AI Text Detector

Paste a passage to see whether it reads like ChatGPT, Claude, Gemini or another AI wrote it. Tuned to avoid false accusations, and honest about what it can't tell.

At least 80 words. Longer passages give steadier results.

How it decides

The free check runs an open-source classifier (desklib v1.01, MIT licensed) on our own servers. It reads the text in chunks of about 500 words and scores how much each reads like a language model wrote it. We only say likely AI-written when the score is high enough that, in our tests, no human writing reached it.

Signed-in users can add a second opinion from Pangram, which also marks which passages look AI-written or AI-assisted. Pangram is the stronger of the two, so its read leads; when they disagree the verdict says so, and if only our classifier sees AI the result is inconclusive.

Where it struggles

  • Text from the newest models, which the free classifier often scores as human.
  • AI text that has been paraphrased, heavily edited, or run through a “humanizer”.
  • Human text polished with AI tools such as grammar rewriters.
  • Short passages. Under about 150 words every detector is shaky.

Reading something online? The Chrome extension lets you select text on any page, right-click, and check it here. It does the same for images; see the AI image detector.

Frequently asked questions

How accurate is it?
On our labeled test set of 139 human-written and 146 AI-written passages, the free check called none of the human texts AI and caught 38% of the AI ones. It is deliberately strict: we would rather miss AI text than wrongly accuse a person. It catches older and default-style output (GPT-4.1, Claude Haiku 4.5) almost every time and misses much of the text from the newest models. The threshold was set on that same test set, so expect a few false alarms in the wild.
Can it prove someone used AI?
No. No text detector can. A high score is evidence worth a conversation, not proof. Please don't use it alone to accuse a student, a colleague or anyone else.
Does it flag non-native English writers?
Many detectors do: simpler vocabulary and textbook grammar look 'AI-like' to them. We tested on 64 texts by non-native writers (English-learner essays and papers by non-Anglophone scientists) and the free check called none of them AI. Two of the highest-ranked open detectors we tried flagged most English-learner essays, which is why we didn't use them.
What about Google SynthID and other watermarks?
Google, OpenAI and Anthropic watermark some of their models' text, but none of them offers a public way to check it: the detectors are private or limited to approved partners. Until that changes, every AI text checker, this one included, is making a statistical judgement.
What is the second opinion?
Signed-in users can also run a passage past Pangram, a commercial detector that independent studies rate well ahead of open models on text from current models. It costs us money per word, so free accounts get a few a month.
Do you keep my text?
No. The text is sent to our server to be checked and then discarded. We keep a fingerprint (a hash) of it and the scores, so checking the same text again is instant. A second opinion also sends the text to Pangram.