Why AI Detectors Flag Human Writing (And Which Habits Trigger Them)

Detectors score how predictable your text looks, not who wrote it. Here are the signals they read and the honest writing habits that can set them off.
AI detectors don't know who wrote your text. They score how statistically predictable it is, compared with examples of human and machine writing. That is why AI detectors flag human writing: clear structure, common words, even sentence rhythm and heavy grammar-tool polish all look predictable, and plenty of honest writers produce exactly that.
This post explains what detectors measure, which ordinary habits produce those signals, and why rewriting yourself to "beat" a detector is the wrong fix. The before/after sentences below are illustrations I wrote for this post, not real student work and not test results.
What does an AI detector actually measure?
An AI detector is a piece of software that has learned what machine text tends to look like. It cannot see your browser history, your drafts or your chat logs. It only gets the finished words, and it has to guess from them.
Two ideas come up in almost every explanation: perplexity and burstiness. They matter, but they are only part of the story in 2026.
Perplexity: how surprising each word is
A language model reads a sentence and, at each point, predicts which word is likely to come next. Perplexity measures how surprised the model is by the words that actually appear. Low perplexity means the text kept doing what the model expected.
GPTZero's own explainer uses this kind of example: after "Hi there, I am an AI", the word "assistant" is very predictable, while "detector" is less so. Text built from the most expected word at every step reads as machine-like, because that is roughly how a chatbot writes.
Burstiness: how much that predictability varies
Burstiness measures how much perplexity rises and falls across a document. People tend to write unevenly: a long winding sentence, then a short one, then a strange word choice, then a cliché. GPTZero describes models as choosing words by the same rule throughout, which produces low burstiness.
Put together, the old rule of thumb was simple. Low perplexity plus low burstiness looked like AI. High and uneven looked human.
Most detectors now use trained classifiers, not a single formula
Here is the part many explainers leave out. GPTZero, the company that popularised the perplexity and burstiness framing, says it moved to a deep-learning model in autumn 2023. Its support pages now describe perplexity and burstiness as history, or at most one signal among several.
Modern detectors are mostly classifiers. They are trained on large piles of text labelled "human" or "AI" and learn whatever features separate the two: word choice, sentence shapes, transitions, paragraph structure. GPTZero's machine learning staff describe it in roughly those terms, as learning patterns of style, structure and word choice from both kinds of text.
That matters for you in two ways. First, nobody outside the company knows exactly which features a given detector weighs. Second, any human style that happens to resemble the training examples of AI text can be scored as AI, whether or not it is predictable in the narrow perplexity sense.
Why AI detectors flag human writing in the first place
The core problem is that the signal detectors use is not unique to machines. Predictable, tidy, well-organised prose is also what teachers, style guides and editors ask people to write.
Think about the advice a first-year student gets. Use a clear thesis. Open each paragraph with a topic sentence. Signpost your argument. Avoid slang. Keep sentences readable. Every one of those rules pushes text toward the centre of the distribution, which is exactly where chatbot output sits. Chatbots learned to write from enormous amounts of human text, so the most typical way a careful person writes is, almost by definition, close to what a model produces.
Detectors make a probability guess about a writing style and present it as a judgement about a person. Even small error rates add up at scale. When Vanderbilt University switched off Turnitin's AI detector in August 2023, it pointed out that a 1% false positive rate, applied to the roughly 75,000 papers it had submitted the year before, would have wrongly flagged around 750 of them.
The tool makers have been open about some of this. OpenAI launched its own AI text classifier in January 2023 and withdrew it on 20 July 2023, citing its low rate of accuracy. On OpenAI's own challenge set, it caught 26% of AI-written text and labelled human text as AI 9% of the time, and OpenAI warned it was very unreliable on anything under 1,000 characters.
Why famous old texts get flagged too
A widely shared example is the Declaration of Independence being scored as AI-generated. The reason is instructive. Documents like that appear many times in the data language models learn from, so a model predicts them almost word for word. To a perplexity measure, memorised text looks extremely unsurprising, which reads as "machine". The detector is measuring familiarity, and familiarity is not authorship.
