Undetectable AI Promises a Human Score; Pangram Labs' Own August 2025 Table Reports 90.3 Percent Caught, Its Lowest of Twenty Tools

Sometimes, and the published evidence splits by detector, by version and by passage length. Undetectable AI's home page sells a rewrite that will "Score as human written on AI detectors and improve readability" ². On its 27 August 2025 results page, Pangram Labs reported catching 90.3 percent of that product's rewrites ¹.

Ten of the twenty tools on that table sit at 100.0 percent, and the lowest figure on the list belongs to Undetectable AI ¹. The page does not say how many samples it ran per tool ¹. Read the other way, close to one rewrite in ten did not come back flagged on Pangram's own test.

Two limits travel with the table. It describes one detector, and the school marking a paper next Tuesday may be running another. It also describes a replaced model: Pangram 4 succeeded Pangram 3.3.2 on 29 July 2026 ⁵.

A detection rate is a count on a test set, and Pangram's own model card says as much from the other end. It acknowledges a non-zero error rate, and it states that false accusations of AI use can lead to serious consequences, including reputational damage and emotional trauma ⁵. A percentage published about a corpus is not a probability about one student's essay.

Published abstracts, not essays. A 2026 University of Notre Dame paper retained 642 abstracts of 25 to 500 words, then humanized every variant with Undetectable AI ⁴. Of the rewrites the detector had labelled AI beforehand, fewer than 4 percent stayed flagged ⁴. That is a false negative rate above 96 percent.

The paper's word filter allowed abstracts from 25 words, below the 50-word floor Pangram 4's model card sets, and it does not say how many were that short ⁴⁵.

The scoring ran on Pangram 3.2 ⁴. Pangram 4 replaced Pangram 3.3.2 on 29 July 2026, and the newer model carries a separate humanizer score that calls a segment humanized only once that score clears a threshold set at 0.91 ⁵. The study measured the version available in its own window, which is all any study can do. A year separates two honest measurements of the same product, and only the Notre Dame paper names the detector version it ran ⁴; Pangram's page does not say ¹. The fuller record of what the detector company publishes sits in what Pangram publishes about humanized text.

One finding in that paper matters more in a department meeting than any detection rate. Light AI editing of the kind a guideline permits, the paper's proxy for honest assistance, was flagged more often than fully humanized text ⁴. Students may be flagged while drafting their own ideas and using AI only for clarity or grammar, yet others who generate inorganic drafts can evade detection after humanization ⁴.

A percentage, from whichever checker the institution licenses, with no field naming the tool that produced the text. Where that checker is Turnitin, its own page says the AI checker is meant to surface text spinners and AI bypassers, also called humanizers ⁷.

Pangram's model card describes segment labels and a humanizer flag ⁵, and Turnitin's page names no product ⁷. Neither describes a rewrite history. The arithmetic under a verdict is worth knowing before a conversation starts. Pangram labels each stretch of a document Human, AI-Assisted or AI-Generated. It then reads the document as AI only when at least 80 percent of its characters carry the AI label, and as Human at 90 percent human characters ⁵. A rewrite that moves a handful of segments can move the verdict, when the document sits near one of those lines, without touching the argument.

The market behind these rewrites is not a niche one. NBC News reported that Turnitin keeps a list of 150 tools that adjust text so a detector will not flag it, some charging as much as $50 for a subscription ⁶. The same report cited an academic integrity company that counted 43 humanizer sites drawing 33.9 million visits in a single month ⁶. The general version of the question sits here: whether AI humanizers work.

Because both sides ship, and the pages that sell the rewrite carry no version. The August 2025 table describes a detector generation since replaced and names no version of its own; the home-page promise carries none either ¹². The one test that recorded versions is the Notre Dame paper, which names Pangram 3.2 and Undetectable AI v11 ⁴. A detection figure is a reading of one tool against one model on one date.

A screenshot from last spring doesn't describe the model running this term.

The seller says so too. Undetectable AI's ethics page states that the company has never condoned cheating and never will. The same page ³ also states that "A single AI text detection score alone is NOT enough to rule out cheating or deception." A vendor selling rewrites and a researcher studying false positives land on the same sentence.

Whether a rewrite gets past a checker and whether it is allowed are separate questions. The University of Southern California's academic integrity office states that work created by a generative AI tool and presented as a student's own counts as plagiarism, whether it was copied word for word or paraphrased ⁸. Write the tool, the version and the date beside any score that lands in a file. A hearing six months from now may have no way to reconstruct which model produced a number.

Run the whole paper rather than the paragraph that reads oddly. Run it at its real length, on whatever version the school licenses today. Pangram's own page quotes a study finding its miss rate rises a bit on shorter passages while staying low ¹, and its model card refuses anything under 50 words ⁵. Then ask the questions a percentage cannot answer, starting with where the draft came from.

The draft history answers a question a second detector run cannot, where the text came from, and it costs one email: reading a Google Docs version history.

An oral follow-up is the other place to look, and it works whether or not a rewriting tool was ever involved: the questions worth asking after a flag.

A flag describes text. A finding about a person needs evidence about that person.

For the paper already on the desk rather than the product category behind it, Human at human.olive.is reads up to 1,000 words pasted in, three times a day, with no account and nothing stored. It reports how much of that text reads as machine-written or machine-edited, sentence by sentence, with a document verdict of Human, Mixed or AI. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. It is tuned for college-level academic writing, and below 50 words it declines to answer.

One in-house finding on rewriting tools, measured on Human's own held-out sets rather than by an independent audit, names no figure: whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. The student's side of that question is free humanizer tiers against paid ones.