When a writer is accused of using AI, the interesting question is rarely whether the accusation is true. The interesting question is who made it, how quickly, on what evidence, and whether the same evidence would have been read differently coming from a different name. YouWrite builds tools that help people draft with AI, so we have a professional interest in this question being asked seriously rather than dismissed as sour grapes.
The Falade case, briefly
In November 2024, novelist Rex Ogle publicly accused Nigerian debut author Fikayo Falade of using ChatGPT to write her Substack essay about racism in publishing. The accusation spread on Bluesky and X within hours. Falade, a Black woman writing about a Black experience, was asked to prove her humanity to strangers who had decided her sentence rhythms were suspicious. Ogle later deleted his posts and apologized. By then the essay had been reframed in thousands of feeds as probably fake.
Compare that to how the industry handled Raven Leilani, Sally Rooney, or any number of establishment-adjacent writers whose prose has been called, at various points, clean, spare, precise, or workshop-inflected. Those are the same features that get called "AI-sounding" when they appear under an unfamiliar name. The descriptor changes with the byline.
A contrasting example
When British Vogue faced questions in 2023 about AI-generated imagery in a shoot featuring model Alexsandrah, and separately when questions surfaced around AI use in various establishment outlets, the framing was almost always about process, ethics, and disclosure. The accused party was extended the courtesy of a stated position before judgment. The Falade thread opened at the verdict.
What "sounding AI" actually encodes
Large language models were trained overwhelmingly on English-language internet text weighted toward American and British sources. Their default register is a kind of polished, mid-Atlantic explanatory voice. When a writer, especially one who learned English through formal instruction rather than through American cultural immersion, produces prose that is grammatically clean and slightly formal, it can pattern-match to that default register.
This is the trap. The features people flag as AI, such as balanced clauses, careful transitions, absence of slang, are also features associated with:
- Writers for whom English is a second or third language and who were taught to write formally
- Writers trained in traditions where rhetorical polish signals respect for the reader
- Writers who edit heavily because they cannot afford a public mistake
All three categories disproportionately describe writers from outside the Anglo-American publishing core. The detection heuristic, human or algorithmic, punishes them for the exact craft habits their circumstances taught them.
The detector literature is worse than most people think
A 2023 study by Weixin Liang and colleagues at Stanford, published in Patterns, tested seven GPT detectors against essays by non-native English speakers. The detectors flagged more than half of the TOEFL essays as AI-generated. The false positive rate for essays by native English speakers was near zero. The paper is titled "GPT detectors are biased against non-native English writers" and it appears in Patterns, volume 4, issue 7.
Turnitin's own published false positive rate, as reported by The Washington Post in April 2023, sits around one percent at the sentence level, which the company frames as low. In a class of 100 students writing 20 sentences each, that is 20 false accusations. OpenAI quietly retired its own AI text classifier in July 2023, citing low accuracy.
So the tools are unreliable in a patterned way. And the humans invoking them tend to trust their gut before they check the tool, and trust the tool before they check with the writer.
Who the gatekeepers are
Roxane Gay has written repeatedly, in her newsletter The Audacity and in her New York Times column, about the way publishing's gatekeeping functions are concentrated in a small demographic slice that then decides what counts as authentic, what counts as derivative, and what counts as suspicious. Lee and Low's Diversity Baseline Survey has documented for a decade that U.S. publishing staff remain roughly 75 percent white, with editorial and marketing skewing higher.
The accusation economy on literary social media inherits that composition. The people whose reactions travel fastest are often the ones with the largest platforms, and those platforms were built inside the same networks. When they say a piece of writing feels off, the feeling has a demographic shape whether or not the person feels comfortable saying so.
The self-critical part
YouWrite exists because AI drafting is a useful practice. That fact complicates the story I am telling. Some accusations are correct. Some writers do pass off unedited model output as their own work and misrepresent the process. Pretending otherwise would insult everyone reading. What we push back on is the enforcement pattern, not the existence of enforcement. Our own product contributes to the ambient anxiety about authorship, which we should own rather than dodge. A tool that makes drafting faster also makes accusation easier, because the accuser can always say "it could have been you."
What honest scrutiny would look like
A fair standard for public AI accusation would require, at minimum, that the accuser name specific textual features rather than gesture at a vibe, disclose whatever detector they ran and its known error profile, and contact the writer privately before posting. None of this is exotic. It is the standard any newsroom applies before publishing a plagiarism allegation. Social media strips those steps out and calls the result accountability.
The deeper fix is slower. It involves widening the pool of people whose taste sets the default for what "real writing" sounds like, so that a Nigerian debut novelist writing in careful English is not read against a template that was never built with her in mind. Roxane Gay's point about gatekeeping was never that gatekeepers are villains. It was that a narrow gate produces a narrow sense of what belongs on the other side of it.
What to do if you are the one who is skeptical
Before posting, ask yourself what specifically triggered the read. If the answer is "the prose is too clean," sit with that a while. Cleanness is not evidence. Ask whether you would have flagged the same paragraph under a byline you already trusted. If the honest answer is no, the problem is not in the paragraph.
If you are the writer being accused, keep your drafts, your notes, your voice memos, your browser history for the research session. The burden of proof has been informally reassigned to you. That is unfair and it is also the current weather. Writers who tell their own story on their own terms, with receipts, tend to weather it better. That is one reason we built YouWrite's story tools around drafting with a preserved trail rather than a black-box output.