ChatGPT detector: what it can see, and what it cannot

There is no signature inside the text to find

A model emits ordinary words in ordinary characters, and nothing is stamped on them. Two identical paragraphs, one typed and one generated, are the same file byte for byte, and no tool can separate them, because there is nothing left to separate.

Published research does describe watermarking: schemes that bias a model's word choices so the key holder can test a passage later. That is a property of the generator, not something an outside website can check, and the signal weakens as text is edited or retyped.

OpenAI's published guidance for educators says AI writing detectors do not reliably distinguish machine text from human text, and it retired the classifier it once offered for its own output, citing a low rate of accuracy. We are not affiliated with OpenAI, so quote its current documentation rather than a summary.

Detection is a style question, not an authorship questi

Two statistics sit under most of these classifiers. The first is average predictability: given the words so far, how expected was the next one. A model picks likely continuations, so its output runs smooth. The second is how much that predictability rises and falls, and human drafting spikes and dips.

Both are properties of the writing, so the honest reading of a high score is that the text reads the way generated text reads. Who produced it is a fact about an event outside the file, and the file did not record it.

That is why human writing gets flagged. Liang and colleagues at Stanford HAI reported in 2023 that 61% of TOEFL essays written by non-native English speakers were falsely flagged as AI across seven detectors. Every one had a human author. Plain, even, correctly punctuated prose organised to a template is what these scores respond to.

Why no detector can name the model

Ask whether a paragraph came from ChatGPT rather than another assistant and no classifier can answer. Chat models are tuned toward the same even, helpful register, so their output converges on the page. A detector scores that shared register, so the same reading comes back for any of them.

Mixed drafts are the harder case and the common one, and a paragraph you wrote and a model tightened lands in the middle. What a score supports is narrower than authorship and still useful: these sentences read this way. Here is what it does not carry.

• Which assistant produced the text, or which version of it.
• Whether the text was translated, dictated or run through a grammar checker, all of which flatten prose.
• Which sentences inside one document were generated and which were written.

What ChatGPT output looks like when you read it closely

Reading beats scoring when you have the time, because a sentence you can point at is worth more than a number you cannot explain. Each of these takes one edit.

• The negation pivot. "Remote work is not just a perk, but a structural change." Keep the true half, delete the setup.
• The rule of three. Three adjectives, three parallel clauses, three-item lists where two specifics would land harder.
• Announced significance. "This underscores the importance of early screening", "plays a crucial role". Each asserts that something matters instead of showing it.
• Stock openers and idle connectives. "In an increasingly competitive marketplace", "Additionally", "Furthermore". If the next sentence follows, it follows without an usher.
• Flat rhythm. Every sentence within a word or two of the same length, and headings all cast in the same grammatical form.
• Confidence with nothing checkable in it. No proper nouns, no figures, and citations that do not resolve when you search.

How to check a draft

Check the finished text rather than an outline. The floor is 65 words, and a run costs one credit whatever the length. Then put the same text through the registry of named AI writing tells, free and unlimited in your browser. It names the exact phrase and the reason it fired.

Nothing in that registry is proof on its own, since every pattern appears in human writing too. What separates a generated paragraph from a written one is density: several tells in a short span, on a flat rhythm.

If the draft needs rewriting, the humanizer rewrites every sentence in one pass, up to 12,000 characters or roughly 2,000 words per run, from a 25-word minimum, at one credit per 50 words. Then check the rewritten version here.

• Paste prose. Reference lists and long block quotations are not your sentences.
• Treat Borderline as unresolved rather than as a pass.
• When the checker and the registry point at the same paragraph, you have found the work.

If you are checking someone else's writing

A score is the start of a conversation, not a finding. Ask about process instead: what they read, what they cut, where the argument came from. A writer who did the work answers in specifics. Decide what your policy permits and what evidence you accept before the first case, not during it. Our page for teachers works through classroom process, and the false positives page covers who gets flagged and why.

If your own writing is the thing being checked

Keep the record while you work. The signed drafting record here captures your session as you write, including timestamps, how much you deleted and whether a large block arrived in one paste, then exports a file signed with ECDSA P-256 that anyone can verify without contacting us. It is free and unlimited.

If an accusation has already been made, the falsely accused page has the questions to ask, written as sentences you can send: ask to see the report rather than a screenshot of a number, and ask whether the score is the only evidence.

On permission, briefly. Rewriting a draft you were allowed to generate is ordinary editing. Where AI is not permitted, that rule covers the draft however it is edited. Read the policy that applies to you.