Does WriteHuman Bypass GPTZero? Its Own Data Says It Varies (2026)

Does WriteHuman bypass GPTZero? Its own published data says 61% one month and 88% the next — and it is the only company in this market that lets you check.

By Vlad Ivanov — publisher of Detection Drama, tracking AI-detector and humanizer behaviour across 245 published tests and teardowns. Last updated: August 2026

Does WriteHuman bypass GPTZero?

Almost every page in this series ends the same way: the vendor claims a number, nobody can reproduce it, and the honest answer is that nobody knows.

This one is different, and the difference is worth stating up front. WriteHuman publishes its raw data — every source passage, every humanized output, every detector verdict, and the scoring code — in a public repository under an open licence. We did not take its word for anything below. We downloaded the files.

What they show is not the number on the marketing page. It is something more useful: the answer moves a lot, month to month, on the same test.

What does the raw data actually say?

WriteHuman operates HumanizerBench, a monthly benchmark scoring humanizers against five detectors — GPTZero, Winston AI, ZeroGPT, Copyleaks and Originality.ai. We pulled its per-tool data file on 3 September 2026. Here is WriteHuman against GPTZero, cycle by cycle:

Source: data/humanizers/writehuman.json, HumanizerBench public repository, downloaded and parsed 3 September 2026. Figures are bypass rates — the share of outputs the detector rated human.

A 27.6-point swing in four months, on the same benchmark, the same five detectors, the same stated method. Nothing about the product changed that much. What that range really measures is how unstable this question is.

Can you check GPTZero yourself?

Turnitin sells no individual licences and shows students no AI report, so nobody outside an institution can produce a genuine Turnitin screenshot. GPTZero removes that excuse entirely. Anyone can paste up to 10,000 characters into gptzero.me right now with no account, no card and no signup; a free account raises the per-scan limit to 150,000 characters with a 10,000-word monthly quota, and paid tiers start at $9.99 a month.

So when a humanizer claims to beat GPTZero and shows you nothing, that is a choice, not a constraint.

It also means you can check your own text. A genuine current GPTZero result shows a three-class classification — human, AI or mixed — a confidence category rather than a bare percentage, a probability breakdown across all three classes, and sentence-by-sentence highlighting. An image showing only “X% AI”, or one displaying perplexity and burstiness, is either fabricated or dates to roughly 2023.

What happened when GPTZero tested it?

On 30 July 2026, GPTZero published its own hands-on review of WriteHuman. It ran a single 800-word ChatGPT-generated sample through WriteHuman and then through four detectors.

Source: gptzero.me/news/writehuman-ai-review, Mehal Rashid, 30 July 2026. Published by GPTZero about a tool that competes with GPTZero’s interests; n=1.

GPTZero’s verdict line was “We are highly confident this text was originally AI, but rewritten by AI or human” — the Paraphraser Shield firing — and its conclusion was that WriteHuman “couldn’t bypass GPTZero as it claims.” It went on to argue that “only poor AI detectors fall for AI humanizers.”

Take that for exactly what it is: one sample, published by the detector under test, on a page arguing its own product is superior. On its own it would not be worth much.

How much should you trust a benchmark its winner owns?

Both halves of that sentence are true and both matter.

The creditable half: HumanizerBench publishes samples.json, tests.json, detector-scores.json and a frozen scoring.js for each cycle, under MIT and CC BY 4.0 licences. The stated method is that every tool is “bought and paid for by us and run by hand on the most evasion-focused setting it advertises”, with “no affiliate arrangements and no vendor-supplied numbers.” The September cycle covers fourteen tools across seven writing categories. Nobody else in this market has published anything close. We were able to verify every figure in this article against those files, which is not something we have been able to do once in the previous forty-five.

The other half: WriteHuman ranks first in every cycle it has published — 83.59, 73.07, 76.69 and 78.29, ahead of Stealth Writer, HIX Bypass and Undetectable AI. A scoring formula weights bypass at 42%, meaning preservation at 32%, readability at 16% and consistency at 10%, with penalties deducted for quality failures; those weights were chosen by the company that wins under them. And the July write-up disclosed the conflict of interest while the August one — the one headlining WriteHuman’s own top placement — did not.

