Signs of AI writing, and why none of them is proof

The most reliable signs of AI writing are not the ones that get shared most. Research points to grammar habits most readers never notice and to vocabulary that shifts with each new model. In fiction, the stronger signs are in how the story is built. The quickest practical clues are plainer, such as a leftover chatbot line or a source that doesn't exist. Each sign nudges the odds a little. None of them, alone or in combination, proves who wrote a particular text.

A sign is not proof

A sign of AI writing is a pattern that turns up more often in AI text than in human text. "More often" is doing a lot of work in that sentence. People produce every pattern that language models produce, because the models learned to write from people.

A little arithmetic shows why this matters. Imagine a class of 30 essays, two of them written with AI. Suppose some pattern appears in half of AI essays and in one human essay in ten. You would expect to see it in one of the AI essays and in about three of the human ones. Most of the essays showing the sign would be the students' own work.

That doesn't make signs useless. It means they are reasons to look more closely, and the research below helps separate the stronger ones from folklore.

Can you tell if text is AI-written just by reading it?

Most people can't, at least not reliably, and they tend to overrate how well they do. In a 2024 study, Johanna Fleckenstein and colleagues asked 89 novice teachers and 200 experienced teachers to pick out ChatGPT essays from essays written by students. Neither group managed it reliably, and both were overconfident in their judgements.

Practice changes the picture. At ACL 2025, Jenna Russell, Marzena Karpinska and Mohit Iyyer asked annotators to judge 300 non-fiction articles, some written by GPT-4o, Claude 3.5 Sonnet or o1-Pro. Annotators who rarely used AI for writing did about as well as chance: on average they caught 56.7% of the AI articles and wrongly flagged about half of the human ones. People who used it often for their own writing did very well. A majority vote of five of them got only one article wrong out of 300.

The experts explained every decision. Vocabulary came up in 53.1% of explanations and sentence structure in 35.9%, but originality, the use of quotes and clarity mattered too. They mentioned predictable sentence templates, items grouped in threes and upbeat but vague conclusions. One caution: the panel's strength came from voting together. Individual experts varied, and on articles rewritten by a humanizer tool, one expert caught none of them.

Clues you can check in a few minutes

Wikipedia's volunteer editors keep a field guide called "Signs of AI writing", drawn from what they have seen while cleaning up undisclosed AI text. These are the clues from it that are quickest to check:

• Leftover chatbot talk. Lines a person would normally delete: "I hope this helps", an offer to expand on a point, a note about a knowledge cutoff, or "As an AI language model" (Google's search quality rater guidelines list that one too). Stray Markdown, such as asterisks around words, is a related tell.
• Facts and sources that don't check out. References that don't exist, DOIs that lead to an unrelated paper, quotes nobody said.
• Inflated importance and vague attributions. Ordinary things given lasting significance, brochure words like "nestled" in neutral writing, and "experts say" with no expert named.
• A change of voice. Writing that sounds nothing like the person's emails, earlier essays or in-class work, with no example from their own experience. This only helps if you have genuine samples to compare against.

The guide itself warns that not all text with these indicators is AI-generated. A leftover chatbot line is strong evidence that a chatbot was involved somewhere, though not of how much.

Grammar habits: the signs of AI writing most readers miss

The most careful work on this comes from a 2025 PNAS study by Alex Reinhart and colleagues at Carnegie Mellon. Humans and language models (GPT-4o, GPT-4o Mini and four versions of Llama 3) responded to the same prompts across six genres, from academic writing to TV scripts, and the team counted grammatical features in each text.

Some of the clearest differences:

• Present participial clauses. Instruction-tuned models used them at 2 to 5 times the human rate, and GPT-4o at 5.3 times. These are openings such as "Glancing at the clock, she reached for her coat."
• Nominalizations. Models used 1.5 to 2 times as many, GPT-4o 2.1 times. Think "the implementation of the policy" where a person might write "putting the policy in place".
• Agentless passives. GPT-4o used them at about half the human rate. Sentences like "the results were published" turned out to be more of a human habit than a machine one.

The differences were larger for instruction-tuned models than for base models, and they did not shrink as the models got bigger. Classifiers built on these features typically told human text from a given model's text with 93% to 98% accuracy. That is a statement about thousands of texts, though. One participial clause in one essay means nothing.

Vocabulary: real signals with a short shelf life

Word choice is the sign people notice first, and it is real. In the same PNAS study, GPT-4o used "camaraderie" about 162 times as often as human writers, "intricate" about 119 times as often and "palpable" about 95 times as often.

The effect is visible at scale. Dmitry Kobak and colleagues, writing in Science Advances in 2025, looked at more than 15 million biomedical abstracts from 2010 to 2024. In 2024, "delves" appeared about 28 times more often than earlier trends predicted, "underscores" about 14 times and "showcasing" about 11 times. From the excess, they estimated that at least 13.5% of 2024 abstracts were processed with a language model.

The catch is that vocabulary goes out of date fast. Each model generation has its own favourite words, writers pick up words they keep reading, and a list of "AI words" from 2024 may say little about text written today.

In fiction, the signs are in the story

Stories leave a different trail. The StoryScope study (Russell and colleagues, COLM 2026) compared 10,272 human stories with versions written by five AI models from the same prompts. AI narrators spelled out the story's theme 77% of the time, against 52% for human authors. AI stories conveyed emotion through the body (a tight chest, a cold sweat) in 81% of cases against 38%, and 79% of them had no subplots, against 57% of human stories. We summarise the paper in how AI fiction differs from human stories.

