Methodology

The short version

Every detector on this site estimates how closely a piece of work resembles the output of a language model. None of them detects authorship, and none of them has access to a record of what wrote your text. A score is a similarity measurement, and similarity is circumstantial. That is why every result here ships with the passages, blocks or regions behind it: the evidence is the part you can check, and the number is the part you cannot.

What the text detector measures

Text is split into passages and each one is scored on its own by an evaluation model, so a single polished paragraph cannot drag a whole document upward and a single rough one cannot rescue it. Passages shorter than a sentence or two lean on the document-level result, because a handful of words carries almost no signal. The features that move a score are concrete, and they are listed on each tool page:

• Formulaic openers and connectives — the transitions generators reach for again and again.
• Low variance in sentence length and paragraph shape — human writing is rhythmically uneven; generated writing often is not.
• Balanced three-part structures used as a default rather than for emphasis.
• Hedged non-claims — sentences that sound careful but commit to nothing checkable.
• Absence of specific, verifiable detail where an author with direct knowledge would name something.
• Summarising closing lines that restate a paragraph without adding to it.

Note what is not on that list: an impersonal tone on its own, a formal register on its own, or the absence of a first-person voice. Those describe most technical, legal and academic writing ever produced by a human being.

What the image detector measures

The image scan runs entirely in your browser. A vision model scores the whole image and five regions, which is what produces the heatmap, and the file itself is read for provenance: C2PA Content Credentials, EXIF camera fields, and the markers some generators write into their own output. Provenance is the strongest evidence available here, because it is a record rather than an inference — but it is also trivially removed. A screenshot, a re-encode or an ordinary social upload strips it, so its absence means nothing at all.

Known error characteristics

These are the failure modes we can name. They are not rare edge cases.

• Writing by non-native English speakers is over-flagged. This is the best documented failure in the field, and the one that does the most damage, because the people it affects are the least able to contest an accusation.
• Short passages are unreliable. Below roughly a paragraph there is not enough text to measure variance in, and the score should be read as close to uninformative.
• Editing defeats detection. Generated text that has been paraphrased, or run through a rewriting tool, loses most of the signal this relies on. A low score is not evidence that a human wrote something.
• Formal and templated genres score high. Lab reports, legal summaries, policy documents and technical documentation are impersonal, evenly structured and heavily conventional by design.
• Code is the hardest case of all. Idiomatic code converges. Boilerplate, generated scaffolding and the house style of a popular framework all look alike no matter who typed them.
• Scores are not stable across models or over time. As models are updated their output changes, so a threshold that looked right last year does not stay right.

How a score should be used

As a prompt to look closer, and nothing beyond that. Read the flagged passages and ask whether the specific thing the detector objected to is actually a problem. If you are assessing someone else’s work, a score is a reason to open a conversation, never a finding to present. Several institutions have reached the same conclusion and switched their detectors off rather than let a percentage stand in for judgement. Our disclaimer sets out the limits in full, and about explains why the site is built this way.

Sources

The claims above are not ours alone. The research they rest on:

• Liang, Yuksekgonul, Mao, Wu & Zou (2023). GPT detectors are biased against non-native English writers, Patterns. Detectors classified a majority of TOEFL essays written by non-native speakers as AI-generated, while classifying comparable writing by native speakers correctly.
• Sadasivan, Kumar, Balasubramanian, Wang & Feizi (2023). Can AI-Generated Text be Reliably Detected? On the theoretical limits of detection, and how far paraphrasing degrades it.
• Weber-Wulff et al. (2023). Testing of detection tools for AI-generated text, International Journal for Educational Integrity. An independent evaluation of fourteen tools, none of which performed well enough for the use they are marketed for.
• Vanderbilt University (2023). Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector. One institution’s reasoning for switching detection off entirely.
• C2PA — the Content Credentials specification the image scan reads.

Try the AI text detector