Our reading: use the raw files, treat the leaderboard ordering with the caution any self-scored ranking deserves, and note that publishing data which visibly undercuts your own marketing is not the behaviour of a benchmark designed purely to flatter.

Which detector is actually the hard one?

This is where the usual story breaks down, and it is worth knowing before you buy anything on the strength of a GPTZero number.

Across most of this series, GPTZero is the softest detector on record — the UChicago BFI paper found it missing roughly half of humanized text while Pangram missed almost none. For WriteHuman in June and July, that held: Originality.ai was the hardest of the five, at 46.3% and 53.6%.

By September it had inverted. GPTZero, at 71.8%, was the hardest of WriteHuman’s five detectors — below Winston AI at 84.1%, ZeroGPT at 77.9%, Copyleaks at 86.7% and Originality.ai at 91.4%. One tool, one cycle, so do not over-read it. But it is a direct, checkable counter-example to “GPTZero is the easy one”, and it came from data the tool’s own maker published.

What matters more than the tool you pick?

What you are writing. The September category breakdown for WriteHuman, from the same file:

Source: HumanizerBench September 2026 cycle, category breakdown for WriteHuman, downloaded 3 September 2026. Aggregated across all five detectors, not GPTZero alone.

Same product, same month, same detectors: near-certainty on one genre and roughly a coin flip on another. The spread between categories is wider than the spread between most tools. Any vendor headline that gives you one number for “bypass rate” is averaging across a range that looks like this.

And what does GPTZero’s own paper say about it?

GPTZero’s February 2026 paper names nine bypass services in an appendix table. WriteHuman is rated “Medium”, with Grubby AI and TwainGPT; Undetectable is the only “High”, and StealthGPT, Quillbot, StealthWriter, HIX and GPTinf are “Low”.

How reliable are the detector numbers on either side?

GPTZero’s own FAQ concedes: “Our classifier is not trained to identify AI-generated text after it has been heavily modified after generation.” Its developers page simultaneously claims a “Paraphraser Shield” means “even if AI content has been altered to look more human-like, GPTZero can detect it.” Both statements are currently published.

The two leading detector vendors publish opposite numbers on the same question. GPTZero’s February 2026 paper reports 93.5% recall on 1,000 texts run through nine bypasser services, and puts Pangram at 49.7%. Pangram’s DAMAGE paper reports GPTZero at 60.04% on humanized text, and 34.53% at GPTZero’s own default threshold. Each benchmark shows its publisher winning, and neither is peer-reviewed.

On false positives, GPTZero was one of seven detectors in Liang et al. (2023), which reported a 61.22% average false positive rate across those seven on 91 TOEFL essays. No per-detector figure was ever published, so that number is not GPTZero’s. GPTZero now claims it has cut its own TOEFL false positive rate to 1.1% — a figure no third party has verified.

What should you take from this?

WriteHuman did the thing this market never does: it published the evidence, including the parts that make it look worse. That deserves saying plainly, and it is the reason this page could be written from data rather than from claims.

The data says its GPTZero performance is real but unstable — between roughly 61% and 88% across four consecutive months — that GPTZero’s own adversarial test failed it during its weakest month, that GPTZero was the hardest of five detectors for it in the most recent cycle, and that the genre you are writing swings the outcome by nearly fifty points on its own.

None of that supports betting a submission on a number. It supports checking your own specific text, which GPTZero lets you do free, in about a minute, with no account — and keeping your drafting history either way.

If your real worry is being wrongly flagged on your own writing, keep your drafting history — it costs nothing and it survives a detector being wrong. It matters most for the writers detectors treat worst: ESL writers and AI detection and false positives for neurodivergent students. Check before you submit with the best pre-submission check and the Turnitin self-check, and strip the obvious tells by hand first via what to remove before using an AI humanizer.

For the product itself see our WriteHuman review, and the same question against Turnitin and Pangram. Also in this cluster: StealthGPT, Clarity Bubble, GPTHuman.

Frequently asked questions