In April 2023, the editors of Science journal flagged an unusual pattern: peer-review reports submitted by multiple independent reviewers contained nearly identical phrasing — phrases like "commendable" and "meticulous" appearing in clusters. Researchers at Stanford traced the pattern to ChatGPT. Reviewers under deadline pressure had used the model to draft feedback, and the model had, predictably, converged on its statistical favorites. The result was not plagiarism in any traditional sense — each reviewer had typed different prompts — but the output voices had collapsed into one.
The Stanford study, published in PLOS ONE in January 2024, estimated that between 6% and 16.9% of peer-review text submitted in late 2023 had likely been AI-generated. The signal was unmistakable precisely because the model's vocabulary peaks are so narrow.
A large language model (LLM) does not "write." It predicts the next most statistically likely token — a word fragment — given everything that came before it. That training corpus is enormous: Common Crawl, Wikipedia, GitHub, books, Reddit, news archives. But enormous does not mean diverse in the way human culture is diverse. The web over-represents English, over-represents certain registers (formal, mildly journalistic, listicle-style), and over-represents the kind of competent-but-forgettable prose that gets published in volume.
The consequence is what researchers call distributional homogenization. Ask ten people to independently describe a sunset and you will get ten genuinely different sentences. Ask ChatGPT, Claude, or Gemini to do the same and, across runs and across models, the outputs cluster. Words like "tapestry," "delve," "testament," "showcasing," and "nuanced" appear at rates far above their frequency in human writing because they sit near the peaks of the model's probability distribution for formal descriptive prose.
Temperature — the randomness parameter — can spread the distribution, producing wilder word choices. But a high-temperature output is random, not personal. Randomness is not a voice.
If you use AI output directly, you are not expressing your voice — you are publishing the average of the internet's voice. The goal of this module is to make AI a drafting tool that you edit back into your own idiom, not a ghostwriter that replaces it.
Ethan Mollick, a Wharton professor who studies AI adoption, ran an informal experiment in 2023 documented in his Substack "One Useful Thing": he asked GPT-4 to write in twenty distinct author styles. Across all twenty, certain grammatical constructions appeared regardless of style — compound sentences joined by semicolons in a very specific rhythm, a preference for the em-dash as a pause, and a resolution toward optimism at paragraph's end. Style prompting shifted surface vocabulary but rarely the underlying cadence.
This is the vocabulary gap: the distance between the statistical center of mass that AI defaults to and the specific, idiosyncratic vocabulary that makes a writer recognizable. Cormac McCarthy's refusal to use commas, Joan Didion's fragmentary syntax, David Foster Wallace's footnote-as-thought — none of these are behaviors a prompted LLM sustains under its own momentum for more than a few sentences before reverting to the mean.
Think of AI-generated prose as a rough clay shape — roughly right in form, but not yet carrying your fingerprints. Your job as the writer is to press those fingerprints in at every level: word choice, sentence rhythm, what you leave out, and the specific angle from which you see the world.
Paste a short AI-generated paragraph into the chat and ask the assistant to identify any "distributional homogenization" markers — words or structures that appear at AI-frequency peaks. Then ask it to suggest replacements drawn from a style or vocabulary you actually use. This lab trains the editorial eye you need to reclaim your voice from AI drafts.
In September 2022, novelist Robin Sloan — author of Mr. Penumbra's 24-Hour Bookstore — published a detailed essay in the Atlantic describing how he used GPT-3 during the writing of his novel Moonbound. Sloan's method was specific: he fed the model a scene description and asked it to generate ten possible sentences that a character might say. He used none of them verbatim. He used them as "a tuning fork," his phrase — something to push against. If the model produced a flat declarative, he wrote a run-on. If it produced optimism, he reached for irony. The AI was not a co-author; it was a negative space he wrote into.
Sloan was explicit about what made this work: he had already written two novels. He had enough command of his own idiom that he could feel immediately when the AI was pulling him toward the generic. Writers without that established self-knowledge, he warned, risk "learning the AI's voice instead of their own."
The single most important conceptual shift for creative writers using AI is this: never treat AI output as a candidate for publication. Treat it as a quarry — a source of raw stone from which you cut your actual work. This is not a moral position about AI authorship; it is a practical one about voice and quality.
When you submit an AI draft directly, you are inheriting all its statistical defaults: its vocabulary peaks, its preferred sentence rhythms, its tendency to resolve ambiguity toward clarity, its aversion to the kinds of syntactic risk that distinguish memorable prose. When you use it as quarry, you extract what is useful (a structural skeleton, a list of possible metaphors, a factual summary to paraphrase) and discard the rest.
The workflow distinction looks like this: Submitting AI drafts → you become an editor of someone else's voice. Using AI for material → you remain the writer, with a richer palette of raw options.
