They are asked whether a human wrote it. They cannot tell, they are confidently wrong at rates that ruin people, and they are worst on exactly the writing that matters most.
overall accuracy for the two leading detectors - Originality and Turnitin - across 192 texts of student, professional, AI and hybrid authorship. Int. Journal for Educational Integrity, 192-text evaluation, 2026
recall on HYBRID human-and-AI text, which is how most real documents are now written. Originality caught 2 in 100; Turnitin caught 31. Int. Journal for Educational Integrity, 192-text evaluation, 2026
of TOEFL essays by non-native English writers were misclassified as AI-generated, while the same detectors were near-perfect on native-writer essays. Liang et al., Patterns (Cell Press), 2023
both move detector accuracy on their own - it falls as texts get longer, and is lower on scientific writing than on humanities writing. Int. Journal for Educational Integrity, 192-text evaluation, 2026
OpenAI withdrew its own classifier for low accuracy. The honest conclusion is not that detection needs a better model - it is that authorship is not recoverable from the text, and a tool that claims otherwise is selling a verdict it cannot support.
Six that close the biggest gaps. 1 in the full content integrity library.
The gap it closes. Answers a different question, and one that has an answer. It measures properties the text actually has - entropy, compressibility, lexical diversity, burstiness, claim checkability - and emits a content hash, so a document can be verified as unchanged and characterised on stated measures. It does not guess who wrote it, because that is not knowable from the text.
See it run in your browser →This is deliberately NOT an AI detector and must not be used as one. It reports measurable structural properties of a document and a tamper-evident hash. It makes no claim about authorship, and no score it produces should be used to accuse anyone of anything. The studies above are the reason that boundary exists.
Every tool above runs entirely in this browser tab. Nothing is uploaded, nothing is sent to a model, and each result cites the rule it applied.
Free tier, no card, no install. An account saves your work and lets you export a citable record of a run. See all 1 content integrity tools.