Pangram's Own Guidance After a Flag: Drafts, Notes, a Conversation

Pangram's guidance for exactly this moment is a page on its own site, published in February 2025 under its chief executive's name, and it is written to the teacher rather than to the student ¹. It doesn't ask for another scan. It asks what the student can show about how the paper got written, and it names the notes, the early drafts and the file's version history.

The page puts two questions to the teacher directly: "What research did they do for this assignment, and did they take notes? Do they have early drafts saved?" ¹ What a drafting record is worth depends on what it shows: a paste, or a draft typed straight through from top to bottom, reads to the page as a sign of AI help ¹. "If they have a robust multi-hour writing history, then that's some very compelling evidence that they wrote the work themselves," the page says ¹. That wording came from the company that built the detector.

The page recommends talking to the student, and says a detection warrants further investigation before any punitive measure ¹. The comparison it borrows from The Cheat Sheet, an academic integrity newsletter, is the metal detector at an airport: the alarm starts a search, and it doesn't end one ¹.

Pangram's September 2026 post on false positives puts its overall rate at roughly 1 in 10,000, and says the figure moves with the kind of writing ². It carries two academic figures: the summary puts academic writing at 0.02 percent, about 1 in 5,000; the domain table and the Turnitin comparison put academic essays at 0.004 percent, about 1 in 25,000 ². All three are the company's own measurements, not an independent audit ².

The post also names writing the company doesn't want screened at all: short bullet lists or outlines, single-sentence responses, and formulaic text such as data lists and templates ². The model card sets a floor of 50 words in complete sentences ³. A 40-word discussion post is under that floor, so a reading reported on one needs explaining before anything else ³.

The rate and its cost both sit in Pangram's own material. The guidance page says a positive reading "could be real, or it could be the statistically anomalous one-in-ten-thousand situation where Pangram gets it wrong" ¹. The model card adds that the model carries a non-zero error rate, and that a false accusation of AI use does real harm ³.

The Atlantic set that rate against the scale of American schooling in May 2026, writing that Pangram "would produce plenty of false accusations even with a one-in-10,000 error rate" ⁴. That is the magazine's estimate, and it assumes every paper is screened and that the published rate holds on all of them. A false flag becomes an accusation only when someone treats it as one.

A high percentage is a confidence figure about part of the document, not a count of machine-written sentences. Pangram's guidance page spells this out: a 99 percent score on a segment means the model is 99 percent confident that AI was used somewhere in that text, not that every word came from a machine ¹. The figure says how sure the model is. It doesn't say how much of the essay a person wrote.

Long documents get split into segments before scoring, and the verdict is arithmetic on characters. The model card gives the thresholds: Human at 90 percent human characters, AI at 80 percent machine characters, Mixed in between ³.

A Mixed paper is pointing at a stretch of text, and what the AI-Assisted label marks on a mixed paper is a different question from what the document verdict says.

Segmentation cuts the other way too. Pangram's chief executive told Nature in August 2026 that a paragraph judged human on its own can switch to AI once it sits inside a larger segment ⁵. A student who pastes one flagged paragraph into a free checker and gets a clean result has not settled anything, since the paragraph was scored as a different input. A teacher who pastes it and gets a worse one has not settled anything either.

A second opinion from a different detector is a different kind of input, and a fragment gives it less to read than a whole essay does. Human's public check reads up to 1,200 words, one paper a day, with no account, so a longer essay is a partial read there. Human's detector scores the text in memory and de-identifies any personal information for additional training. Human reports its false-flag count as its own in-house measurement on its own held-out documents, as of September 2026, not an outside audit: 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. Treating that count as the odds on one particular essay is an assumption. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. A second reading does not replace the drafts.

Pangram names five of them on its own guidance page: grammar checkers such as Grammarly that write as well as correct, translation tools built on language models, the Help me write feature in Google Docs, phrasing carried over from a brainstorming chat, and asking ChatGPT for wording advice ¹. The page lists these as uses of AI that may trigger detection, and whether any of them breaks a course's rules depends on that course's policy ¹.

Bloomberg Businessweek, testing detectors other than Pangram in 2024, reported that Grammarly's rewriting features turned work that read as fully human into fully machine-written, while plain spelling and grammar fixes had only a marginal effect ⁶. Whether Pangram reads the same edits the same way is not something that report shows. Which Grammarly feature was switched on separates an edited sentence from a generated one.

Pangram's own page calls this a mismatch ¹: what a teacher counts as cheating, what a student counts as cheating and what sets off a detector are three different lines. The page's fix is an AI-use policy, and it links a tier system so that students and teachers agree on which assistive tools are allowed ¹.

A reply that works is short and attaches the record. It names the file the essay was drafted in, offers the version history, and asks for a meeting. The notes and the sources go in as attachments. If a vendor-stated limit applies, one sentence citing the vendor's own page carries it; a general argument about detector accuracy is weaker ground for a student, and it moves the conversation off the page the vendor wrote. Attach that page instead.

The record also has to arrive in readable form, and what a Google Docs version history can and cannot show is worth checking first.

A clean typing record is not proof by itself. Process documentation, argues a 2026 framework in Frontiers in Artificial Intelligence, is evidence of engagement rather than proof of independent authorship, since an artifact can be staged ⁷. That limit lands better coming from the student than from the instructor.

So the record goes over as the account of how the essay was made, next to the essay itself. The email after an AI flag and what a student can actually show cover the wording and the evidence in detail.

One line on Pangram's guidance page cuts the other way. A record of repeated flags counts, and the benefit of the doubt thins each time ¹.