When Stephen Marche published "The Computers Are Getting Good," his Atlantic essay on AI writing, he disclosed that he had used GPT-3 to draft passages he then edited extensively. The resulting text was, by his own account, indistinguishable in published form from his usual prose — not because AI wrote it, but because his editing voice absorbed it. The AI gave him raw material; he gave it direction.
The episode provoked debate precisely because readers could not tell which words originated with the machine. What they could tell was that the final piece sounded like Marche. That asymmetry — undetectable origin, detectable author — sits at the heart of what AI drafting partnership actually means.
The phrase gets used loosely. In practice, AI as a drafting partner occupies a specific role distinct from AI as an editor, AI as a research tool, or AI as a content generator. A drafting partner produces raw text that the human writer then shapes. The critical variable is who controls selection.
When you ask an AI to generate three different openings for a piece, you are the editor choosing among options. When you use its output as scaffolding you will tear apart and rebuild, you are the architect using temporary supports. Both are legitimate — but they produce different cognitive experiences and, crucially, different final voices.
The drafting partner model works best when the writer is a more demanding editor than generator. If you find yourself accepting AI sentences wholesale because they are "good enough," the partnership has drifted toward ghostwriting. If you use AI output to prime your own thinking and then write over it, the partnership is working as intended.
Psychologist Mihaly Csikszentmihalyi documented in his 1990 research on flow states that initiation tasks — starting a creative work — consume disproportionately high cognitive load relative to continuation tasks. Writers do not struggle to continue; they struggle to begin. AI drafting directly addresses this load asymmetry by eliminating the standing-start problem.
In 2023, journalist Charlie Warzel documented his process at The Atlantic: he would describe to an AI the argument he wanted to make, receive a rough 600-word draft, read it once — specifically looking for what was wrong — and use his objections as the first paragraphs of his actual piece. The AI's wrongness was generative. His disagreement with it became his voice.
This technique exploits a known feature of creative cognition: it is far easier to identify what you don't want than to produce what you do want from nothing. Reaction is faster than creation. The AI produces something to react to, and the writer's reactions cohere into a genuine point of view.
First drafts from AI share the same fundamental limitation as first drafts from any source: they are wrong in interesting ways. They make arguable claims too confidently, they use safe phrasing where specificity is needed, and they default to the generic shape of whatever genre they are imitating. None of this is fatal — first drafts are supposed to be wrong. The problem arises when writers treat AI first drafts as drafts two through five. That is where voice disappears.
A sustainable drafting partnership treats AI output as draft zero — earlier than first draft, the raw deposit of language before the author has arrived. The author arrives in revision.
The drafting partner model preserves authorial voice not by limiting what AI produces, but by maximizing the editorial pressure the writer applies afterward. The AI starts; the writer decides what starting was for.
Describe a piece you want to write — an essay, article opening, or argument — in two or three sentences. Ask the AI for a rough 200-word draft. Then push back: tell the AI specifically what it got wrong, what voice is missing, what claim is too vague. The AI will help you sharpen your actual position through disagreement.
When Hua Hsu won the Pulitzer Prize for memoir in 2023, discussions of his process noted something instructive: Hsu had written Stay True across six years, during which AI writing tools became available. He did not use them — but the reason he cited in interviews illuminates the core issue. He needed the wrongness of his early drafts to find out what he was actually trying to say. His bad first drafts were diagnostically his.
The AI's default drafts are not wrong in the same way. They are wrong in generic ways — smooth where the writer should be awkward, confident where the writer should be uncertain. A writer who prompts for their own voice rather than AI's default is asking for a different kind of wrong: wrong in a useful, personal direction.
Large language models are trained to predict probable next tokens. In practice this means they trend toward the most frequent phrasings, structures, and arguments in their training data. For expository prose, this produces text that is competent, clear, and deeply uninteresting — the statistical center of published writing, which is to say the prose of writers who were not trying to distinguish themselves.
This is a documented feature, not a malfunction. OpenAI's 2023 system card for GPT-4 explicitly notes that the model tends toward "safe, non-controversial" outputs by default — a deliberate alignment choice. The result is prose that offends no one and moves no one.
Distinctive voice is, almost by definition, deviation from the statistical center. It is the choice to be awkward when smooth would do, specific when general would serve, emotional when clinical would protect. AI defaults pull against all of these deviations unless you instruct it otherwise.
When a writer submits a rough draft for AI to "improve," the improvement removes the deviations that constitute voice. The smoothed result is technically better by certain metrics and expressively diminished by the only metric that matters: does it sound like you?
