AI writing recognised
90.7%
68 of 75 texts
95% range 82% to 95.4%
All machine writing in the test: answers written by a chat model, continuations of a human opening and paraphrases.
Measured accuracy
AI detector accuracy by writing type
AI detector accuracy varies by language and by kind of writing. Every figure below was measured on held-out text the detector was never fitted on, at the line the result card uses, with its sample count and its range. Explore AI recognition and false flags for each language below.
- 13 languages with separate calibration
- A score, a confidence rating and sentence marks
- Measurements repeated when the model or threshold changes
Language results
Compare the supported languages
How much AI writing each language's held-out test recognised and how much human writing it flagged, counting a text at a score of 65 or more.
English
AI recognised in the test90.7%Native writers flagged0%Non-native writers flagged0.5%Measured on held-out text, 2026-10-12Spanish
AI recognised in the test82.9%Native writers flagged0.7%Non-native writers flagged0.7%Measured on held-out text, 2026-10-12French
AI recognised in the test68.8%Native writers flagged0%Non-native writers flagged1%Measured on held-out text, 2026-10-12German
AI recognised in the test90.7%Native writers flagged0%Non-native writers flagged0%Measured on held-out text, 2026-10-12Italian
AI recognised in the test94.7%Native writers flagged0%Non-native writers flagged0.7%Measured on held-out text, 2026-10-12Portuguese
AI recognised in the test96%Native writers flagged0.7%Measured on held-out text, 2026-10-12Dutch
AI recognised in the test97.3%Native writers flagged1.3%Measured on held-out text, 2026-10-12Polish
AI recognised in the test91.9%Native writers flagged0%Measured on held-out text, 2026-10-12Turkish
AI recognised in the test89%Native writers flagged0%Measured on held-out text, 2026-10-12Russian
AI recognised in the test90.7%Native writers flagged0.7%Measured on held-out text, 2026-10-12Arabic
AI recognised in the test95.9%Native writers flagged0.7%Measured on held-out text, 2026-10-12Hindi
AI recognised in the test94.5%Native writers flagged0.7%Measured on held-out text, 2026-10-12Indonesian
AI recognised in the test98.6%Native writers flagged0%Measured on held-out text, 2026-10-12
Choose a language to see how much AI writing the held-out test recognised and how much human writing it flagged, for native and non-native writers separately.
Results by language
AI detector accuracy from both sides
English test results
AI chat answers recognised
100%
54 of 54 texts
95% range 93.4% to 100%
Texts written whole by a chat model.
Native writers wrongly flagged
0%
0 of 150 texts
95% range 0% to 2.5%
A lower rate means fewer human texts were flagged in the test. Check the notes for the writers and writing types included.
Non-native writers wrongly flagged
0.5%
3 of 600 texts
95% range 0.2% to 1.5%
Essays and answers by people writing in a language they learned.
These rates come from held-out texts, none of which the detector was fitted on. A text counts as flagged at a score of 65 or more, the line the result card uses. The rates describe the test, not the certainty of an individual score.
Measured on held-out text, 2026-10-12Text length
Give the detector enough context
Longer passages provide more evidence about writing patterns. Read the confidence rating alongside the score, since language and writing type also affect the result. What should you paste? Choose a continuous passage from the document you want to review. Keep its wording intact for the first check, then use the sentence marks to decide which passages deserve closer attention.
Under 50 words, the detector gives no reading.
Below 100 wordsLow confidence, with a short-text note
100 words and aboveMore context for the reading
Compare the shorter-text tests
English test results
Shorter-text measurements for this language are pending.
How it works
From your text to a useful reading
The detector reads how predictable the writing is, sentence by sentence. It uses patterns learned from human and machine writing, then applies the calibration curve fitted for the text's language. It assesses writing patterns rather than searching for banned words, a watermark or a match in a database.
- 01
Add your text
Paste your writing or upload a file in a language listed on this page.
- 02
Read each sentence in context
Each sentence is assessed alongside the surrounding passage, so the reading reflects more than isolated phrases.
- 03
Assess writing patterns
Trained models score how closely each passage follows patterns associated with human or machine writing.
- 04
Apply the language calibration
The score is interpreted through a curve fitted for the text's language, including writing by non-native speakers.
