directions at once, and both are real.
Table 2. Evidence on generative AI and writing quality, diversity and ownership
| Source | Design | Finding |
|---|---|---|
| Noy and Zhang, Science, 2023 | Preregistered experiment, 453 professionals, occupation-specific writing tasks | With ChatGPT, time taken fell 40% and output quality rose 18% |
| Doshi and Hauser, Science Advances, 2024 | 293 writers, 600 evaluators; short stories written with or without GPT-4 story ideas | One AI idea raised rated novelty 5.4% and usefulness 3.7%; up to five ideas raised them 8.1% and 9.0%. But AI-assisted stories were more similar to each other, an increase equal to 10.7% (one idea) and 8.9% (five ideas) of the full range of similarity scores |
| Doshi and Hauser, Science Advances, 2024 | Same study, evaluator judgments | Evaluators imposed an ownership penalty of at least 25% on writers who used AI ideas; 88.4% of writers offered AI ideas used them at least once |
| Padmakumar and He, ICLR 2024 | Controlled experiment: argumentative essays written alone, with a base model (GPT-3) or with a feedback-tuned model (InstructGPT) | Writing with InstructGPT, but not base GPT-3, significantly reduced the diversity of content across essays; the loss came from the text the model contributed |
| Anthropic Education Report, April 2025 | Analysis of 574,740 anonymized conversations from higher-education accounts | Nearly half (about 47%) of student conversations were Direct, seeking answers or content with minimal engagement |
Read together, these results describe a trade. AI raises the floor of each individual draft and narrows the range of drafts across people. For a routine status report, a narrower range is fine. For anything where your value depends on being different, a strategy, an argument, a position, a story, the narrowing is the cost.
The ownership penalty is the second cost, and it is easy to miss. Readers in the Doshi and Hauser study discounted the work of writers who used AI ideas. Whether that penalty persists as AI becomes normal is an open question. For now, being seen to think for yourself is still worth something.
The sibling research on cognitive effort points the same way. We covered it in detail in is AI making you worse at thinking: the pattern across those studies is that the more a person relies on AI for the thinking part of a task, the less of that thinking they do.
Why the skill is getting scarcer, not more common
A skill becomes more valuable when it is useful and scarce. On the scarcity side, the most recent large-scale measurement of adult literacy is not encouraging.
The OECD’s Survey of Adult Skills, run in 31 countries and economies in 2022 and 2023 and published in December 2024, found an OECD average literacy score of 260 points. Across OECD countries, 26% of adults scored at or below Level 1 in literacy, which the OECD describes as being able to understand short texts and organised lists when information is clearly indicated. Over the past decade, only Finland and Denmark saw significant improvements in adult literacy. Some 14 countries recorded an increase in the share of low-performing adults in literacy, and no country saw a reduction.
Table 3. Adult literacy in selected countries, OECD Survey of Adult Skills 2023
| Country | Mean literacy score | Share at or below Level 1 | Change in mean since 2011-12 |
|---|---|---|---|
| Finland | 296 | 12% | Up |
| Japan | 289 | 10% | Similar |
| England (UK) | 272 | 18% | Similar |
| Canada | 271 | 19% | Similar |
| Germany | 266 | 22% | Similar |
| OECD average | 260 | 26% | Not applicable |
| France | 255 | 28% | Down |
| Spain | 247 | 31% | Similar |
Source: OECD Survey of Adult Skills 2023, country notes. Adults aged 16 to 65.
In the United States, the National Center for Education Statistics reported that the share of adults at Level 1 or below in literacy rose from 19% in 2017 to 28% in 2023. NCES notes two changes that matter for that comparison: in 2023 all participants took the assessment on tablets, and items designed to assess more basic skills contributed to the scores only in 2023.
The survey measures reading literacy, not writing, so it is an indirect signal. But the two are linked in the evidence above: in the Graham and Hebert meta-analysis, writing about a text improved comprehension of it. A skill that fewer adults practise and master is a skill that differentiates the people who do.
Where writing matters most at work: our O*NET analysis
On the usefulness side, we went to the most detailed public database of what jobs require. ONET, sponsored by the US Department of Labor, rates the importance of 10 basic skills for each of hundreds of occupations on a 1-to-5 scale. We used the ONET 31.0 release (August 2026) and computed the figures below ourselves.
