Signs of AI Writing: Six Tells, Why None Is Proof, and the Craft Move That Answers Each
Teachers have learned to read for the signs of AI writing the way they once learned to read for a copied paragraph: a change of register, a claim without a page number, a quotation that is not quite in the book. This guide names the six tells that experienced readers actually notice, explains why each is a sign rather than proof, and pairs every sign with the craft move that makes an essay unmistakably a person's. It is written for students who want to write better and for readers who want to judge fairly.
This is not a guide to slipping past a detector. Every tell below is also a weakness in the writing, and the cure is the same whether a machine or a tired student produced it.
In short: The six signs of AI writing are uniform sentence length, summary where analysis belongs, hedged lists of three, quotations that do not exist or are misattributed, vocabulary tics, and the absence of a position. Each appears in honest human writing too. A cluster starts a conversation; only a fabricated quotation comes close to proof.
The six signs of AI writing, and why each is only a sign
1. Uniform sentence length
Generated prose settles into a rhythm: sentence after sentence of roughly the same length, each carrying one idea, none interrupting the last. Readers feel it as evenness before they can name it. It is a sign because language models optimise for local coherence, and it is not proof because every student writing a timed exam, a lab report or an essay in a second language produces the same evenness. Liang and colleagues at Stanford showed that seven detectors flagged more than half of a set of essays by non-native English writers precisely because constrained vocabulary and sentence structure look statistical.
2. Summary where analysis should be
The paragraph retells the plot and then announces what it "shows." Nothing is argued; events are listed and a theme is attached. Generated text does this because summary is the safest thing to say about any book. Students do it because summary is easier than argument, and every English teacher has been correcting it since long before 2022.
3. Hedged lists of three
"The character could be seen as isolated, conflicted and ultimately tragic." Three adjectives, one hedge, no evidence. The triplet is a rhythm models fall into, and the hedge protects a claim the writer has not tested. It is a tell because it recurs sentence after sentence; it is not proof because the five-paragraph essay taught a generation of students exactly this move.
4. Quotations that do not exist or are misattributed
This is the strongest sign and the only one that approaches proof. Generated text quotes what is commonly quoted and sometimes quotes what was never written: a line that sounds like Fitzgerald but appears nowhere in Gatsby, or the famous "to thine own self be true" credited to Hamlet, when it belongs to Polonius, the play's windbag, in the first act. A student with the book open does not invent a line. A student writing from memory sometimes misattributes one, which is why even this sign calls for a question before a conclusion.
5. Vocabulary tics
Certain words cluster in generated text at rates human writers rarely match: "multifaceted," "underscores," "nuanced," "pivotal," "showcases," "it is important to note," "in the modern world." Readers who see essays in bulk notice the clustering. It is a sign, and a weak one on its own, because these words also live in thesauruses, teacher handouts and the model answers students are shown.
6. No position
Read the paragraph and ask whether anyone could disagree with it. Generated analysis is engineered so that nobody can: it ends on "this highlights the complexity of the human condition," which is true of every book ever written. Human writing that has been through a draft or two takes a side. Human writing under time pressure often does not, which is why this sign, too, is a prompt for conversation rather than a verdict.
| Sign | Why it appears in generated text | Why it is not proof | The craft move |
|---|---|---|---|
| Uniform sentence length | Models optimise for local coherence | Exam writing, lab reports and second-language prose are even too | Let a short sentence land what a long one set up |
| Summary instead of analysis | Summary is the safest statement about any text | Most weak human essays summarise | Replace "this shows" with a claim about how a specific line works |
| Hedged lists of three | Triplets are a learned rhythm; hedges avoid error | The five-paragraph essay teaches the same move | Pick one adjective, prove it with a quotation, drop the hedge |
| Fabricated or misattributed quotations | Models reproduce what is commonly quoted | Memory misattributes too, though it rarely invents | Quote from the page, with act, scene or chapter, and check it twice |
| Vocabulary tics | High-frequency phrases in training data | Thesauruses and model answers spread the same words | Use a concrete noun where an abstract one sits |
| No position | Unfalsifiable claims are never wrong | First drafts and timed essays often lack a stance | State the reading a classmate could argue against, then admit the counter-reading |
How teachers actually read for the signs of AI writing
Experienced readers do not run a checklist. They read the whole class's essays in one sitting, which makes the tics visible by repetition, and against what they already know of each writer, which makes a change of register visible by contrast. They check quotations against the text. Then they ask questions, because a student who wrote a paragraph can explain a word choice in it and a student who did not cannot. Our guide to whether teachers can tell if you used ChatGPT covers that conversation from the teacher's side.
