When AI Produces the First Draft, Has the Writer’s Work Really Decreased?

Geoffrey Chen

After AI Enters the Workflow · Season One: “From Answering Questions to Participating in Work” · Article 3

The hardest part of an article is sometimes the blank page.

The subject has been chosen and the material exists across several files, but the opening, sequence and destination of the argument remain unclear. A writer might once have spent hours pushing those fragments into an unsatisfactory first draft. Now the same writer can provide the subject, sources and requirements to an AI system and receive a fluent, apparently complete article within minutes.

Measured by the clock, work has plainly been reduced. The blank page is gone. Paragraphs, headings and transitions already exist.

Then another set of questions arrives. Is the opening example real? Has the second section turned correlation into causation? Which source supports a confident summary? Does the central conclusion belong to the writer, or has the model supplied a familiar position because the brief left the judgement unresolved? If half the draft should be deleted, which half actually carries the argument?

The faster a first draft appears, the more these decisions are concentrated in the later stages. AI reduces the friction of producing sentences and structure, but it may move the difficult work from “How do I write this?” to “What is worth retaining, how can it be supported, and who is willing to stand behind it?”

The effect cannot be expressed as a simple percentage of words written by each party. We must first separate the different kinds of labour that writing contains.

Research has found substantial gains in speed

The productivity effect on some writing tasks is supported by more than anecdote.

Shakked Noy and Whitney Zhang asked hundreds of university-educated professionals to complete occupational writing tasks such as press releases, short reports, analytical plans and sensitive emails. Participants with access to ChatGPT completed the tasks about 40 per cent faster on average, while independent evaluators rated the outputs about 18 per cent higher in quality. MIT Economics, “Study finds ChatGPT boosts worker productivity for some writing tasks”

The distribution of time also changed. AI users spent less time producing a rough draft and relatively more time planning and editing. The system did not merely accelerate typing; it shifted attention among stages of work. Science, “Experimental evidence on the productivity effects of generative artificial intelligence”

The scope of the result matters. These were mid-level professional writing tasks, but they did not require participants to master a complex body of internal organisational knowledge, and rigorous source verification was not the central challenge. The researchers also noted that prompting, fact-checking and highly specific context in real work can reduce the time saved.

The justified conclusion is therefore narrower than the slogan. AI can sharply reduce the cost of drafting in some forms of professional writing. The study does not show that every article requires 40 per cent less authorship, or that faster text is ready for publication.

“First draft” does not mean the same thing in every genre

A routine notice, a factual investigation and an essay of original judgement all have first drafts, but those drafts perform different functions.

In a stable administrative notice, the relevant facts have already been settled. The writing task is to convey them accurately, appropriately and consistently. If AI receives the correct facts and template, it may perform much of the linguistic work. The reviewer focuses on recipient, dates, commitments and authority to send.

In a research report, sentences must be connected to evidence. A draft must show which claims come from data, which from literature and which are the author’s inferences. AI can generate a coherent narrative quickly; it is less reliable at constructing a fully traceable evidentiary chain without deliberate system design and checking.

In an argumentative essay, writing may itself be a mode of thought. The author does not always possess a complete view and then translate it into prose. The view emerges while testing formulations, finding contradictions and abandoning concepts that initially appeared useful. If AI supplies a complete architecture too early, it may help—but it may also close the problem before the author has discovered what remains unresolved.

A first draft can therefore be at least three things:

  • a language draft, expressing content that is already settled;
  • a structural hypothesis, proposing one way to organise the material;
  • a record of thinking, preserving difficulty, revision and discovery.

AI is most directly useful in the first. In the second it can generate alternatives. In the third, its smoothness can either reveal relationships or conceal unfinished thought.

A fluent draft creates a new editorial burden

Traditional early drafts often reveal their incompleteness through broken sentences, repetitions and visible gaps. Writers know where further work is needed.

An AI draft can be dangerous precisely because it looks finished. The title, transitions and conclusion arrive together. Consistent style creates an impression that the argument is already coherent. Errors rarely appear only as absurd inventions. More often, a real statistic loses its date, a limited study is generalised, a qualification disappears during rewriting, or two correct facts are connected by an unsupported causal claim.

Editing therefore changes character. The writer is no longer improving an obviously incomplete personal draft. The writer is auditing a persuasive text whose sources of knowledge and choices are partly opaque.