The Washington Post's technology team documented their internal AI writing guidelines in October 2023. Reporters were permitted to use AI to generate initial research summaries and to draft structural outlines, but all prose submitted for publication had to be written by the reporter. The guideline explicitly stated: "AI-generated sentences should not appear in published text." The rationale was not just accuracy — it was voice. Editors had noticed that AI-drafted ledes flattened the tonal signature that distinguished Post reporting from wire copy.
The prompts that generate useful raw material differ structurally from prompts that generate finished prose. Useful material prompts tend to:
1. Request plurality. "Give me twelve possible opening lines for this essay" produces a range from which you choose and modify. "Write the opening line" produces one answer you either use or reject wholesale.
2. Request raw ingredients, not assembled dishes. "List ten concrete images that could represent isolation in a city" gives you material. "Describe isolation in a city" gives you prose you will edit forever without making it yours.
3. Set explicit constraints that force departure from defaults. "Draft a paragraph about grief that uses no abstract nouns and no metaphors" pushes the model away from its statistical comfort zone and produces something stranger and more usable.
4. Request structural scaffolding, not prose. "Give me a three-section outline for this argument, with one counterargument per section, as bullets only" is infinitely more useful than asking for a draft essay, because you will write the sentences yourself.
The most experienced writers use AI output as something to write against, not from. If the AI produces the obvious, write the non-obvious. If it resolves toward hope, consider whether your piece needs ambiguity. The model's defaults can clarify your own choices by opposition.
In this lab you'll work with the assistant to build plurality prompts and constraint prompts for a writing project you have in mind. The goal is to produce raw material — not finished prose — and then identify which elements you would actually use and why.
In December 2022, Stephen Marche — author and cultural critic — published an essay in The Atlantic with an unusual credit line: the piece had been written collaboratively with GPT-3, Sudowrite, and Cohere. Marche documented the process in detail. He found that AI could produce competent sentences quickly but that those sentences had a specific defect: they resolved. They concluded. They moved toward the summarizing gesture. His own writing style — recursive, looping back on itself, comfortable with unresolved tension — had to be actively imposed on every paragraph through revision.
Marche's editing method was systematic: he identified every sentence that used an abstract noun where a concrete one would serve, every paragraph that ended with a claim instead of an image, every transition word ("therefore," "moreover," "additionally") that appeared in the AI draft and deleted it. Then he rewrote each of those positions from scratch. The result, he said, read like him. The process took longer than writing from scratch — but produced a richer draft to push against.
Editing AI text back into your voice requires working at five distinct levels simultaneously, because AI defaults operate at all five:
1. Vocabulary. Replace the model's statistical peak words with your actual vocabulary. This means having a clear enough sense of your own word preferences to know that you would never write "delve," that you prefer "look into" or "dig through" or sometimes just "read." If you don't yet know your vocabulary preferences well, read ten pages of your best previous writing and list the words that appear there but not in AI drafts.
2. Sentence rhythm. Count the syllables in three consecutive AI sentences. They will often be close to each other — the model prefers rhythmic regularity. Your own prose likely has more variation: a short sentence. Then a longer one that accumulates clauses in a specific way. Then another short one for emphasis. Rewrite to match your rhythm, not the model's.
3. Resolution habit. AI defaults toward closure and clarity. If your writing is more comfortable with ambiguity, look at every paragraph ending. If it ends with a declaration, consider replacing it with a question, an image, or a withheld conclusion.
4. Transition vocabulary. "Furthermore," "moreover," "additionally," "in conclusion," "it is worth noting" — these are AI transition staples. Delete all of them and either write your own connection or, if the logic is obvious, cut the transition entirely. Your own transition vocabulary is more specific and probably more idiosyncratic.
5. Specificity level. AI tends toward the category level: "a bird," "a building," "an emotion." Your prose lives at the instance level: "a red-tailed hawk," "a pre-war tenement on Myrtle Avenue," "the specific anxiety of someone about to speak in front of a crowd they have underestimated." Go through the AI draft and push every abstraction toward a specific instance.
Turnitin's AI detection research team, in a white paper released in August 2023, documented that the most commonly flagged AI prose patterns were not rare vocabulary items but structural signatures: paragraphs that averaged similar lengths, consistent use of three-part sentence structure, and transition word density above the 90th percentile of human academic writing. Editors at major magazines reported independently that AI copy "felt smooth in a way real writing rarely is" — meaning the absence of idiosyncratic rough edges was itself the tell.
The Turnitin finding points to something important: your voice is partly constituted by what a statistical model would never generate — the specific imperfections, departures from convention, and personal tics that feel wrong by average standards but right by the standards of your particular sensibility.
David Sedaris's run-on sentences that break grammatical rules. Zadie Smith's intrusive narrator who interrupts the fiction to comment on it. Joan Didion's refusal to provide context before emotion. These are rough edges. They are what editing AI prose toward smoothness removes, and what editing AI prose toward your voice requires you to add back in.
