Pangram and false positives
A detector with a strong reputation for accuracy, used by publishers and increasingly by institutions.
- What it claims
- Its published technical report claims a false positive rate of 0.0041%. Independent testing in 2026 found it flagged none of 495 genuine human passages.
- Where it fails
- Even a very accurate detector is probabilistic, and it performs distinctly worse on text that imitates a specific author, missing around 10% of that in testing. It is also the detector least likely to produce a false accusation, so if it flagged you, treat the result seriously and assemble evidence rather than arguing about the tool.
- How to respond
- Do not attack the tool. If Pangram flagged a document you wrote, get the process evidence in front of a person quickly and calmly.
- Worth knowing
- Being accused by an accurate detector is a different situation from being accused by an inaccurate one, and it deserves a different response.
Nothing here detects anything either
Inkprint reports how a document was written and issues a signed certificate. It publishes no AI score, deliberately.
Analyze a document Read the full guideMore in ai detectors
Other signposts