The key insight, documented in writing by practitioners including Ann Friedman (newsletter, 2023) and discussed at length in the Nieman Lab's 2023 series on AI and journalism, is that specificity in the prompt about voice characteristics overrides defaults. Generic prompts produce generic text. Voice-specific prompts produce voice-adjacent text you can edit toward specificity.
Effective voice-preserving prompts do several things: they name the emotional register (not "professional" but "blunt, slightly impatient, credible"); they specify structural habits ("I use short sentences after long ones to land arguments"); they identify what to avoid ("no hedging words like 'perhaps' or 'it might be argued'"); and they provide a few sentences of the writer's own prose as stylistic anchors.
Rather than asking for a single draft, requesting three versions of the same passage with different voice characteristics forces the AI to range across options. You then identify which version is closest to your voice and describe why — that description becomes the voice specification for subsequent prompts.
This method was informally documented by several journalists at the 2023 Online News Association conference, who reported that the act of choosing between AI versions helped them articulate voice characteristics they had never previously put into words. The AI's range became a mirror for their own preferences.
Caution: This technique requires that you make the choice — not that you average the three versions or ask AI to blend them. The averaging produces the generic result you were trying to avoid.
AI defaults to the statistical center of published prose. Your voice lives at the periphery — in your specific deviations. Prompting for your voice means explicitly describing those deviations, providing textual anchors, and resisting the pull toward the smooth and the safe.
Start by pasting two or three sentences of your own writing. Ask the AI to describe the voice characteristics it detects. Then ask it to draft a short paragraph on a topic of your choice using those characteristics. Compare the result to the generic version — ask the AI to produce both so you can see the difference.
The Reuters Institute Digital News Report 2023 documented patterns among journalists at 46 news organizations who had adopted AI drafting tools. A subset had developed what researchers described as iterative feedback loops: they would submit an AI draft, mark the paragraphs that failed, describe why they failed, and request a targeted revision. Three to five passes typically produced material the journalist considered publishable after light editing.
The report noted something troubling alongside this: journalists who ran more than six or seven revision passes often reported that the resulting text had grown smoother but felt less like them. The AI was optimizing for internal coherence and surface fluency — qualities that, pursued past a certain point, erased the productive roughness that had characterized the journalist's own style.
The productive version of iterative AI drafting follows a specific pattern. The writer submits a targeted critique — not "make this better" but "the third paragraph assumes the reader already accepts the premise; rebuild it as if they are skeptical." Each pass addresses a specific failure. The AI's response to specific criticism is substantially better than its response to general improvement requests.
Specificity of feedback drives quality of revision. This is well documented in human editing contexts (Gordon Lish's editorial letters to Raymond Carver, for instance, were extremely specific about what to cut and why), and the pattern holds with AI. Vague feedback — "this doesn't sound right" — produces vague improvement. Named problems produce targeted solutions.
The workflow that produces the best results typically involves the writer maintaining a running document of criteria: what they are asking for in each pass, what the AI got right, what it still gets wrong. This document becomes a progressively refined voice specification.
Gordon Lish's 1970s edits of Raymond Carver's manuscripts — cutting some stories by more than 50% — are now extensively documented at the Lilly Library. What made those edits productive rather than destructive was their specificity: Lish could name exactly what he was removing and why. Iterative AI feedback works by the same principle: named removals beat vague improvement.
The Reuters Institute finding points to a structural risk in iterative drafting: each pass the AI makes optimizes for the local problem the writer described, but it also incrementally smooths surrounding text. After several passes, the cumulative smoothing can strip out the idiosyncratic rhythms, abrupt transitions, and opinionated word choices that constitute distinctive voice.
This is the drafting equivalent of overfitting in machine learning — the AI has optimized so precisely for the stated criteria that it has lost generalization, which in writing terms means it has lost life. The writer ends up with text that answers every critique but feels like no one in particular.
The practical solution is to return to the writer's own prose after three to four passes rather than continuing to iterate. Use what the AI gave you, but write the next section yourself using the AI's structure as a scaffold. Then submit that section for a new first pass rather than continuing to revise the same material.
Writers in the Reuters Institute sample described a recognizable moment when iterative drafting stopped being productive: they noticed they were critiquing the AI's revisions using criteria that couldn't be articulated. "This is technically fine but wrong" is the loop-break signal. It means the problem is no longer specifiable — it is a voice problem, and voice cannot be specified into existence through iterative critique. At that point, the writer must write.