Signs that mislead

Some popular signs cause more false accusations than they catch:

• Em dashes. Plenty of human writers love them, and newer models use fewer. When an old version of this site relied on them, it took a chapter of Moby-Dick for AI writing. More in em dashes don't prove AI.
• Clean grammar, a neutral tone, formal words. In the Russell study, the inexperienced annotators assumed people write more correctly than AI, labelled even-toned text as AI and treated any "fancy" word as a giveaway. The experts knew that humans make more grammatical errors than models, and looked only for the specific words models overuse.
• Plain, simple English. In a 2023 Stanford study, seven detectors flagged an average of 61.22% of TOEFL essays written by people learning English, and asking ChatGPT to simplify the word choices in US eighth-graders' essays raised the share flagged from 5.19% to 56.65%.
• Short texts. A paragraph doesn't hold enough evidence. In a Becker Friedman Institute study, results for most detectors became uneven on passages under 50 words.
• The absence of signs. In a 2026 study, leading detectors flagged just 3% of excerpts from a model fine-tuned on an author's books, against 97% of ordinarily prompted excerpts. Clean prose is not proof of a human writer.

How to tell if text is AI-written: a quick checklist

Work through these in order. The early steps are quick; the later ones take more effort and carry more weight.

1. Decide what's at stake. The higher the stakes, the more steps you need, and the more weight belongs on direct evidence such as drafts.
2. Make sure there's enough text. Our AI detector needs at least 50 words and works best from 150. Turnitin won't score fewer than 300 words of prose, and OpenAI said its early classifier was very unreliable below 1,000 characters. With only a paragraph, hold any conclusion loosely.
3. Clean out hidden characters. Text copied from PDFs, websites and chat apps can carry zero-width spaces or look-alike letters from other alphabets, which can confuse detectors. The invisible character remover strips them out. They tell you how the text travelled, not who wrote it.
4. Look for the clues above, and look up two or three of the citations, quotes or statistics.
5. Compare with the writer's known work. A sudden change in voice or fluency is worth a question. People do improve, and some get help from friends or editors, so treat it as a reason to ask.
6. Run a detector and read the whole report. Look at the outcome (Human, Mixed, AI or Inconclusive) with its confidence word, then at the sentence map to see which passages drive it. Our report shows style signals as explanations only; the verdict comes from a classifier that scores the whole text. Like any detector, it can be wrong in both directions, so treat a result as a reason to look closer, not as proof (our methodology page explains how it is tested).
7. Look at the process and talk to the writer. Version history in Google Docs or Word shows how a document grew. Drafts, notes and sources show the work behind it. Someone who wrote the piece can usually explain a paragraph or a source without effort.

What to do with what you find

Most checks end in one of three places, and each calls for something different.

• Nothing stands out. The usual result. It isn't proof of human writing, but with no leftovers, no bad sources and no change of voice there is usually no real question to ask.
• A detector flag, and nothing else. This is the weakest case. False positives gather in short texts, plain or non-native English and formulaic writing; Turnitin, for one, lists repetitive text and text without much structural variation among the human writing its model can mistake for AI. Look at the drafts before you say anything.
• Leftover chatbot lines or sources that don't exist. This is the strongest case. Show the writer exactly what you found and ask how the text was produced. Check the rules first: if some AI help was allowed, the honest answer may be perfectly acceptable.

Whichever case you are in, write down what you checked and what you saw, and never act on signs alone. A conversation costs little and can prevent a serious mistake. For student work, our guide to checking an essay for ChatGPT walks through that conversation step by step.

Sources

1. Reinhart et al., Do LLMs write like humans? Variation in grammatical and rhetorical styles (PNAS, 2025), checked October 2026
2. Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Science Advances, 2025), checked October 2026
3. Fleckenstein et al., Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays (Computers and Education: Artificial Intelligence, 2024), checked October 2026
4. Russell, Karpinska and Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025), checked October 2026
5. Wikipedia: Signs of AI writing (WikiProject AI Cleanup), checked October 2026
6. Google: Search Quality Rater Guidelines (September 2025 version), checked October 2026
7. Russell, Rajendhran, Pham, Iyyer and Wieting, StoryScope: Investigating idiosyncrasies in AI fiction (COLM 2026), checked October 2026
8. Liang et al., GPT detectors are biased against non-native English writers (Patterns, 2023), checked October 2026
9. Jabarian and Imas, Artificial Writing and Automated Detection, BFI Working Paper 2025-116, checked October 2026
10. Chakrabarty, Ginsburg and Dhillon, Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers (arXiv 2510.13939), checked October 2026
11. Turnitin Guides: File requirements for an AI Writing Report, checked October 2026
12. Turnitin Guides: Turnitin's AI writing detection capabilities FAQs, checked October 2026
13. OpenAI Help Center: How can educators respond to students presenting AI-generated content as their own?, checked October 2026
14. OpenAI: New AI classifier for indicating AI-written text (31 January 2023, updated 20 July 2023), checked October 2026
15. Karr, Khvatskii, Hua and Chawla, Why AI Detection Fails for Academic Integrity (ACM AI Summit 2026), checked October 2026