What makes text look AI generated to detectors?
Each signal a detector relies on has a perfectly ordinary human cause. This table pairs them up.
Signal a detector reads as "AI" | What it looks like on the page | Ordinary human reasons for it |
|---|---|---|
Low perplexity (predictable words) | Common, safe word choices; stock phrases | Writing in a second language, plain-language training, writing for a young audience |
Low burstiness (even rhythm) | Sentences of similar length and shape | Following a template, writing under time pressure, trained academic style |
Formulaic structure | Thesis, three body points, recap conclusion | Five-paragraph essay teaching, report templates, exam technique |
Heavy signposting | "Firstly", "Moreover", "In addition", "Overall" | Taught as good practice in many schools and ESL courses |
Generic, hedged claims | "Plays a crucial role", "various factors" | Writing about a topic you know only from reading, word-count padding |
Over-polished surface | No typos, no fragments, perfectly regular grammar | Grammar checkers, careful proofreading, professional editing |
Familiar or memorised text | Famous quotations, standard definitions | Quoting sources, legal or technical boilerplate |
None of these is evidence of anything. Human text flagged as AI almost always shows several of them at once. A sentence can be predictable because a person chose the clearest way to say something. The detector has no way to tell that apart from a chatbot choosing the most likely token.
Which writing habits look robotic to a machine?
These are the habits that most often produce the signals above. I'm not listing them so you can scrub them out. Several are good habits. The point is to understand why an honest writer can be misread, and what to say if it happens to you.
1. Writing to a template
Five-paragraph essays, lab report formats and "problem, solution, benefit" business templates all produce the same skeleton every time. Chatbots love that skeleton too, because they saw millions of examples of it.
Illustration, before: "In this essay, I will discuss three reasons why remote work benefits employees. First, I will examine flexibility. Second, I will consider cost savings. Finally, I will explore wellbeing."
Illustration, after: "My sister saves about an hour a day by not commuting, and she spends most of it asleep. That trade, time for rest, is the strongest case for remote work, and it's the one I'll start with."
The first version is fine for an exam. It is also the shape a model produces when asked for an essay outline.
2. Stacking stock transitions
"Furthermore", "Moreover", "Additionally", "In addition", "Overall". Many students are explicitly taught to use these, especially in English-as-a-second-language courses. Models use them constantly, so they have become a strong statistical marker.
Illustration, before: "Furthermore, social media affects sleep. Moreover, it can increase anxiety. Additionally, it reduces face-to-face contact."
Illustration, after: "Social media eats into sleep, which feeds anxiety, which makes face-to-face plans feel like more effort than they're worth."
The second version links the ideas through cause and effect instead of announcing each new point.
3. Keeping every sentence the same length
Some writers settle into a comfortable rhythm of 15 to 20 words per sentence, subject first, one clause each. It reads smoothly. It also has very low burstiness.
Illustration, before: "The experiment tested plant growth under blue light. The plants grew faster than the control group. The leaves were also darker in colour. This suggests blue light affects chlorophyll production."
Illustration, after: "Under blue light, the plants outgrew the control group within a week. Their leaves darkened too. That second result surprised me more than the first, because it hints that blue light changes chlorophyll production rather than just speeding growth."
4. Choosing the safest, most common words
This is the habit with the strongest research behind it. In a 2023 study published in the journal Patterns, Stanford researchers led by Weixin Liang ran seven popular detectors on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth graders. The detectors were close to perfect on the eighth-grade essays but labelled the TOEFL essays as AI about 61% of the time on average. Eighteen of the 91 essays were flagged by all seven detectors, and 89 were flagged by at least one.
Then the researchers did something telling. They used ChatGPT to enrich the vocabulary of the TOEFL essays, and the average false positive rate fell to about 12%. They also used ChatGPT to simplify the word choices of the native speakers' essays, and those essays started getting flagged far more often. Same ideas, same authors, different vocabulary, different verdict.