This is not a failure of the process. It is the process working correctly: the AI has handled the specifiable problems and delivered you to the boundary of what only you can do. That handoff point — where specification ends and authorship begins — is the most important skill in AI-assisted drafting.
Three to four targeted revision passes produce better material than one general request and better material than seven passes. The optimum is specific, bounded, and followed by the writer's own pen. Over-iteration smooths the voice out of text that the earlier passes helped create.
Ask for a first draft of a short argument. Then critique it specifically — not "make it better" but name exactly what fails: "the second sentence hedges when it should be direct," or "the opening is too general — it needs a specific example." Ask for a targeted revision. Do this three times. On the third critique, try to identify whether you've hit the loop-break signal.
In February 2023, Clarkesworld Magazine — one of the most prominent science fiction publications in the world — temporarily closed submissions after receiving a flood of AI-generated story submissions. Editor Neil Clarke documented the numbers publicly: the magazine had received more AI-generated submissions in January 2023 than in all previous years combined. He closed the submissions window entirely.
The Clarkesworld crisis crystallized something the industry had been quietly arguing about: the difference between using AI as a drafting tool and submitting AI output as original work. Clarke himself distinguished between the two in subsequent interviews — he objected not to AI assistance but to undisclosed AI generation presented as original writing. That distinction, between tool use and authorship substitution, has become the operational line most publishers are trying to draw.
Following the Clarkesworld crisis, a rapid industry response produced a patchwork of policies that is worth examining directly rather than summarizing. The Authors Guild (May 2023) released principles stating that AI-generated text in a submitted work must be disclosed, and that AI should not replace human creative authorship. Penguin Random House updated its author contracts in mid-2023 to require authors to warrant that submitted manuscripts do not consist substantially of AI-generated content without disclosure. The New York Times updated its ethics guidelines to require journalists to flag any AI-assisted drafting to editors.
Notably, none of these policies ban AI use. They regulate disclosure and degree. The operative question is not whether AI was used but how substantially and whether that use was disclosed to the relevant party — editor, publisher, or reader.
In March 2023, the US Copyright Office issued guidance stating that works containing AI-generated material are copyrightable only to the extent of human authorship. In the Zarya of the Dawn case (February 2023), it determined that AI-generated images in a comic book could not be copyrighted, while the human-authored text and arrangement could be. The threshold for "sufficient human authorship" in prose remains legally unresolved but the principle is established: substantial human creative contribution is required for copyright protection.
The ethical questions around AI-assisted drafting differ depending on the relationship between writer and reader. A novelist submitting to a publisher operates under a different disclosure regime than a blogger writing for their own platform. A journalist at a newspaper is subject to editorial policy; a newsletter writer is subject only to their own stated commitments to readers.
What is consistent across relationships is the principle that undisclosed AI generation creates a false impression of authorship when authorship was implicitly promised. A byline implies a human wrote the piece. A personal essay implies a person's experience. A memoir implies a life. Where these implicit promises exist, AI substitution without disclosure is a breach — regardless of the quality of the editing applied afterward.
AI assistance, by contrast, has a long tradition under different names: researchers, research assistants, editors, ghostwriters, developmental editors, and collaborators have all contributed to published works without byline credit. The question is one of degree and transparency.
The industry has drawn a working line between tool use and authorship substitution. What it has not resolved — and what courts, publishers, and scholars continue to debate — is where heavily edited AI text sits relative to that line. If a writer generates ten pages of AI draft and edits it to 800 words that read as entirely their own, is that authorship substitution or tool use? The Copyright Office's "sufficient human authorship" standard was articulated for images; its application to prose has not been formally tested.
The honest answer is that current norms are evolving faster than current policies. Writers operating in good faith should: disclose AI use to any publisher with a stated policy; clearly distinguish their own creative contributions from AI-generated material in their own records; and be prepared for norms to continue shifting as the technology matures and legal precedents accumulate.
The ethical line is disclosure, not use. AI as a drafting tool is consistent with authorship when the human writer substantially shapes the final work. Undisclosed AI generation presented as original writing violates the implicit contract of the byline — regardless of how much editing was applied afterward.
Describe a specific writing scenario involving AI assistance — a personal essay, a journalism piece, a social media post, a book proposal — and ask whether that level of AI use requires disclosure, to whom, and what form that disclosure should take. Push back on the AI's first answer to test the edges of the principle.