Pangram's August 2025 Table Scores QuillBot at 100.0 Percent

Pangram Labs keeps a page on how well its detector does against humanizers, updated on 27 August 2025. The table has a Humanizer column and an Accuracy column, and the row the table spells Quillbot reads 100.0 percent ¹. The post sums it up as Pangram performing above 90 percent on all the notable humanizers it tested ¹. So a detector company is grading itself here.

What the page leaves out matters as much as the row. There is no sample size. Nothing on which model drafted the passages or how long they ran, and nothing on which QuillBot setting produced the rewrite ¹.

The post calls the column detection accuracy on text run through the tools and stops there, with no statement of what counted as a hit ¹. A number with no denominator under it can't be rechecked. On a humanizer benchmark that word usually means the share of rewritten machine-written passages the detector still called machine-written, which is an inference rather than a quotation. Look at the spread. Undetectable AI sits at 90.3 percent, TwainGPT at 92.7 percent, and ten of the twenty rows land on exactly 100.0 percent ¹. Grammarly is one of those ten. That says something about rewriting tools and nothing about a student who ran a spell check.

Pangram's 2026 technical report, written with the University of Maryland, gives ranges instead of a row. In it, humanized text was detected as AI 97.67 percent of the time and as Mixed or AI 98.83 percent. Per-system AI recall ran from 92.78 to 99.70 percent across thirteen commercial humanizers, which the report names only as Commercial A to M, so no row can be matched to QuillBot ². One tool's score doesn't transfer to another. And a number from August 2025 describes the detector and the rewriting tools of that month, not this one.

Competing humanizers, mostly, plus detector vendors scoring themselves. A check of this query on 18 September 2026 returned QuillBot's own product page and four rival humanizer sites. None of them was an independent test. One of the four, a competing humanizer, ran QuillBot's output through three detectors, reported it flagged each time, and reported its own output passing a paragraph later. The other three are reviews published by competing humanizers.

The publisher chose the passages and the detectors, and sells the tool that wins. That's an advertisement with a methods section. Originality.ai, another detector company, publishes its own humanizer figure of up to 97 percent ⁹, one more vendor scoring itself.

Why so many of these pages? Scale. NBC News reported in 2026 that Turnitin keeps a list of 150 tools charging as much as 50 dollars a subscription to adjust text so a detector won't flag it ³. The same report counted 43 humanizer sites and 33.9 million visits in one month, figures that came from an academic integrity company ³.

That's a large market with no referee in it.

An evidence-grade test needs nothing exotic: a document count, the model that wrote the documents, the detector version with its run date, and an author with nothing to sell on either side. QuillBot's own pages returned an error to the automated request made for this article, so nothing here reports how the company describes its own feature.

What the rewrite does to a paragraph is a separate question from who is scoring it, and it has its own answer: what an AI humanizer does to writing.

The surface, mostly. Pangram Labs defines a humanizer as a tool that rewrites machine-written text to evade detectors, and names synonym swapping and inserted nonsense among the documented techniques ⁴. Pangram's August 2025 page lists what it looks for in humanized text: tortured phrases, unnatural spacing, repetitive phrases and non-standard characters ¹.

Rewriting used to work well. A 2023 study built a paraphraser called DIPPER. It drove DetectGPT's accuracy from 70.3 percent down to 4.6 percent at a fixed 1 percent false positive rate, without appreciably changing what the text meant ⁵. Those detectors are three years and several model generations old.

Detection of rewritten text has been broken and then repaired since, on the record cited here. That's why an honest answer carries a date. A page that answers with a flat yes or a flat no, and names no month, is describing a moment it hasn't identified. There's also the cost to the writing. A synonym pass trades precise terms for near-synonyms, and inserted filler reads as filler to a marker long before any tool runs ⁴. A paragraph that survives a detector and loses the argument has already lost the grade.

The artefacts a rewrite leaves behind are visible without software: what humanized text looks like.

At USC, a rewrite that no detector flags is still covered by the plagiarism rule. The University of Southern California's academic integrity office, on its page as retrieved on 18 September 2026, states that material created by a generative tool and presented as a student's own work counts as plagiarism whether it was copied verbatim, near-verbatim or paraphrased ⁶. A paraphrase pass is the paraphrased case, so the policy reaches the rewrite along with the draft it came from.

A student at another school has that school's own policy to read.

The vendor pages fold two questions together. Whether a detector flags a passage is a measurement, and it moves with the detector and the month. Whether a submission breaks the rule is a policy question, and people decide it.

That distinction decides what a rewrite is actually worth, which is the argument under a humanizer is not a plagiarism remover and under is QuillBot cheating.

Keep the version history, then check the draft: both are worth more than another pass through a tool. A dated trail of drafts is evidence a detector reading can't produce. Read what a detector says about the draft itself, knowing what that reading is and what it isn't.

No detector score is a finding. Turnitin's guide names QuillBot as an example of the paraphrasing tools behind one of its report categories, and gives no catch rate for it ¹⁰. In a June 2023 post, Turnitin put the sentence-level false positive rate of its own detector at around 4 percent, which works out at roughly one highlighted sentence in 25 that may be human-written after all ⁷.

Length changes the reading more than most people expect. In 2023 Turnitin raised its own minimum from 150 to 300 words, reasoning that accuracy improves with more text ⁸. Human declines to score anything under 50 words at all, and 50 to 149 words gives it little to read. A reading on a short excerpt is weak in either direction.

The one false-flag figure Human publishes for its own detector is an in-house measurement on its own held-out set, not an independent audit, as of September 2026: 0 wrongly flagged of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. That figure counts human writing misread as machine-written, so it says nothing about how often a QuillBot rewrite slips through.

A dated folder of earlier drafts outranks a screenshot of a vendor's table.