What an AI detection score actually means
The most common misreading
A detector returns 87%. Most people read that as 'about 87% of this was written by AI'. That is not what it means. The number is a single classifier output for the whole submission: how strongly the piece as a whole matches statistical patterns the model associates with AI-generated content. It says nothing about which parts, or what proportion. A document where every paragraph is mildly AI-like and a document with one heavily AI-like section can land on the same score.
What the model is actually measuring
Text detectors look at properties that distinguish generated prose from human prose in aggregate. The main ones are predictability — how well each word is anticipated by what came before — and variation. Human writing is bumpy. Sentence lengths vary, a sentence occasionally goes somewhere odd, word choice is sometimes specific in a way no averaging process would produce. Language models optimise for likely next words, which produces prose that is smoother and more uniform than most people write. Image and video detectors work on the same principle with different features: texture distribution, frequency-domain artefacts, edge and compression consistency. In every case the model is asking 'how typical is this of generated content', not 'was this generated'.
Why the same text scores differently on different tools
Each detector was trained on a different corpus, against different generation models, with a different threshold. There is no shared standard and no calibration between vendors. So a paragraph scoring 30% on one tool and 80% on another is not a contradiction or a bug — the two numbers are not on the same scale. Comparing across tools is close to meaningless. What is meaningful is whether several independent tools agree on the direction.
Why we show bands instead of exact numbers
A displayed 87% implies a precision the underlying model does not have. Re-run the same text with a paragraph removed and it might read 79%. Nothing about the authorship changed; the estimate moved. Bands acknowledge that. AI Catcher reports five: low likelihood, mostly human-like, mixed signals, likely AI-assisted, high likelihood. Each has fixed wording that does not escalate. A band is honest about resolution in a way a two-digit number is not — and it prevents the false precision that gets people into trouble when they quote a score in a disciplinary meeting.
- 10–19% — Low AI likelihood: mostly natural patterns
- 220–39% — Mostly human-like: a few signals, likelihood stays low
- 340–59% — Mixed signals: both natural and AI-like patterns present
- 460–79% — Likely AI-assisted: several sections match generated writing
- 580–100% — High AI likelihood: strongly matches AI-generated patterns
The reliability problem nobody advertises
Two failure modes matter, and they are not symmetric in consequence. False positives flag human work as AI. These are concentrated in specific populations: non-native English writers, whose vocabulary range is narrower and therefore more predictable; people writing in technical or formulaic registers; anyone using grammar-correction tools, which smooth exactly the variation detectors look for. A 2023 Stanford study found detectors flagged over half of TOEFL essays by non-native speakers as AI-generated, against near-zero for native-speaker essays. False negatives miss AI content, and are trivial to produce — light editing, a paraphrasing pass, or simply prompting for a less standard style will usually do it. So the tool is weakest exactly where it is most often used: deciding whether a specific person cheated.
What makes a score less trustworthy
Some inputs are genuinely harder to assess, and a good detector will say so rather than returning a confident number anyway. Short text is the biggest factor — under a few hundred words there is not enough signal, and scores swing wildly. Translated text carries the structure of the source language and confuses the model. Heavily edited text is a mixture the classifier was not trained on. For images, compression, screenshots and filters all degrade the signal; for video, platform re-encoding destroys most of it. If a tool returns the same confident number whether you feed it 80 words or 8,000, that tells you something about the tool.
How to actually use a result
Use it as one input, weighted by the stakes. For low-stakes questions — is this listing copy generated, does this post read like a bot — a high score is enough to act on, and being wrong costs nothing. For anything involving a person's standing, a score is a reason to start a conversation, never a reason to conclude one. Ask for drafts, version history, or a discussion of the material. Those establish process, which is what you actually care about; the score only ever gestured at it. And if you are on the receiving end of a false positive, the same logic works in your favour: document your process, because process is the evidence a score can never be.
Weiterlesen
Your work was flagged as AI and it wasn't. What now?
False positives are common, and they hit some writers far more than others. The defence that works is process evidence — and most of it has to exist before you need it.
How to tell if an image is AI-generated
The obvious tells — six fingers, garbled text — are mostly gone. Here is what still works, what does not, and how to combine checks instead of trusting any one of them.