Across 910 occupations, writing has a mean importance of 3.19. It ranks sixth of the 10 basic skills, below active listening (3.60), speaking (3.54), critical thinking (3.52), reading comprehension (3.49) and monitoring (3.31), and above active learning, learning strategies, mathematics and science. Writing is rated at least 3.0 in importance for 65.2% of occupations, at least 3.5 for 38.2%, and at least 4.0 for 16.2%.
Table 4. Importance of writing by occupation family, O*NET 31.0 (CEOtudent calculation)
| Occupation family (US SOC major group) | Occupations rated | Mean writing importance (1 to 5) |
|---|---|---|
| Legal | 7 | 4.03 |
| Life, physical and social science | 59 | 3.87 |
| Educational instruction and library | 62 | 3.83 |
| Community and social service | 14 | 3.79 |
| Management | 55 | 3.72 |
| Business and financial operations | 46 | 3.64 |
| Architecture and engineering | 55 | 3.62 |
| Computer and mathematical | 36 | 3.51 |
| All 910 occupations | 910 | 3.19 |
| Production | 107 | 2.52 |
| Construction and extraction | 61 | 2.34 |
| Food preparation and serving | 16 | 2.34 |
The pattern is the point. Writing matters most in exactly the families where work consists of judgment, argument and instruction: law, science, teaching, management and business. Those are also the jobs where AI drafting is most available. The skill is concentrated where the temptation to delegate it is highest.
Employers say something similar. In the National Association of Colleges and Employers’ Job Outlook 2025 survey, 77.1% of responding employers said they look for written communication skills on a candidate’s resume, behind only problem-solving and teamwork among the attributes listed.
The four jobs of writing, and which ones to hand to AI
The confusion in the “should I still learn to write?” debate comes from treating writing as one thing. It is at least four. The framework below is our editorial synthesis; the evidence column shows what each recommendation rests on.
Table 5. The four jobs of writing (CEOtudent editorial framework)
| Job | What the writing is for | Examples | Default AI role | Evidence behind the default |
|---|---|---|---|---|
| Discover | Finding out what you think | Journals, working notes, first drafts of an argument, reflections on what you do not understand | None during the first draft; afterwards, as a questioner that challenges your draft | Writing-to-learn meta-analyses: learning comes from the writer’s own effort; metacognitive writing works best |
| Decide | Committing to a position others can inspect | Decision memos, proposals, strategy documents, post-mortems | You write the argument; AI checks it, finds gaps, argues the other side | Amazon’s narrative-memo practice; AI narrows the range of ideas across people |
| Direct | Telling people and agents what to do | Briefs, specifications, prompts, delegation instructions | You write it; AI can help you test whether it is unambiguous | The quality of delegated work depends on the brief; see our agent briefing framework |
| Deliver | Communicating something already decided | Routine emails, status updates, summaries, formatting | AI drafts; you edit for accuracy, tone and anything that commits you | Noy and Zhang: 40% less time, 18% higher rated quality on such tasks |
Two points about the table.
Direct is the job that grows with AI. The more work you hand to agents, the more of your day becomes writing instructions. A vague brief produces confident, plausible, wrong work. Writing a precise brief is writing as thinking in its purest form: you cannot specify what you have not worked out. Our piece on the meeting that should have been an agent shows the same shift in team communication: when agents take over the information-moving meetings, what remains is written decisions and written briefs.
Decide is the job that separates people. Jeff Bezos described Amazon’s practice in his 2017 letter to shareholders: “We don’t do PowerPoint (or any other slide-oriented) presentations at Amazon.” Instead, meetings start with narratively structured six-page memos read silently, and he noted that a great memo can take a week or more to write. That is one company’s practice, not a controlled study. But it illustrates the mechanism: a narrative forces the writer to connect claims into an argument, and gaps that a bullet list hides become visible. If AI writes the memo, the gaps are still there. They just stop being visible to you.
The write-first routine
A simple rule follows from the evidence: write first, prompt second for anything in the discover, decide or direct jobs. In practice, for a piece of thinking work:
- Write the ugly version yourself. Ten to twenty minutes, no AI, no formatting. State your position, your reasons and what you are unsure about. The evidence favours short, reflective writing over long polished writing.
- Name what you do not understand. One paragraph listing your open questions. This is the metacognitive step that increased writing-to-learn effects.
- Then bring in AI as a critic, not an author. Ask it to find the weakest claim, argue the opposite position, list missing evidence, or point out where your reasoning jumps.
- Revise it yourself. Accept or reject each critique consciously. The ideas that survive are now yours, tested.
- Only then let AI help with delivery. Formatting, tightening, translating, adapting for different readers.
