Introducing HumanizerBench

If you have shopped for an AI humanizer in the last year, you already know the problem. Every product claims the highest bypass rate against the most detectors, usually with a number like 99.9% set in big type on the homepage, and almost none of them show the test behind the number. The articles that rank for “best AI humanizer” are mostly affiliate pages. The reviews under them are often seeded. So you buy one tool, it underdelivers, you buy another, it underdelivers in some new way, and eventually you settle or give up.

The space got crowded fast, and the louder it got, the harder it became to tell a real result from a marketing line. We built HumanizerBench to answer the question with data instead of adjectives.

What it actually does

Once a month we run the same set of prompts through every AI humanizer we can buy access to. We take each humanized output and submit it to five commercial AI detectors: GPTZero, Originality.ai, Copyleaks, Winston AI, and ZeroGPT. A deterministic scoring script turns those detector verdicts, plus how well each rewrite kept the original meaning and how clean the grammar stayed, into one score. Most of the weight sits on whether the output actually beat the detectors. Meaning preservation and readability carry most of the rest, and a short list of penalties docks tools that pad length, drift off topic, or hand back the input unchanged.

Then we publish the whole cycle: the input prompts, every humanized output, every detector response, and the exact scoring code that produced the ranking. If we say a tool placed third, you can download that cycle and reproduce the third place finish yourself.

We pay for everything, the same as any customer

Here is the part nobody asks about, and it is the part that matters most. We pay for every detector and every humanizer ourselves, in full, every month. No vendor gives us a free account. No vendor hands us a private API key. We don’t take comped access from any company on the board, including the detectors we test against.

That sounds like an accounting footnote. It is actually what keeps the whole thing honest. The moment a humanizer can tell which account belongs to the reviewer, it can treat that account differently: route it to a better model, waive a rate limit, return a cleaner result than a paying customer would ever get. So we don’t announce ourselves. We sign up and pay across a number of ordinary separate accounts, every cycle, and we test through the same front door everybody else uses. What you see on the leaderboard is what you would get if you swiped your own card. It is slower and more expensive to run a benchmark this way, and it is the only way the numbers mean anything.

How we keep ourselves honest

We should be upfront about the obvious conflict. HumanizerBench is operated by WriteHuman, which sells one of the humanizers on the board. WriteHuman is tested under the exact same prompts, detectors, scoring, and penalties as every other tool, the scoring code is public, and the longer version of that argument lives on the why we built this page.

Two design choices back it up. First, the scoring is mechanical. Rankings come out of a published script run over published data, not out of a meeting, so if a number on the board didn’t come from that script running on that data, anyone running npm run verify against the public repository would catch it. Second, we can’t quietly tune the test toward our own product, because we don’t know the test in advance either. Each cycle’s prompts are seeded from a secret random value, and the only thing we publish when a cycle opens is a hash of that value. The prompts become reproducible when the cycle closes and we reveal the original. Nobody, us included, can pre-compute next month’s prompts and train against them.

If you make a humanizer, get on the board

If you build an AI humanizer and you aren’t on the leaderboard yet, we want to add you — submit it for a future cycle. The fairness page covers how to dispute a result if you think a past one was wrong. We read and answer every one of those.

Where to start

• The leaderboard for this cycle’s rankings.
• The methodology page for the formula, the weights, the penalties, and the steps to reproduce any score by hand.
• The public repository for every input, every output, every detector response, and every scoring script we have ever run.

We publish a new cycle every month, and we’ll post here whenever one lands or the methodology changes.