What we will test
- Human writing across genres: student essays, business writing, legal and technical prose, creative writing, and writing by non-native English speakers.
- AI writing from several current models, unedited.
- AI writing that has been lightly edited by a person.
- AI writing run through a paraphrasing tool.
- Short (under 150 words), medium, and long samples.
What we will measure
- False positive rate: how often human writing is flagged. We report this first, because it is the error that harms people.
- Detection rate: how often AI writing is caught, for each kind of AI text.
- Fairness: false positive rates for non-native English writing compared with other human writing.
- Consistency: whether the same text gets the same result on repeat checks.
How we will publish
- The date of each test and the detector settings used.
- Results by category, not just one overall accuracy number.
- A description of the test set, and the human samples’ sources with permission.
- Reruns on a regular schedule, because detectors and AI models change.
Until then
The best independent study we know of is a 2025 University of Chicago working paper comparing Pangram, Originality.ai, and GPTZero. Read our summary in How accurate are AI detectors? and our detector reviews.