- Keep a small record. Once a week, look at one decision you wrote up and ask whether the writing changed your mind about anything. If it never does, you are writing to perform, not to think.
This is the CEO and student lens applied to a single skill. The CEO half: you remain the author of your decisions and accountable for them, so the argument must be yours. The student half: the act of writing is how you keep learning, and a tool that removes the effort also removes the learning.
When AI-first writing is fine
The framework is a default, not a moral rule. AI-first drafting is reasonable when:
- The content is already decided. A status email reporting a decision made elsewhere is delivery, not thinking.
- The format is standard. Meeting notes, routine replies, form letters.
- You will not need to defend or build on it. If nobody will ask you why, the thinking behind it matters less.
- You are writing in a language you do not master. Using AI to express your own ideas in another language is a translation aid, as long as the ideas came first.
And AI-first is a poor choice when your name goes on an argument, when the text will instruct someone else’s work, or when you are trying to learn the subject.
Frequently asked questions
Is writing still a valuable skill if AI can write?
Yes, for the parts of writing that are thinking. Meta-analyses of writing-to-learn studies show that writing about content improves learning, with an effect size of 0.30 across 56 experiments. AI can take over routine delivery writing, but deciding, arguing and instructing still depend on your own thinking, and writing is how you do that thinking.
Does using AI to write make you a worse thinker?
It can, if AI writes the parts where the thinking happens. Studies of AI-assisted writing find that drafts become more similar across people, and in a survey of knowledge workers, higher confidence in AI was associated with less self-reported critical thinking. Writing your own first draft and using AI as a critic avoids most of this.
What kind of writing should I delegate to AI?
Delivery writing: routine emails, status updates, summaries of decisions already made, formatting and adaptation for different audiences. In one experiment, professionals using ChatGPT on such tasks took 40% less time and produced work rated 18% higher in quality.
What kind of writing should I keep doing myself?
Writing to discover what you think, writing that commits you to a decision, and writing that instructs people or AI agents. These are the jobs where the draft is a by-product of your reasoning.
Why does writing matter more when working with AI agents?
Because delegating to an agent is done in writing. The quality of an agent’s work depends on how precisely the task, constraints and success criteria are written. The more you delegate, the more of your job becomes writing clear instructions.
How can I practise writing as thinking?
Write a short, unassisted first draft of your position before asking AI anything, list what you do not understand, then use AI to critique the draft rather than write it. Short, frequent, reflective writing has stronger evidence behind it than long, polished writing.
Sources
Graham, S., Kiuhara, S. A., and MacKay, M. The effects of writing on learning in science, social studies, and mathematics: a meta-analysis. Review of Educational Research, 2020, volume 90, issue 2, pages 179 to 226.
Bangert-Drowns, R. L., Hurley, M. M., and Wilkinson, B. The effects of school-based writing-to-learn interventions on academic achievement: a meta-analysis. Review of Educational Research, 2004, volume 74, issue 1, pages 29 to 58.
Graham, S., and Hebert, M. A. Writing to Read: Evidence for How Writing Can Improve Reading. A Carnegie Corporation Time to Act Report. Alliance for Excellent Education, 2010.
Flower, L., and Hayes, J. R. A cognitive process theory of writing. College Composition and Communication, 1981, volume 32, issue 4, pages 365 to 387.
Noy, S., and Zhang, W. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 2023, volume 381, issue 6654, pages 187 to 192.
Doshi, A. R., and Hauser, O. P. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 2024, volume 10, issue 28, article eadn5290.
Padmakumar, V., and He, H. Does writing with language models reduce content diversity? International Conference on Learning Representations (ICLR), 2024.
Anthropic. Anthropic Education Report: How University Students Use Claude. April 2025.
OECD. Do Adults Have the Skills They Need to Thrive in a Changing World? Survey of Adult Skills 2023. OECD Skills Studies, December 2024; and the accompanying country notes.
US National Center for Education Statistics. Highlights of the 2023 U.S. PIAAC Results. December 2024.
National Center for ONET Development. ONET 31.0 Database, August 2026, skills importance ratings.
National Association of Colleges and Employers. Job Outlook 2025.
Bezos, J. 2017 Letter to Shareholders. Amazon, April 2018.
The ONET means, rankings, threshold shares and occupation-family averages in the ONET section and Table 4 were computed by us from the O*NET 31.0 skills file (importance scale). Table 5 is a CEOtudent editorial framework.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