- 05
Review the score and evidence
You get a figure, its range, a High, Medium or Low confidence rating, and sentence marks showing which passages point toward AI or human writing.
Building and testing
Measured with human writers in mind
You get more useful evidence when false flags are measured alongside AI recognition. We set the decision line with human writing in mind, then publish dated results and the context needed to read them. These are Blurwrite's own measurements. The research links and external reference check provide separate context, so you can distinguish our tests from outside evidence.
- 01
Learn from varied writing
The models learned from many corpora and sources, covering varied human writing and machine writing from current models.
- 02
Calibrate each language
Each language has its own fitted curve, using texts in that language and including writing by non-native speakers.
- 03
Test beyond calibration
Held-out tests cover writing excluded from calibration, language learners, older model generations, other model families and machine continuations of human openings.
- 04
Apply a publication rule
Our publication rule sets a ceiling of one false flag in twenty on the kinds of writing we publish figures for, with non-native writing measured separately. The held-out results below show how current readings compare with that rule.
How the decision line works
Human and machine writing can produce overlapping scores. The decision line balances AI recognition with protection against false flags. Writing near that line needs closer attention to the confidence rating and sentence evidence.
Measuring accuracy for non-native writers
Research has found that AI detectors can flag non-native English essays more often than native English essays. We measure non-native writing separately so a language-wide average does not hide those mistakes.
AI detector accuracy: common questions
Is 20% AI detection bad?
A 20% score is a reading of the text, not proof that you used AI or that a fixed share of your words was generated. Review the confidence rating and the marked passages in context. What is the 30% rule for AI? Treat any percentage cutoff as a rule to verify with the organisation requesting your work. Ask whether it refers to an AI score, a similarity score or an allowed use policy. Those describe different things. A tool's decision threshold also serves a different purpose from an assignment's requirements. Check your text with the Blurwrite AI detector to see the evidence behind its score.
Can an AI detector be wrong?
Yes. Human writing can be flagged, and the rates vary by language and writing type. A 50-word excerpt gives less context than a longer passage; compare a sample of 50 words with one of 100 words or more. Why was my own writing flagged? Human and machine writing can share patterns, so an authentic passage can cross the decision line. Start with the exact text that was checked and inspect the marked sentences. If you are discussing a disputed result, record the tool, date, language, excerpt and confidence rating. That keeps the review focused on the actual reading rather than assumptions about your writing process. Check your text with Blurwrite and read the confidence rating alongside the score.
How do I prove I did not use AI?
Drafts, version history, notes and source records can help demonstrate your writing process. Keep that evidence and explain how you developed the text; a detector score is supporting context rather than proof of authorship. What should I share during a review? Build a clear timeline from your outline to your submitted draft. Identify where you added sources, changed an argument or revised a passage. Preserve the original files and timestamps where available. Ask the reviewer which passages raised concern and what evidence their process accepts. If you used permitted writing assistance, explain its role and follow the relevant disclosure policy. Check your text with Blurwrite to review passages that may need an explanation.
How good is the AI detection used by my institution?
Judge it by published tests for the relevant version, language and writing type. The measurements on this page describe Blurwrite, not another tool's performance. How should I compare AI detectors? Look for dated recognition rates, false flags, sample sizes, writing types and held-out tests on comparable material. Consider how tools handle rewritten passages and whether measurements are refreshed as AI writing changes. For an essay, prioritise essay results over a broad average. Ask whether the reported threshold matches the one used in practice and whether non-native writers were tested separately. Agreement between tools is useful context, but shared mistakes can also produce agreement. Check your text with the Blurwrite AI detector and use the score-reading guide on this page to interpret the result.
Read the results with their context
What the score and rating mean
The score describes how the text reads, while the High, Medium or Low rating describes the strength of the evidence. Use the sentence marks to identify passages worth reviewing rather than treating the score as proof of authorship. How do the test rates differ? AI recognition counts machine-written texts correctly flagged. The false-flag rate counts human-written texts wrongly flagged. Their denominators describe different groups, so the rates are not opposites and should not be added together. A useful accuracy comparison reports both, with the sample counts, writing types and decision threshold. An individual result is a separate reading of your text.