How our AI detector works, and how accurate it is
Are AI detectors accurate? Ours makes mistakes both ways, and this page shows how often, on texts it never saw in training, with counts, 95% ranges and the test date.
How the AI detector reads a text
How do AI detectors work? Ours asks of every sentence how predictable its wording is, given what came before. Machine text picks the expected word more often than people do. It runs on our own model, with no watermark search and no database match.
Paste or upload
Paste text or upload a file. The language is read from the text.
Read in context
Each sentence is scored with the passage around it.
Score the patterns
Our model measures how closely each passage follows the habits of human or machine writing.
Apply the language's line
The score is read against a line fitted for that language. Other languages are marked not calibrated.
Show the evidence
A score with a range, a confidence rating and sentence marks. High confidence reads "Reads as AI-written", medium reads "Likely", and "Mixed" means the marks point both ways.
AI detector accuracy, from both sides
Machine texts caught and human texts flagged, per language. A text counts as flagged at a score of 65 or more, the result card's line.
Pick a language. French is our weakest at catching machine text, and its row shows it.
English: held-out results
Held-out test, 2026-10-12All machine texts caught
68 of 75 texts90.7%95% range: 82% to 95.4%
Chat answers, continuations of a human opening and paraphrases. Continuations are the hardest to catch.
Chat answers caught
54 of 54 texts100%95% range: 93.4% to 100%
Texts a chat model wrote start to finish.
Native writers flagged by mistake
0 of 150 texts0%95% range: 0% to 2.5%
Everyday web writing, at most 2 per site, across 8 kinds of writing.
Non-native writers flagged by mistake
3 of 600 texts0.5%95% range: 0.2% to 1.5%
People writing in a language they learned, from learner collections.
None of these texts was used in training. A text counts at 65 or more. The rates describe the test, not one result.
Tested with human writers in mind
A detector that flags a student's own essay does real harm. So we count both kinds of error, set the line with human writers first and date every result. All figures here are our own.
Why some texts land in the middle
Human and machine scores overlap. Our line sits near the top of the human curve, so few human texts cross it and some machine text stays under. Near the line, read the sentence marks.
Train on varied writing
Many collections of human writing, and text from current chat models.
Calibrate each language
Each language gets its own line, with non-native writing kept as a separate group.
Test on unseen text
150 human texts per language, plus text from 9 chat models, continuations and paraphrases, all kept out of training.
Hold to a publication rule
A language is calibrated only if at most 1 human text in 20, and 1 non-native text in 20, crosses the line.
Short texts give it less to read
A short passage can swing on a few words. Paste one continuous passage as it was written, then check which sentences pulled the score.
Short-text results No short-text figure yet.
Common questions on accuracy
Is a 20% AI score bad?
Not on its own. A score describes how the text reads, not the share a machine wrote, and it comes with a range and a rating. If a school uses a cutoff, ask whether it means an AI score or a similarity score. Check the marked sentences in the Textlume AI detector.
Can an AI detector be wrong about my own writing?
Yes. Human text can cross the line, more often in some languages than others. Below 50 words there is no reading, and 50 words give far less to go on than 100. Paste a longer stretch into Textlume and compare.
How do I show I wrote it myself?
With your process, not a second score. Drafts, version history and notes show how the text grew. Textlume marks the sentences that read as machine-like, so you know which parts to explain.
How should I judge another AI detector's accuracy?
Look for dated figures with the number of texts behind them, tested on text the tool never trained on, with false flags next to the catch rate. For an essay, essay results beat a broad average. Our figures describe Textlume alone.
More on reading these figures
English by non-native writers
On held-out English by writers whose first language is Arabic or Spanish, the detector flagged 3 of 600 texts: 0.5%, 95% range 0.2% to 1.5%. For native writers it flagged 0 of 150.
What the score, range and rating mean
The score says how machine-like the text reads, the range how far it could move, and the rating how strong the evidence is. None of it proves who wrote a text. Caught and flagged rates come from different texts, so never add them up.