How it works

1. Quick checks first

Before anything is scored, we check that the text has at least 80 words, that it is mostly prose rather than code, lists or tables, and that it is in English or Spanish. If not, you get a clear message instead of a guess.

2. The Engine reads the text

Our Engine is a language model trained to tell human writing from AI writing. Long texts are read in overlapping slices, and their scores are combined into one score for the whole text. Every result names the Engine that produced it, so it can still be explained after we improve the Engine.

3. A Verdict in three Bands

• Likely AI: the score is higher than all but about 1 in 100 human texts in our tests.
• Likely human: the score is lower than all but about 1 in 100 AI texts in our tests.
• Unclear: the score sits between those two lines. When parts of an Unclear text look AI-written, we say so.

The thresholds are set separately for English and Spanish. The AI probability is shown underneath, but it is the Band that carries the meaning. See Accuracy for the measured error rates.

4. The evidence

For the parts that did not read clearly human, we score runs of whole sentences (at least 80 words each) and highlight the ones that look AI-written. Hover over a highlight to see its confidence.

What happens to your text

We score it and throw it away. Only the result is kept: the scores, where the highlights fall, the word count and the language. Never the text itself. See Privacy.

What a Verdict does not mean

• It is not proof of who wrote a text. A careful human can write like a model, and a model can be edited until it reads like a person.
• It should never be the only reason for a grade, a sanction or an accusation.
• We cannot predict what Turnitin or any other detector will say. They use different models and thresholds.