Illustration, before: "The city has a big problem with traffic. Many people use cars. This makes the air bad and the roads very busy."
Illustration, after: "Traffic is the city's worst habit. Nearly everyone drives, so the roads jam by 8 a.m. and the air on Main Street smells of exhaust by noon."
The "before" here is perfectly honest writing from someone with a limited vocabulary. It deserves feedback on language, not a misconduct investigation.
5. Hedging and generalising instead of being specific
"Technology plays a crucial role in modern education." "There are various factors to consider." "This has both advantages and disadvantages." Generic claims are what you write when you're unsure, padding toward a word count, or writing about something you only know second-hand. They are also the default register of a chatbot asked about a topic it has no stake in.
Illustration, before: "Climate change has many impacts on agriculture, and farmers face various challenges as a result."
Illustration, after: "Wheat farmers in our county planted two weeks earlier this spring because the last frost came in March, not April."
Specific details, local names, numbers you checked and personal observations are hard for a model to produce by chance, so they tend to read as human. They also make for better writing, which is the real reason to use them.
6. Summarising conclusions that repeat the introduction
A conclusion that restates the thesis and lists the three points again is a classic school pattern, and a classic chatbot pattern.
Illustration, before: "In summary, remote work offers flexibility, cost savings and improved wellbeing, making it a beneficial option for employees."
Illustration, after: "Remote work won't suit everyone. But for my sister, getting an hour of sleep back each day mattered more than any office perk."
7. Writing in a formal, conventional genre
Abstracts, legal summaries, policy documents, product descriptions and technical documentation are supposed to be predictable. Readers want standard terms in standard order. A clear methods section in a scientific paper will never look "bursty", and it shouldn't.
If you write in these genres, expect higher scores and don't try to make a methods section quirky. Note the genre when you respond to a flag.
8. Very short pieces
Short text gives a detector little to work with. OpenAI said its classifier was very unreliable under 1,000 characters, and Turnitin says it needs at least 300 words of prose before it will produce an AI writing report. A discussion-board post or a 150-word reflection gives far less evidence than a 2,000-word essay, so treat any score on it with extra suspicion.
9. Writing next to AI-assisted passages
If part of a document came from AI with permission (an outline, a summary of a source) and the rest is yours, your own sentences can get swept in. Turnitin said in 2023 that its sentence-level false positive rate was around 4%, and that 54% of wrongly flagged human sentences sat right next to actual AI writing, with another 26% two sentences away.
That is a strong argument for marking clearly which parts came from AI, because a detector will blur the boundary.
Do grammar tools make writing look like AI?
It depends on what the tool does to your text.
Spelling, punctuation and basic grammar fixes usually don't. Turnitin says in its AI writing FAQs that its detector is not tuned to target corrections from Grammarly and similar tools, and that in its own tests those corrections mostly weren't flagged. That makes sense: fixing "their" to "there" barely changes how predictable your text is.
Generative features are different. Turnitin says text produced with Grammarly's generative features, such as drafting, paraphrasing and summarising, will likely be flagged. Tone rewrites, "improve this paragraph" buttons and full-sentence rephrasing all replace your choices with the model's choices. After a few rounds, the paragraph really does contain AI-written sentences, even though you started with your own.
Professional human editing can shift scores too. A 2026 preprint by researchers at Chung-Ang University and the editing company Wordvice compared 135,389 non-native manuscripts with their professionally edited versions across 13 detectors. False positive rates on the human-written texts ranged from 0% to 100% depending on the detector, and the same edits pushed AI scores up in some detectors and down in others, with bigger shifts for heavier editing.
