AI Writing vs Human Writing: What a Close Reader Notices That a Detector Cannot
Most discussions of AI writing vs human writing are written from the detector's chair: perplexity, burstiness, a probability score. This one is written from the reader's chair, the one your teacher sits in. A close reader does not count sentence lengths. A close reader notices how a paragraph handles evidence, quotation, hedging, rhythm, error and position, and those six habits separate generated text from human-written text far more reliably than any percentage.
Below, one claim about Charlotte Perkins Gilman's "The Yellow Wallpaper" is written both ways.
In short: Generated text lists, hedges, summarises and stays safe; human-written analysis quotes a specific line, risks an arguable claim, varies its rhythm, and sometimes gets something wrong in an interesting way. The difference is a relationship between a writer and a text, and that relationship shows on the page.
AI writing vs human writing: six things a close reader checks
Evidence
Generated prose gestures at evidence. It says a character "is portrayed as controlling" and moves on, because the model has learned that essays contain such sentences. A human writer who has read the story points at a moment: the scene, the line, the word. Evidence in human writing is located; in generated writing it is asserted.
Quotation
The clearest single difference is the quotation mark. Generated analysis quotes rarely, and when it does the quotation is famous, short and often slightly wrong, because the model is reproducing what is commonly quoted rather than what is on the page. A student with the book open quotes the unglamorous line nobody else uses, and gets it exactly right because it was copied from a numbered page.
Hedging
Generated text hedges by reflex: "could be seen as," "may suggest," "in many ways." The hedges protect the model from being wrong about a text it has not read. Human writers hedge selectively, at the one point where they are genuinely uncertain, and commit everywhere else.
Sentence rhythm
Machine paragraphs tend toward the same medium-length declarative sentence repeated, each carrying one idea, each ending on a full stop. Human paragraphs speed up and slow down. A twenty-word sentence sets up a claim; a four-word sentence lands it. This is the feature detectors try to measure as burstiness, and a reader hears it without measuring anything.
Error
Human writing contains the errors of someone thinking: a claim pushed one step too far, a quotation read against its obvious meaning, a paragraph that changes its mind halfway through. Generated writing contains the errors of nobody thinking: a wrong act number, a character given the wrong line, a confident summary of a scene that does not exist. The first kind of error is a sign of authorship; the second, of its absence.
Position
Ask of any paragraph: could a reasonable reader disagree with this? Generated analysis is built so that nobody can. It ends on "this highlights the themes of gender and mental illness," which is true, unfalsifiable and empty. Human analysis takes a side, which means it can lose, which is what makes it worth reading.
The same claim written both ways
Both paragraphs below claim that John, the narrator's husband in "The Yellow Wallpaper," harms her. The first is written in the flat, hedging, list-everything register generated essays fall into; the second is a close reading. Neither was produced by a tool. The point is the difference a reader sees.
| Generated-style paragraph | Close-read paragraph |
|---|---|
| In "The Yellow Wallpaper," John is an important character who represents the medical and patriarchal attitudes of the time. He is a physician and the narrator's husband, and he does not believe that she is really sick. Throughout the story, John treats the narrator in a way that could be seen as controlling, dismissive and condescending. He does not allow her to write, work or socialize, which may contribute to her decline. Ultimately, John's treatment of the narrator highlights the themes of gender roles, mental illness and the dangers of ignoring women's voices. This shows that the story is a critique of how women were treated in the nineteenth century. | John's cruelty in "The Yellow Wallpaper" arrives entirely in the vocabulary of love, and that is what makes it so hard to name. When the narrator asks to move to a prettier room downstairs, he "took me in his arms and called me a blessed little goose," and then explained why she could not have it. He is a physician "of high standing," and his prescription is that she is "absolutely forbidden to 'work'"; her own quotation marks around "work" are the only protest she is allowed. I think the story's real target is the tone rather than the treatment. A rest cure delivered as a command could be resisted. Delivered as tenderness, by a man she calls "very careful and loving," it cannot be, because you cannot lock a door against someone who is carrying you upstairs. Gilman never lets John raise his voice. That is the horror. |
Run the six checks on this AI writing vs human writing pair. The left column asserts ("controlling, dismissive and condescending") and never locates; the right column quotes four phrases a reader can find on the page. The left hedges twice in one sentence; the right hedges once, with "I think," and then commits. The left column's sentences are all between fifteen and twenty-five words; the right runs from four words to thirty-eight. The right column risks an error a teacher could argue with (is tone really the target rather than the treatment?) and takes a position that could be wrong, which is exactly why it reads as written by somebody.
