What an AI detector actually measures
An AI detector does not recognise an author. It measures the surface of a passage, how vocabulary, sentence rhythm, punctuation and structure are distributed across it, and compares that shape with what a trained classifier learned from large collections of human and machine writing. What comes back here is one number for the whole passage, an AI-writing signal score from 1 to 99, and the zone that number falls into: Likely human-written up to 35, Uncertain from 35 to 64, Likely AI-written from 65.
A high score says the writing looks statistically like the machine-written text the system was trained on. It does not say who typed it. Two passages can score alike for opposite reasons: one because a language model produced it, another because a person wrote to a taught template in a narrow register. Nothing in the measurement separates those two cases. That is why every result here arrives with a zone, a score stability range and one sentence of evidence for that length of text from our sealed test, instead of as a number on its own.
It also follows that editing moves the reading. Rewrite an opening, split long sentences, or put a draft through another round of revision, and the score shifts, because the surface has shifted. The measurement describes the text in front of it. It is not a record of how that text came to exist, and it cannot tell a careful writer from a careful model when both produce the same shape of prose.
The engine behind the score is V4, released 17 October 2026: a fine-tuned transformer text classifier that reads a text in overlapping windows of 512 tokens, averages the window scores, and compares the result with a threshold frozen for the text’s length. The displayed score is a fixed map of that internal value, built so that 65 means “at the threshold” at every length. Scores of 65 or more are treated as AI-written, at every text length; a chip on the card says whether the signal is moderate (65-84) or strong (85-99).
One sealed test, opened once
Every accuracy figure on this site comes from a single sealed test, scored once on 17 October 2026 by the frozen engine, with the thresholds, the score map and the data rules written down and hash-stamped beforehand. The sealed human set holds 4,000 texts from 8 source families the model never trained on, all dated 2021 or earlier, before writing assistants were in common use. The sealed AI set holds 6,113 texts from 9 model families, plus two open-weights families the engine never trained on at all. The previous engine, V3, was scored on the same texts so the two can be compared honestly.
- 0.14% of human texts of 150 words or more reached 65 or more: 4 of 2,863, about 1 in 700. V3 on the same texts: 0.80%.
- 0.97% of human texts of 100 to 149 words: 11 of 1,136, about 1 in 100. V3: 5.81%.
- 98.97% of AI texts of 150 words or more from the tested current models were caught, about 99 in 100. V3 caught 30.46%.
- 98.50% of texts from gpt-oss, a model family the engine never trained on, were caught (998 texts); Gemma, the second unseen family, 98.55% of a small sample of 69.
The places where the sealed result is worse than the development result, and the one kind of writing where V4 is worse than V3 (modern formal speeches, 1.83% flagged against 0.97%), are published with everything else on the evidence page, with raw counts and Wilson 95% intervals for every family and every length band.
What changed with V4
Three things changed at once on 17 October 2026, and it is worth knowing which is which. The engine changed: V3 was a fusion of classical text models, and V4 is a single fine-tuned transformer classifier, trained once and frozen. The scale changed: the old display ran from 0 to 100 with a different cut point for every length, so the same number meant different things on different texts; the new scale runs from 1 to 99 with the AI line at 65 everywhere. And the card changed: instead of a false-flag sentence alone, every result now carries the zone, the strength chip, the fixed sentence about the AI line, the evidence sentence for that length, and the disclaimer.
What did not change: the rules. The thresholds were frozen before the sealed test, the sealed sets were scored once, and nothing was tuned afterwards. If V4 is ever unavailable for a moment, the card says so and asks you to try again in a few seconds; no other engine answers in its place. Every response carries the release id and a request id.
One honest caveat belongs here rather than in a footnote. V4 is better than V3 on every length band and every model family we measured, with one exception: modern formal speeches, where it flags 1.83% of texts against V3’s 0.97%. If you check a parliamentary speech, a central-bank address or a formal statement written to a template, read a flag on it with that in mind. The number is published on the evidence page with its interval, next to everything else.
How to check a text, step by step
- Paste at least 100 words of English, or upload a PDF, DOCX, TXT or RTF file. The counter under the box tells you when you have enough; below 100 words the tool declines rather than guessing.