A practical way to think about it:
What the tool did | Likely effect on a detector | How to describe it if asked |
|---|---|---|
Fixed spelling and punctuation | Little or none | "I used a spell checker." |
Suggested grammar fixes you accepted one by one | Usually small | "I used Grammarly for grammar suggestions." |
Rewrote sentences for tone or clarity | Can raise scores noticeably | "I used a rewrite feature on some sentences." Say which ones. |
Paraphrased, summarised or drafted text | Likely flagged, and it is AI text | Disclose it under your AI-use policy. |
If your school or workplace allows grammar tools but not generative rewriting, this is the line that matters. It is also the line that is hardest to remember a month later, so keep track as you go.
Why does my writing sound like AI?
Usually because you were trained to write well in a particular, conventional way, or because you're writing in a language you learned later. Some groups get flagged more often for reasons that have nothing to do with honesty.
Non-native English writers
The Stanford study above is the clearest evidence. Writers who learned English as an additional language often rely on a smaller set of words and sentence patterns, which lowers perplexity. Liang's team concluded that the design of many detectors penalises writers with limited linguistic expression and cautioned against using them in education.
Neurodivergent writers
A 2025 article in the journal Teaching in Higher Education, on neurodiversity and academic integrity, noted that detection software can produce false positives that fall disproportionately on neurodivergent students. It pointed to a 2024 study by Gegg-Harrison and Quarterman, while stressing that the study was small and needs follow-up.
Academic, technical and professional writers
Researchers, lawyers, engineers and people writing corporate copy all work within conventions that reward consistency. Their prose often sits close to the patterns detectors associate with AI because both are aiming for clear, standard, unsurprising language.
People who were taught by formula
If you learned to write from rubrics that reward topic sentences, transition words and recap conclusions, you were taught a style that sits close to chatbot output. That is not your fault. The rubric and the chatbot learned from the same body of tidy, conventional writing.
Why your score can change from one detector to the next
Detectors don't agree with each other, and sometimes they don't agree with themselves. The Chung-Ang and Wordvice study found detectors pulling in opposite directions on the same edits, which suggests they rely on different signals rather than a shared definition of AI text.
Thresholds also differ. Turnitin, for instance, says its under-1% document-level false positive rate applies to documents where 20% or more of the text is flagged, and it no longer shows a specific percentage between 1% and 19% because low scores are where false positives cluster. A free online checker might show you "38% AI" with no such caveat. Two numbers from two tools are not two pieces of evidence; they are two guesses using different rules.
Should you change how you write to avoid being flagged?
My view: no, not as a goal. Write clearly, be specific, vary your sentences when it helps the reader, and drop filler. Those are good habits on their own merits, and as a side effect they make your writing harder to confuse with a chatbot. But contorting your style to chase a lower score is a bad trade, for four reasons.
It doesn't work reliably. The same Stanford study showed that one simple rewriting prompt made ChatGPT-written essays slip past the detectors most of the time. If a single instruction can move the score that far, the score is measuring style, not truthfulness. Chasing it means chasing a moving target.
It can make your writing worse. Adding odd words for "perplexity" or random short sentences for "burstiness" produces prose that is less clear. A methods section or a cover letter is not improved by being made unpredictable.
It points you toward the wrong tools. "Humaniser" tools rewrite your text with AI to dodge AI detectors. If you wrote the text yourself, running it through one means submitting AI-rewritten text, which may break the very policy you were worried about.
It doesn't answer the real question. A teacher or editor who suspects AI use wants to know how the text was made. A lower score from a different tool doesn't tell them that. Evidence of your process does.
There is one style change I'd recommend regardless of detectors: put more of yourself into the work. Name the case study you read, the conversation that changed your mind, the figure you checked. That is what readers value, and no style tweak replaces it.
What should you do if an AI detector flagged your essay?
If your thought right now is "an AI detector flagged my essay and I wrote every word", stay calm and get practical. A flag is a prompt for a conversation, and Turnitin itself says its scores should start a conversation rather than settle a decision.
Ask what tool was used and what it said. Get the score, the tool name and which passages were highlighted. A 25% score on a 300-word reflection means something different from 95% on a long essay.
Check the highlighted passages against this list. Are they template sections, stock transitions, definitions, quotations or passages you ran through a rewrite tool? Being able to explain why a passage reads as uniform is useful.