Why the human version reads as authored
Shakespeare gives the best account of this in the second scene of Hamlet. Pressed on why his grief seems so particular, Hamlet lists the outward signs, the inky cloak, the "customary suits of solemn black," the sighs and tears, and dismisses them all: "These indeed 'seem,' / For they are actions that a man might play; / But I have that within which passes show." The stylistic tells of human writing are the suits of woe. They can be imitated, which is why a detector that counts them is fooled in both directions, flagging the plain, careful student and passing the polished machine. What cannot be imitated is "that within": the relation between a reader and a text, visible as the choice of that quotation, the willingness to be wrong about it, and the sentence that changes speed because the thought did.
The research on AI writing vs human writing supports the reader over the detector. Liang and colleagues at Stanford found that seven detectors flagged more than half of a set of essays by non-native English speakers as machine-written, because limited vocabulary produces low perplexity whoever is typing. Turnitin's own documentation names the human writing most likely to be flagged: "content without a lot of structural variation, text that literally repeats itself." Those describe the left column above, and a human being can write it. Surface tells are evidence of register; authorship is evidence of reading, and only a reader can see it.
This is also why close reading and generated analysis diverge so sharply on literature in particular. A close reading, in the sense our guide to close reading as a method lays out, is an argument that could only have been made by someone who noticed a specific word. Generated analysis is an argument that could have been made without opening the book. Our reading of The Yellow Wallpaper builds from single words: "creep," "smooches," "gnawed." No summary of the story contains them.
What AI writing vs human writing means for a student who wants their writing to be unmistakably theirs
The answer is craft, the same craft teachers asked for before any of this existed. Quote the line nobody quotes, and quote it exactly, with the page. Make one claim a classmate could argue against, and defend it rather than softening it. Let a short sentence follow a long one when the thought calls for it. Admit the strongest reading against your own in a sentence, then answer it. Keep the drafts in which your thesis got narrower, because the trail from "John is controlling" to "John's tenderness is the cure's real instrument" is the record of an author at work, and it exists nowhere in generated text.
Two things follow from that. First, it is worth learning the signs of AI writing as a reader learns them, so that you can recognise the flat register when it creeps into your own drafts and revise toward the specific. Second, the way to make writing on a particular work provably yours is to do the reading in public, on the page, which is what our guide to writing a Yellow Wallpaper essay that is yours walks through step by step. None of this is a technique for passing a detector; it is what good literary writing has always been, and it reads as human for that reason.
Frequently asked questions
What is the biggest difference in AI writing vs human writing in an essay?
Position and quotation. Human analysis quotes a specific, often unfamous line and takes a stance a reader could dispute; generated analysis paraphrases, hedges, and ends on a theme statement nobody could disagree with.
Can a human write in the "AI" register by accident?
Yes, and students do it constantly when summarising under time pressure or writing in a second language. That is why detectors produce false positives and why the reader's six checks are more reliable than a score.
Is varied sentence length enough to make writing read as human?
No. Rhythm is a surface feature and can be imitated. What reads as authored is the relationship to the text: the chosen quotation, the arguable claim, the admitted counter-reading. Rhythm follows from thinking; it cannot substitute for it.
Conclusion: AI writing vs human writing is a question for readers
Set side by side, AI writing vs human writing stops being a question about statistics. One paragraph lists and hedges and summarises; the other quotes "blessed little goose," argues that tenderness is the instrument of harm, and lets a four-word sentence carry the weight. A detector sees sentence lengths. A reader sees a person who read the story. The whole of the Authorship pillar rests on that difference: write from the text, keep the trail, and the writing will be yours in a way no score can take from you.
The essays on this site are study companions for reading and writing about literature; they are never submission-ready coursework.
Sources
- Gilman, Charlotte Perkins, "The Yellow Wallpaper" (1892), Project Gutenberg: all quotations from the story, including "blessed little goose," "of high standing," "absolutely forbidden to 'work'" and "very careful and loving"
- Folger Shakespeare Library, Hamlet, Act 1, scene 2: "These indeed 'seem,' / For they are actions that a man might play; / But I have that within which passes show"
- Liang, Yuksekgonul, Mao, Wu and Zou (2023), "GPT detectors are biased against non-native English writers", Patterns (arXiv preprint): seven detectors flagged more than half of 91 non-native TOEFL essays, driven by low perplexity
- Turnitin Guides, "Turnitin's AI writing detection capabilities FAQs": the human writing types most likely to be flagged, "content without a lot of structural variation, text that literally repeats itself"