That audit has at least four layers.

The first is factual review: names, dates, numbers, institutions, rules and quotations must return to primary or authoritative sources.

The second is argument review: the evidence must actually support the conclusion; qualifications and counterexamples must survive; verbal flow must not substitute for logical connection.

The third is selection review: why these facts, examples and sections rather than others? A model’s comprehensive-looking coverage may dilute the question the article is supposed to answer.

The fourth is voice review: does the text express a judgement the named author is prepared to own publicly, or merely a statistically familiar form of reasonable-sounding prose?

This work requires active doubt. The editor has to resist the appearance of completion and break the draft back into claims, evidence, inference and rhetoric.

Authorship is not determined by the proportion of keystrokes

If a model generated most of the sentences, is the result still the human writer’s article? Counting keystrokes is a poor answer.

Authorship has never meant that every visible word was physically entered by one person. Editors reshape manuscripts; translators choose language; research assistants gather material. What matters is whether the named author controlled the purpose, selected and verified the evidence, made the substantive judgements, approved the final form and can answer for errors.

AI complicates this because it can contribute language, structure and apparent judgement simultaneously. It cannot, however, accept the obligations attached to publication. The International Committee of Medical Journal Editors states that AI tools should not be listed as authors because they cannot be responsible for accuracy, integrity and originality; human authors must disclose relevant use and remain accountable. ICMJE, “Use of AI by Authors”

Springer Nature’s editorial policy similarly does not accept large language models as authors and places accountability with human authors. Springer Nature, “Artificial Intelligence editorial policy”

These policies concern scholarly publishing, but the underlying distinction applies more broadly. Assistance may be delegated; authorship responsibility cannot be transferred to a system that cannot inspect allegations, correct a public record or accept professional consequences.

AI may reduce drafting labour while increasing publication volume

Even when AI reduces the effort required for one article, it can increase the total work surrounding publication. If an organisation responds by producing three times as much content, the demand for source checking, legal review, editing, metadata, translation and maintenance also grows.

Cheap generation can create an editorial queue whose apparent completeness disguises the cost of validation. A weak draft written slowly may never leave the writer’s desk; a polished AI draft is more likely to be circulated, reviewed and mistaken for near-finished work. The cost has not vanished. It has moved downstream.

This is why productivity should not be measured only as words per hour. The relevant unit is an adopted, verified and maintainable publication. A thousand generated pages are not productive if responsible editors can validate only a hundred.

A more defensible AI-assisted writing process

A strong workflow begins before generation.

First, the writer defines the question, intended reader, factual scope and judgement the article must reach. Second, sources are collected and their authority, date and relevance recorded. Third, AI may help compare structures, expose missing links or prepare a language draft. Fourth, claims are separated from their supporting evidence, and important facts are checked independently. Fifth, the writer rewrites the argument in a form that reflects an owned judgement rather than simply polishing the model’s prose. Finally, the article is reviewed as a public commitment: can its claims be defended, corrected and updated?

Disclosure should be proportionate to the context. In formal research, editorial rules may require explicit statements about AI use. In other settings, internal records may be sufficient. What cannot be acceptable is using a model to conceal fabrication, evade responsibility or imply that an automated system can be accountable as an author. Guidance on responsible scientific publishing continues to emphasise transparency, verification and human accountability as capabilities evolve. Nature Methods, “Using AI responsibly in scientific publishing”

Conclusion: the reduced burden is drafting friction, not necessarily authorship

When AI supplies the first draft, it can remove a real obstacle. It can make alternatives visible, give scattered material a provisional shape and release time from routine expression. Those are substantial gains.

But a first draft is not the unit for which an author is ultimately responsible. The author is responsible for the published claims, the selection and treatment of evidence, the integrity of the argument, the effect on readers and the correction of mistakes.

The correct conclusion is therefore not that AI leaves writers with nothing to do, nor that every AI-assisted sentence must be rewritten to become legitimate. It is this:

AI can reduce the labour of producing a draft. It cannot reduce the standard by which a publication becomes the author’s work.

If the saved time is invested in better questions, stronger evidence and more demanding revision, the writer’s work may become both faster and better. If it is used only to multiply fluent pages, authorship has not been reduced so much as displaced into an unmanageable verification burden.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.