A practical exercise: take a paragraph you have written in the past that you consider among your best. Count the "grammatical violations" in it — comma splices, fragments, unconventional punctuation. Now look at the AI draft you are editing. Count the same. If the AI draft has fewer violations, it is too smooth. Add some of yours back in.
Bring an AI-generated paragraph (or ask the assistant to generate one for you to practice on). Together you'll work through the five-layer edit: vocabulary, sentence rhythm, resolution habit, transition vocabulary, and specificity. The assistant will identify which layers need work and suggest revision strategies — but you make the actual word choices.
In January 2023, Clarkesworld — one of the most prestigious science fiction magazines — closed its submissions window after receiving a flood of AI-generated stories. Editor Neil Clarke documented the numbers publicly: in the three months following ChatGPT's public release, AI-generated submission attempts had increased by approximately 800 percent. Clarke could identify most of them by style markers, but not all. He reopened submissions with a new policy: any story suspected of AI generation would be rejected without feedback, and submitters found to have disguised AI authorship would be permanently blacklisted.
Within the same month, three other SFWA (Science Fiction and Fantasy Writers of America) affiliated publications — Beneath Ceaseless Skies, Strange Horizons, and Fantasy & Science Fiction — posted explicit AI prohibitions. The prohibitions were not primarily about copyright; they were about voice. As Strange Horizons stated: "We publish writers, not outputs."
Disclosure norms for AI-assisted writing are evolving rapidly and vary significantly by context. Understanding what each context requires is now a basic professional competency for any working writer.
Academic writing: The Modern Language Association (MLA) and the American Psychological Association (APA) both issued guidance in 2023 requiring disclosure of AI tools used in the preparation of manuscripts. The MLA stated that AI-generated text should be treated as a source requiring citation. The APA required that any use of AI tools be described in the methods or acknowledgements section. Many individual universities have gone further, treating undisclosed AI use as academic dishonesty equivalent to plagiarism.
Journalism: The Society of Professional Journalists (SPJ) added AI guidance to its Code of Ethics in 2023: reporters must disclose AI use to editors, and publications must disclose AI-generated or AI-substantially-assisted content to readers. The Associated Press style guide now includes entries on AI attribution.
Commercial and creative writing: Publishing contracts signed from mid-2023 onward have increasingly included "AI warranty" clauses in which the author certifies that the work is not substantially AI-generated. The Authors Guild's model contract language specifies that the author must disclose if AI tools were used in the composition of more than a de minimis portion of the work.
The phrase "de minimis" — meaning "too small to matter" — appears in multiple emerging AI disclosure frameworks, but no consensus definition exists yet. Using AI to suggest a synonym is almost universally considered de minimis. Using AI to draft a full section and editing it is not. The practical guidance: when in doubt, disclose. A brief acknowledgement ("AI tools were used in drafting this piece") costs nothing and protects against accusations of bad faith.
In February 2023, the U.S. Copyright Office issued a guidance statement on AI-generated works: content generated entirely by AI is not copyrightable. Copyright requires human authorship. In August 2023, the Office clarified for the graphic novel case Zarya of the Dawn (Kris Kashtanova): the AI-generated images were not copyrightable, but the human-authored text and arrangement were. This is the current legal framework in the U.S.: your creative decisions are copyrightable; the AI's text generation is not.
The practical implication for writers: the more you edit, restructure, and rewrite AI output, the stronger your copyright claim to the result. A lightly edited AI draft may have ambiguous copyright status. A heavily revised AI draft that retains few of the original AI sentences is substantially your own work. This is another reason — beyond voice — that the five-layer edit matters.
The instinct to conceal AI use is understandable but strategically counterproductive. Editors, academic reviewers, and readers are becoming more capable of detecting AI-influenced prose. Being discovered to have concealed AI use after the fact causes far more reputational damage than proactive disclosure. The writers who have navigated this period most successfully — Marche at The Atlantic, Sloan with his documented process, journalists at publications that have developed explicit AI policies — are those who have been transparent about their workflow and specific about what AI did and did not contribute.
Before submitting any AI-assisted work, ask three questions: (1) Does the publication or institution have an explicit AI policy? If yes, follow it. (2) Have I substantially transformed the AI output through revision? If yes, the work is primarily yours. (3) Would a reasonable reader or editor consider my disclosure sufficient? If uncertain, add more disclosure. Transparency is never the wrong choice.
Disclosure norms vary by context, and the right decision is not always obvious. In this lab, describe a specific writing situation — academic paper, magazine pitch, fiction submission, commercial copy, blog post — and the degree of AI involvement you used. The assistant will help you determine what disclosure is required, recommended, and sufficient given current guidelines from the relevant professional bodies.