Tool Benchmarks, independent tests of AI tools with every score published
Tool Benchmarks tests free AI tools on material where the right answer is known, and publishes every result, including the raw scores. Our first test ran seven open-source AI text detectors on 756 texts; the best caught 68.1% of AI writing, and 117 of 126 human texts were flagged by at least one detector.
Tested
How we test
Every benchmark here follows the same rules. The test material comes from sources where the truth is known, such as human writing published before ChatGPT existed. Each tool runs on our own hardware with the settings its makers publish. We report the share of mistakes with a 95% interval and publish the data behind every number, so you can check it or reuse it.
We do not accept payment from the makers of the tools we test. Pages are written with AI help and checked against the results by a script that rejects any number it cannot trace; the editorial policy explains how.
Where to start
The AI detector leaderboard ranks seven free detectors. If you are worried about a false accusation, read which human writing gets flagged and how detectors handled famous texts such as the Declaration of Independence. The guides explain what the numbers mean.