About AI Humanizer Benchmark

Once a month, AI Humanizer Benchmark gives 11 AI humanizers the same set of AI-written texts to rewrite. Every rewrite goes to 7 commercial AI detectors, and a fixed, published formula turns the results into one overall score per tool. Each cycle's leaderboard is published together with its inputs, the rewritten outputs, every detector score and the scoring script.

A note on ownership: the people who build UndetectedGPT also run this site, and UndetectedGPT appears in the rankings alongside the other tools. We say this on the homepage, in the methodology, in the terms and on this page. UndetectedGPT is treated the same as every other tool. It is scored and penalized under exactly the same rules. Decisions about the benchmark itself (which tools are tested, how they are scored, what gets published) are made separately from the UndetectedGPT product team. Because the operator also competes on the leaderboard, we ask readers to check the evidence rather than take our word for it.

Humanizer vendors often claim top bypass rates, broad detector coverage and strong meaning preservation all at once, and readers have few ways to test those claims. This site gives one fixed comparison, with the same inputs, detectors and scoring every month. You can reproduce any rank shown here by downloading that cycle and recomputing it. Read more about why we built this.

Open source

Every cycle's input texts, rewrites, detector results and scoring script are stored in our GitHub repository. A copy is also published as a Hugging Face dataset under the AI Humanizer Benchmark organization. It loads directly with load_dataset. Each published cycle is also archived on Zenodo, which mints a DOI for it, so the data stays citable and retrievable independently of this site.

Citing this benchmark

For the project as a whole, cite AI Humanizer Benchmark at https://aihumanizerbenchmark.com.

To cite the data as a whole, which always resolves to the newest published cycle:

AI Humanizer Benchmark. (2026). AI Humanizer Benchmark: monthly cycle data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22966275

Scores change from cycle to cycle, so a citation that quotes a specific number should reference that cycle's own DOI instead. Each one is listed under Versions on the Zenodo record, alongside a ready-made citation for it.

Anti-bias commitments

• All tools are tested with one methodology and one scoring formula. No tool gets special treatment.
• Anyone can read the scoring script, and it returns the same result every time it runs on the same data.
• Rankings come only from the published methodology. Nobody moves a tool up or down by hand.
• Placement is not for sale.
• No tool is ever removed because it outscores another, and every tool we test has its results published.

The fairness page covers these rules in more detail and explains how to challenge a result.

Contact

• Press: contact@aihumanizerbenchmark.com
• Partnerships: contact@aihumanizerbenchmark.com
• To challenge a score or report an error, submit a dispute. The fairness page describes each step.