The fair reader also knows the evidence on the tools. Weber-Wulff and colleagues, testing 14 detectors for the International Journal for Educational Integrity in 2023, concluded they were "neither accurate nor reliable." Turnitin's own documentation lists the human writing most likely to be flagged: "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas." That is a description of signs one, two and six, produced by a human, and it is why a single sign should never carry an accusation. A cluster of signs justifies a question. A fabricated quotation justifies a harder question. Nothing on this list, alone, justifies a finding, and the tools that automate it inherit the same limit; what AI detectors look for sets out what they measure.
The constructive flip: the craft move for each sign
Every tell has an opposite, and the opposite is simply good writing about literature. Take them in turn.
Against the uniform sentence, the deliberate short one. A twenty-eight-word sentence that traces how Ralph's authority in Lord of the Flies depends on an object the boys have agreed to respect needs a four-word follow-up: "Then the conch breaks." Rhythm is a consequence of thinking in stages, and a writer who has stages will have rhythm.
Against summary, the arguable claim. "Golding shows the boys becoming savage" is a summary wearing a thesis costume. "The conch loses its power before it is destroyed, because Jack has already stopped waiting for it" is a claim someone could dispute, which is what makes it a thesis. Our guide to building a thesis that is actually an argument exists for this move.
Against the hedged triplet, one adjective and one quotation. Against the vocabulary tic, a concrete noun: the conch, the glasses, the fire, the pig's head on a stick. Abstract nouns ("savagery," "civilisation") are where generated text lives; concrete nouns are where the book lives, and a paragraph built on them cannot be written without having read it.
Against the invented quotation, the real one, checked. The Folger text of Hamlet puts "This above all: to thine own self be true" in the mouth of Polonius, lecturing Laertes in Act 1, scene 3, a few lines after "Neither a borrower nor a lender be." A student who quotes it as Hamlet's creed has not opened the play; a student who quotes it as Polonius's, and notes that the play's most famous advice comes from its least self-aware character, has written something a reader remembers.
Against the missing position, the admitted counter-reading. State your claim, then give the strongest case against it in one sentence, then answer it. "It could be objected that the conch still commands silence in the final assembly; but Piggy has to hold it up like a shield, which is the opposite of authority." No generator produces that sentence, because it requires being willing to be wrong. For the same comparison worked out on a full paragraph, see AI writing vs human writing.
Frequently asked questions
What are the most common signs of AI writing?
Even sentence length, plot summary with a theme attached, hedged lists of three adjectives, quotations that are invented or credited to the wrong speaker, recurring words such as "multifaceted" and "underscores," and paragraphs that take no position a reader could dispute.
Is any single sign proof that a text was generated?
No. Each appears in honest human writing, especially timed, formulaic or second-language work. A fabricated quotation is the closest to proof, and even that deserves a question before a conclusion.
How can I make my own writing unmistakably mine?
Quote the exact line with its location, make a claim a classmate could argue against, use concrete nouns from the text, admit the strongest counter-reading, and keep your drafts so the growth of the argument is on record.
Do vocabulary tics like "multifaceted" count as signs of AI writing?
They mean the writer has absorbed a register, which may come from a model, a handout, or a thesaurus. On their own they are the weakest sign on the list.
Conclusion: the signs of AI writing are the signs of weak writing, and the cure is the same
The signs of AI writing were the signs of thin writing before any tool existed: evenness, summary, hedging, borrowed quotation, borrowed vocabulary, no stake in the claim. That is why they are useful to a reader and dangerous as proof. A teacher who reads for them fairly asks a question; a student who reads for them in a draft revises toward the specific, the quoted and the arguable. Both are doing the work the Authorship pillar is built around: writing that could only have come from a particular person reading a particular book.
The essays on this site are study companions for reading and writing about literature; they are never submission-ready coursework.
Sources
- Turnitin Guides, "Turnitin's AI writing detection capabilities FAQs": the human writing types most likely to be flagged as false positives
- Weber-Wulff et al. (2023), "Testing of detection tools for AI-generated text", International Journal for Educational Integrity: the 14-tool test and the "neither accurate nor reliable" conclusion
- Liang, Yuksekgonul, Mao, Wu and Zou (2023), "GPT detectors are biased against non-native English writers", Patterns (arXiv preprint): more than half of 91 non-native TOEFL essays flagged by seven detectors
- Folger Shakespeare Library, Hamlet, Act 1, scene 3: Polonius's "This above all: to thine own self be true" and "Neither a borrower nor a lender be"