Pangram Reports 97.67 Percent on Humanized Text; Outside Tests Split

Pangram publishes a catch rate for humanized text and dates it, and the two 2026 figures name the model version as well. An August 2025 update reported above 90 percent on all the notable humanizers Pangram tested ². Pangram 3.3, released 13 May 2026, claimed twice as many commercially humanized texts caught as the version before it ⁴. The Pangram 4 technical report gives a rate: 97.67 percent ¹.

Every one of those figures is Pangram's own, measured on Pangram's own set.

The average is the number that travels worst. In the same report, the share of humanized text read as AI-generated runs from 92.78 to 99.70 percent across thirteen commercial systems, and the share read as Mixed or AI-generated from 95.28 to 100 percent ¹.

Seven points separate the best-caught system from the worst. At the bottom of that range, roughly one humanized document in fourteen reads as something other than AI-generated, and one in twenty-one reads as neither Mixed nor AI. That is a different Tuesday from the one the headline figure describes.

Pangram 4 also scores the processing itself. Next to its provenance labels it carries a humanizer score, and marks a segment as humanized once that score passes a threshold set at 0.91 on release ³. One check yields two readings: whether the writing reads as machine-written, and whether it reads as rewritten to look otherwise. On text already read as AI, the humanizer flag fires between 91.52 and 99.39 percent of the time across the same thirteen systems ¹.

A high humanizer score describes the rewriting, not the first draft. The pattern it reads is the subject of what humanized text looks like.

Three outside results name Pangram specifically, and they do not agree. A 2025 National Bureau of Economic Research working paper evaluated four detectors and reported near-zero error rates for Pangram. Those rates held across models and threshold rules, on passages of 50 words or fewer, and on humanizer output ⁵. The Atlantic ran ChatGPT and Claude output through Walter Writes AI in May 2026, and reported that Pangram called the result human-written every time ⁶.

The date on the second result is the part that falls off when it gets quoted.

May 2026 is before Pangram 4 shipped on 29 July 2026 ³, so the article describes an earlier model meeting one commercial humanizer, not the version a school licenses this term. That doesn't make the finding wrong. It makes it a measurement with an expiry date, which is what the vendor's July table is as well.

The third result names the model version it tested, which is what makes it readable at all. A 2026 University of Notre Dame paper scored Pangram 3.2 and GPTZero on abstracts humanized with Undetectable AI v11, and fewer than 4 percent of AI-labelled rewrites stayed flagged across the two detectors ⁸. The 3.3 release note says some text from the latest OpenAI and Anthropic releases had been labelled human by the model 3.3 replaced ⁴. The cohort matters too. Those were 642 abstracts of 25 to 500 words ⁸, short academic prose and not term papers. The NBER paper's robustness on passages that short was measured on a different model and a different set ⁵.

Nature's 2026 survey of the field states the position plainly. Only the accuracy rates the firms announce from their own internal testing can be current, and those are not externally verified ⁷. Set the vendor's July table beside a humanizer's undated promise and the comparison is between two unaudited documents, one of which at least says what it measured and when.

The one humanizer page cited here sells a promise rather than a measurement, and publishes nothing behind it. Rephrasy's homepage guarantees a 100 percent pass rate on all major AI detectors, with no sample size, no date, no detector version and no method beside it ⁹. The promise names no detector, so reading it as an answer to Pangram's figures is an inference. Pangram's claims can be argued with because they name what was measured; a bare pass rate can't.

Undetectable AI's ethics page states that a single AI text detection score alone is not enough to rule out cheating or deception, and that the company has never condoned cheating ¹⁰. That is a statement about a clean reading: passing a detector clears nothing.

Human's own in-house measurement, as of September 2026, found that whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. That measurement is Human's own held-out testing rather than an independent audit, and no figure is published with it.

The gap between the free tools and the subscriptions is the subject of free humanizer versus paid tier.

Each side of this argument publishes a snapshot of a moving target. Pangram shipped 3.3 on 13 May 2026 and Pangram 4 on 29 July 2026, eleven weeks apart ⁴³. The 3.3 announcement says what prompted it: some text from the latest OpenAI and Anthropic releases had been labelled human ⁴. A humanizer benchmark from 2025 therefore describes a model that has been replaced since.

Two questions settle most humanizer-versus-detector claims. Which detector version was tested, and on what date. A humanizer page naming neither is marketing rather than measurement, and a vendor table naming both is still a vendor table.

For a school the consequence is a procurement question: a licence bought on a figure measured in 2025 rests on a model that has since been replaced, and the renewal question is what the current version was measured on, and by whom.

The wider accuracy picture, including what Pangram reports on ordinary human writing, sits in is Pangram accurate.

Treat the paragraph as a reason to ask rather than a conclusion, which is what the vendor recommends first. Pangram's model card records a non-zero error rate and states that false accusations of AI use can cause reputational damage and emotional harm ³. Its chief executive told The Atlantic the tool should never be the ending arbiter, only the starting point for a more thorough investigation ⁶.

On 642 short abstracts scored with Pangram 3.2 and GPTZero, honest, guideline-compliant AI editing carried a higher sanction risk than humanizer-assisted evasion ⁸. If that carries to coursework, which the paper did not test, a rule policed by a detector rewards concealment.

Nature relays a figure from Pangram's own technical study: when the firm used consumer AI to substantially modify human-written student essays, the model still labelled the results fully human 41 percent of the time ⁷. A real draft reworked by a chatbot is the case that figure describes.

What survives scrutiny is process evidence. The draft history and the document's version log say when the words appeared, and five minutes of conversation about the argument on page two says whether the student can hold it.

A paragraph scoring as humanized is a reason to ask, and the asking is set out in how to talk to a student about AI.