When the Commonwealth Foundation paused parts of its 2025 Short Story Prize process amid AI suspicions, the loudest takes split into two tribes. One side wanted disqualifications and detection tools. The other called the panic a moral hysteria. Both missed the real problem. The literary world has been operating for decades on an unspoken agreement about authorship that AI has now forced into the open.
At YouWrite, we work on the seam between human intent and machine assistance every day, including in our story work, so we have a stake in being honest about this. Nobody, including us, has a clean definition of where a writer ends and a tool begins.
Two definitions of authorship, both old, both incompatible
Long before transformer models, two ideas of authorship coexisted by accident.
The first is craft-as-labor. You typed every sentence. You picked every comma. The text on the page is a direct readout of your nervous system, and that physical act of composition is what makes it yours. This is the model behind most MFA pedagogy, behind the romantic image of the writer at the desk, and behind the gut reaction that prompting a model feels like cheating.
The second is craft-as-vision. Authorship is the chain of decisions that shape a work: what to write about, whose voice to use, what to cut, what to keep, what counts as good. Under this model, a director who never operates a camera is still the author of the film, and a novelist who dictates to a typist is still the novelist.
Both definitions have always been true in practice. Raymond Carver's published stories were heavily reshaped by Gordon Lish, sometimes cut by more than half, as the New Yorker documented in Charles McGrath's 2007 piece on the Lish archive at Indiana University. Thomas Wolfe relied on Maxwell Perkins to the point that critics like Bernard DeVoto attacked him for it in 1936's "Genius Is Not Enough." Ghostwriting is an entire industry. Translators rewrite. Editors restructure.
The craft-as-labor camp tolerated all of this because the labor still happened somewhere, by someone human, under contract. AI breaks that compromise. The labor can now happen with no human hand at the keyboard for long stretches, while the vision remains, theoretically, the writer's.
Why the two camps cannot hear each other
When a craft-as-labor reader sees a sentence they suspect was generated, they feel something close to betrayal. The contract was: you suffered for this, so I will spend my attention on it. Generated prose, even good generated prose, violates the contract before the question of quality even arises.
When a craft-as-vision writer hears that complaint, it sounds like fetishism. They will point out that they revised the output forty times, threw away three drafts, rewrote the ending, and shaped every beat. To them, the keystrokes are bookkeeping. The work is the judgment.
Neither side is lying. They are measuring different things.
Detection is not the answer, and pretending otherwise is dishonest
The Commonwealth response, like most institutional responses, gestured at detection. Everyone should be skeptical here, including writers who want AI out of their competitions.
OpenAI quietly retired its own AI Text Classifier in July 2023, citing low accuracy. A 2023 study by Weixin Liang and colleagues at Stanford, published in Patterns, showed GPT detectors flagging writing by non-native English speakers as machine-generated at rates above 60 percent while clearing native speakers. Turnitin has repeatedly walked back confidence claims on its own detector. Running suspect stories through these tools after the fact is, charitably, theater.
That leaves judges, editors, and prize committees with a problem detection cannot solve. If you cannot reliably tell, the rule against AI use becomes a rule against getting caught, which selects for the most sophisticated users, not the most human ones.
What actually separates a human-directed story from generated filler
Readers can often feel the difference even when they cannot prove it. The difference is not stylistic fingerprints. It is the presence of decisions a model would not make on its own.
- A detail that is specifically wrong in the way memory is wrong.
- A structural choice that hurts the story's readability but serves its meaning.
- An ending that refuses the satisfying shape the setup promised.
- A sentence that exists because the writer could not let go of it, not because it earns its place.
Generated filler optimizes for the average expectation of the prompt. Human-directed work makes specific, often costly, choices against that average. You can use a model heavily and still produce the second kind, and you can type every word yourself and produce the first.
This is the part the loudest voices on both sides leave out. The meaningful axis is not human versus AI. It is intentional versus default.
A working framework for writers
If you want to use AI assistance and still call the work yours, in a way you can defend out loud, three tests are useful.
- The deletion test. Could you cut any AI-suggested sentence and replace it with something better, in your own voice, on demand? If no, the model is not assisting you, it is carrying you.
- The why test. For each significant choice in the piece, can you say why it is there, in terms of the story's meaning rather than the model's habits? If a beat exists because the model produced it and it sounded fine, it is filler.
- The taste test. Did you reject more output than you kept? Working writers throw away most of what they generate, by hand or by prompt. A high acceptance rate is a warning sign.
None of these are detectable by software. All of them are obvious to the writer, and usually obvious to a careful reader within a few pages.
Where YouWrite fits, and where it does not
We are not neutral here, so the disclosure matters. YouWrite is built to help people who have a story but not the time, training, or stamina to draft it cleanly. That is real work we can do well, particularly memoir and personal narrative where the source material is the client's lived experience and our job is shaping.
Where we are weaker: we are not the right tool for someone entering a literary prize that bans AI assistance, and we should not pretend otherwise. Sudowrite is better positioned for fiction writers who want a co-writing surface and full control of every paragraph. NovelCrafter offers more structural scaffolding for long-form planners. ChatGPT and Claude, used directly, are cheaper and more flexible if you already know what you want. Our value is editorial judgment and a finished artifact, not maximal authorial control.
The Commonwealth case will be re-litigated by the next prize, and the next. The institutions will keep reaching for detectors that do not work. The work for writers is deciding in advance which definition of authorship you are operating under, and being able to say it plainly when someone asks.