- Prefer 150 words or more. Short passages are the least reliable: the card says so on every short result and asks you to add more text if you can.
- Press Check text. Most checks finish in well under a second. If many people are checking at the same moment the tool says so and asks you to try again in a few seconds; it never queues your text indefinitely.
- Read the zone first, then the score, then the evidence sentence. The sentence tells you how often human texts of that length reached 65 in the sealed test, and how often AI texts did. It describes the length, not your text.
- Compare up to 3 passages from the same document if you want to see whether sections differ. Each passage is scored on its own; nothing is averaged.
- Save the card as a PDF if you need a record. The report carries the release id and the request id, so the exact engine that produced it can always be identified.
Who this is for
Writers checking their own drafts before sending them anywhere a detector might be used, and keeping the result as a record. Editors and content teams deciding which submissions deserve a closer read, with the writer’s consent. Teachers who want one more input before a conversation, never a verdict: a result is not enough on its own for a decision about a student. Researchers who need a published, reproducible measurement rather than a marketing number. It is not for anyone who needs proof, because no detector can provide it. It is also not for text in languages other than English, for passages under 100 words, or for decisions about a person: the terms of use exclude academic, employment, legal and similar decisions, and the card repeats on every result that a result alone does not prove who wrote a text.
Reading a result responsibly
Length is the one thing you control that changes the answer most. Short human writing is easy to misread, so short passages misfire more often: about 1 in 100 human texts at 100 to 149 words reached 65 or more in the sealed test, against about 1 in 700 at 150 words and above. At 600 or more words, none of the 970 human texts in the sealed test did.
A Likely AI-written result means the passage sits above a threshold that was frozen before the test, for its length. It is a reason to read the text and its history, not a finding about a person. A Likely human-written result means no strong signal was found; about 1 in 100 AI texts of 150 words or more also landed there in the sealed test, so it is not a clearance either. The Uncertain zone, 35-64, is where edited, formal, templated and second-language English lands more often; treat it as no useful answer rather than as a weak accusation.
Whatever the zone, the card ends with the same sentence: the score is not the share of the text written by AI, and it is not a probability. A result alone does not prove who wrote a text. If you are reviewing someone else’s writing, ask for the drafts, the notes and the sources; those are evidence, and a score is not.
Where this detector fails
- Short texts. The 100 to 149-word band is the weakest, at about 1 in 100 human texts wrongly flagged, and it is also the band with the largest gap between the development result and the sealed result (recall 100-149: -1.89 pp).
- Modern formal speeches. Parliamentary debate, the Congressional Record and central-bank speeches from 2024 to 2025 were flagged 1.83% of the time (17 of 931), the one register where V4 is worse than V3.
- Non-native English. 1.34% of learner texts in the sealed test were flagged (22 of 1,641), and a diagnostic on further learner corpora, outside the sealed test, measured 1.80% at 150 words or more. No Arabic-first-language data has been measured.
- Not yet measured for V4: mixed human and AI text, older or small models, and heavily edited or paraphrased machine text. Expect lower scores on edited machine text than the sealed figures suggest.
- Languages other than English are not supported. The engine accepts them but nothing about those results has been validated.
AI detector or plagiarism checker?
They answer different questions. A plagiarism checker looks for text that already exists somewhere else and can show you the match. An AI detector looks for the statistical shape of machine writing and can show you nothing but a score, because there is no original to point at. A passage can be entirely original and still score high; a passage can be lifted word for word from a human source and score low. If your question is “was this copied?”, use a plagiarism checker. If your question is “does this read like machine writing?”, this tool gives you a measured answer with its error rate attached. Neither answers “who wrote this?”.
Guides on this site
Understanding results explains the card word for word, the four zones and the evidence sentence. If your text was flagged is for writers whose own work drew a signal, with the short-text, non-native and formal-speech limits spelled out. The model guides, 10 of them, describe how often texts known to come from each model family were caught in the sealed test, with the exact versions tested. Methodology and evidence carry the frozen numbers and the full sealed test, and language support explains why the detector is English only.