AI Humanizer BenchmarkSeptember 2026 prompt evidence
This cycle used 33 prompts, each with its own AI-written input text. All 11 ranked humanizers rewrote every input text, and 7 commercial AI detectors scored each rewrite. Pick a row below to see a prompt's input text, each tool's rewrite, and every detector's score.
How to read the columns. A tool counts as “Passed” when 6 or more of the 7 detectors gave its rewrite a human-likelihood score of 0.50 or higher. It counts as “Failed” when any detector that returned a score put it below 0.50. So one tool can count in both columns. “Hardest” is the detector that caught the most tools on that prompt. If several are tied, only one is shown.
How to read these scores
Each detector's result is converted to a human-likelihood score between 0 and 1. A score of 0 means the detector is sure the text is AI. A score of 1 means it is sure the text is human. Dots are green at 0.50 or above and red below. Hover a dot to see its exact score. The 0.50 cut-off only sets the dot colors. A tool's bypass rate on the leaderboard takes the median of the 7 detector scores for each test, then averages those medians. That gives a continuous number, so it will not match the share of green dots. The same goes for each per-detector rate. It is the average of that detector's scores (the number detector pages rank by), not a count of green dots.
Raw data
Each ranked tool has its own JSONL file for this cycle, with one row per prompt. A row holds the prompt, its input text, the tool's rewrite, its meaning preservation and readability scores, and the run status. Scores from the 7 detectors are in separate files, one per detector, in the cycle's GitHub folder.