Gather your process evidence. Version history in Google Docs or Word, earlier drafts, notes, outlines, browser research history, emails to a tutor. Anything with timestamps helps.
Be precise about what tools you used. If you used a spell checker, say so. If you used a rewrite feature on two sentences, say which two. Precision builds trust; vague denials don't.
Offer to talk through the work. Explaining your argument, your sources and why you made certain choices is strong evidence that you understand and own the writing.
Point to the published limits of detectors. Vendors themselves describe false positive rates, minimum lengths and caveats, and some universities, including Vanderbilt in 2023 and Curtin University in Australia from January 2026, have switched Turnitin's AI detection off.
Use your institution's process. If the conversation doesn't resolve it, follow the formal appeal route and ask that the detector score not be treated as the only evidence.
Why is a record of your process sturdier than style tweaks?
Every point in this post comes back to one fact: a detector only sees the finished text, so it can only guess. The sturdier answer is a record made while you write that shows where each passage came from.
That is the idea behind Assist Writer, an editor that records which text came from AI. It is not an AI detector and doesn't guess authorship from style. Text you paste with Ctrl+Shift+V (Cmd+Shift+V on Mac) is marked as AI text and shown in colour, while normal typing and normal paste count as your own writing. As you rewrite an AI passage, its colour gradually shifts toward normal text, so heavy rewriting is visible, and a stats bar shows the human vs AI split of the whole document.
That answers the question a detector can't: not "does this look like AI?" but "which parts came from AI, and how much did the writer change them?" If your teacher allows AI for an outline or a summary, you can show exactly where that help ended and your own writing began, which also protects your human sentences from being lumped in with the AI ones.
The quick guide to the AI paste shortcut and editor modes covers the shortcuts, the right-click "Paste as AI Text" and "Paste as Human Text" options, and the Markdown and rich-text modes. If you'd rather just try it, you can open the editor without signing up and write your next draft there. Documents stay in your browser in demo mode, and a free account keeps up to five in the cloud.
Write the way you'd write anyway. If someone questions it, a record of how the text got onto the page will hold up better than any score.
Frequently asked questions
Can an AI detector prove that I used ChatGPT?
No. An AI detector only sees the finished text and estimates how closely it resembles examples of machine writing. It cannot see how the text was produced. OpenAI withdrew its own classifier in July 2023 for low accuracy, and Turnitin publishes false positive rates and says its scores should start a conversation, not settle one. Drafts, version history and your ability to explain the work are far better evidence.
Are non-native English speakers more likely to be flagged by AI detectors?
Yes, according to the best-known research. A 2023 Stanford study in the journal Patterns found seven detectors labelled TOEFL essays by non-native speakers as AI about 61% of the time on average, while scoring US eighth-grade essays almost perfectly. Simpler, more common vocabulary makes text more predictable, and predictability is what many detectors read as machine writing.
What are perplexity and burstiness in AI detection?
Perplexity measures how surprised a language model is by each word in a text; low perplexity means the words were easy to predict. Burstiness measures how much that predictability varies across a document. Early detectors treated low, even scores as signs of AI. GPTZero says it moved to a deep-learning classifier in autumn 2023, so these are now at most one signal among several.
Will spell check or Grammarly get my essay flagged as AI?
Basic spelling, punctuation and grammar corrections usually won't. Turnitin says its detector isn't tuned to target those corrections and that its tests mostly didn't flag them. Generative features are different: Turnitin says text drafted, paraphrased or summarised by Grammarly's AI features will likely be flagged. If you used a rewrite feature, note which sentences so you can disclose it accurately.
Is Assist Writer an AI detector?
No. Assist Writer does not guess authorship from writing style. It records how text entered the document: text pasted with Ctrl+Shift+V (Cmd+Shift+V on Mac) is marked as AI and shown in colour, while typing and normal paste count as your own writing. As you edit AI text its colour shifts toward normal text, and a stats bar shows the human vs AI split.