Does WriteHuman Bypass Pangram? What the Evidence Actually Shows

Vlad Ivanov · LinkedIn
Runs detectiondrama.com, where he has tested and documented 40+ AI humanizers and detectors against Turnitin, GPTZero, Originality.ai and Pangram since 2024. This piece is an evidence review of published third-party data, not a first-party test.
Last updated: July 2026

Does WriteHuman bypass Pangram? The only published 2026 test says yes — a 100% human score — but that is a single run, and Pangram’s own 20-humanizer benchmark, which catches every tool it has tested at 90% or better, has never published a WriteHuman number.

Methodology note: This is an evidence review, not a first-party test. I did not run WriteHuman output through Pangram myself. Every number below is published by a named source and linked to its primary document, with the funding and recency of each source stated plainly so you can weight it yourself. Where a tester earns commission on the tools involved, I say so.

Has anyone actually tested WriteHuman against Pangram?

Yes — once, publicly, and recently. On July 7, 2026, writer Anangsha Alammyan published a four-tool bypass test that ran the same AI-generated article through WriteHuman, GPTHuman, HumaLingo, and Undetectable AI, then scored each output on Pangram Labs and Quetext.

WriteHuman was the outlier. Its output came back 100% human on Pangram and 96% probability human on Quetext — the only tool of the four to pass both detectors. The other three did not just lose; they were flagged 100% AI-generated by Pangram, all of them. Her conclusion: WriteHuman was “the best AI humanizer for AI detector bypass” in that test, with the explicit warning that no humanizer “can guarantee detector bypass every single time.”

That is a real, screenshot-documented result on the current-generation detector — the test ran roughly eight weeks after Pangram 3.3 shipped in May 2026. It is also, as of this writing, the entire public evidence base for the “yes” side.

Why isn’t WriteHuman in Pangram’s own benchmark?

Here is the strange part. Pangram publishes its own humanizer benchmark — updated August 27, 2025 by co-founder and CTO Bradley Emi — covering 20 humanizers and paraphrasers. Every single tool in the table is caught at 90% or better: Undetectable AI is the “best” evader at 90.3% detected, StealthGPT sits at 95.6%, and ten tools including Quillbot and Grammarly are caught 100% of the time.

WriteHuman is not in the table. Not tested, not named, not scored — despite being one of the most-searched humanizers of 2025-2026 and a fixture in every “best humanizer” listicle. Pangram has published a detection rate for tools as obscure as surgegraph.io and DIPPER, but not for WriteHuman.

There are three boring explanations and one interesting one. Boring: the benchmark predates WriteHuman’s current popularity; the sample of 20 was arbitrary; or an updated table is coming. Interesting: vendors generally publish the benchmarks they win. I have no evidence Pangram is sitting on a bad WriteHuman number — but the absence means the single strongest voice on the “gets caught” side of this question has never actually spoken on it.

How much weight can a single test carry?

Less than the headline suggests. Before you act on the 100% human score, price in what it actually is:

It is n = 1. One AI-generated source article, humanized once, scored once. No repeat runs, no genre variation, no length variation, no per-mode breakdown — WriteHuman offers Standard, Professional, Academic, Blog, and Casual modes, and the post does not say which one produced the pass.

The tester has affiliate links on both sides. The post carries referral links for WriteHuman and Pangram (and Quetext, and the losing tools). That does not make the screenshots fake — the same author flagged three of her four affiliate tools as failures, which is more honesty than most of the niche manages — but it is a commercial context, not a lab.

Detectors are moving targets. Pangram ships model updates specifically to close humanizer gaps, and says humanizer detection is an active research area. A pass recorded in July 2026 is a fact about July 2026. As I documented when Turnitin rolled out bypasser detection, the arms race regularly deletes last quarter’s results.

And the result is theoretically anomalous. Pangram’s own research position is that fluency helps detection: “the more readable/fluent the humanizer text is, the more likely it is to be detected by Pangram.” WriteHuman’s entire pitch is fluent, readable rewrites. Either WriteHuman found a way to be fluent without the statistical tells — or this particular sample landed in the tail. One run cannot tell you which.

What does independent research say about Pangram and humanizers?

Everything independent says Pangram is the hardest detector in the humanizer game — which cuts both ways for this question: it makes the 100% pass more impressive if real, and less likely a priori.

A University of Maryland benchmark (Russell et al.) put Pangram at 97% accuracy on humanized text, against GPTZero at 46%, Fast-DetectGPT at 23%, and Binoculars at 7%. Economists Brian Jabarian and Alex Imas at the University of Chicago’s Becker Friedman Institute compared four commercial detectors across 3,984 texts and found Pangram was the only one whose performance stayed robust against humanizers, detecting “nearly 100% of AI-generated text” in longer passages, while GPTZero “largely loses its capacity to detect AI-generated text” once a humanizer touches it. Their false-positive finding matters too: Pangram claims roughly a 1-in-10,000 false positive rate, and the UChicago data backed its dominance at strict false-positive thresholds.

So the “does it bypass X” question is really only interesting for Pangram. If your text only needs to slip past GPTZero, the published data says most decent humanizers manage that much of the time. Pangram is the wall — and one tool just posted one clean climb of it.

How does WriteHuman compare to StealthGPT and Undetectable AI on Pangram?

Pulling both sources into one table — note carefully which number comes from whom:

Sources: Pangram Labs humanizer benchmark (vendor, Aug 2025); Alammyan four-tool test (independent, affiliate-monetized, Jul 2026).

I compared these three tools head-to-head on features and pricing in Undetectable AI vs WriteHuman vs StealthGPT, and each has a full standalone review: WriteHuman review, StealthGPT review, Undetectable AI review. For the detector side, see my Pangram review and the general guide to bypassing Pangram (short version: almost nothing does).

Does passing Pangram mean passing Turnitin or GPTZero?

No. Detectors do not share a brain. The same text can score 100% human on one and 100% AI on the next — that is exactly what happened to GPTHuman in the July test (100% AI on Pangram, 80% human on Quetext). Turnitin runs separate bypasser detection that reports what percentage of AI text was modified by a humanizer tool, and some humanizers that survive one detector get shredded by another. Detection is probabilistic; a pass is a sample, not a property. The false-negative side has the same shape — see the published false negative rates across tools.

What should you do before submitting anything?

If real stakes ride on the answer — a thesis, a visa statement, a job application — do not outsource the risk to one blogger’s screenshot, mine included. Check your own text against the detectors that will actually judge it, read the output yourself, and keep drafts as evidence of process. Detection scores move; a documented writing history does not. The Nature/Pangram academic-humanizer exchange is a good reminder that even researchers disagree about what these scores mean.

FAQ