Can an AI Detector Flag a Provost's Own Writing?: What The Dartmouth Found, and What the Percentage Still Does Not Prove

What The Dartmouth actually measured

The Dartmouth's October 14 investigation looked at academic articles and opinion pieces Schnell published, five from before the November 2022 release of ChatGPT and nine from this year. Pangram 4.0 called every pre-2022 sample 100% human-written. On the nine newer pieces, the median AI-written share was 96%. The student paper is not claiming it watched him type. It is reporting what one vendor's model printed when the files were run through it.

Pangram is a commercial classifier trained to separate human documents from generated ones. The Dartmouth notes an October 2025 University of Chicago audit that described a near-zero error rate, a company figure of about a 0.01% false-positive rate, and a July 2026 company note that version 4.0 mistakes "AI polish" for fully AI-written text about 0.009% of the time, down from 0.18% on 3.2. Those are lab-style rates on the sets the auditors and the company chose. A rate that small is easy to quote and hard to apply to one provost, one genre, and one editing habit.

Schnell's reply, given across statements in September, is narrower than the headline. He said he uses AI for language refinement, copyediting, clearer organization, and a faster writing process, and that the tools do not stand in for his intellectual judgment. He also said the description of a Washington Post essay as 100% AI-written is not accurate. Both things can be reported without picking a winner: the detector saw a strong generated-looking signal, and the author says the signal came from editing rather than from handing the argument to a model.

Polish and authorship are not the same measurement

Detectors score the finished page. They do not score the hours, the outline, the sources, or the decision to keep a sentence. A full draft from a chatbot and a human draft that was then tightened by a chatbot can both come out even, formal, and low on surprise. An August 2026 University of Notre Dame preprint, as described in The Dartmouth's reporting, says current detectors cannot reliably tell AI editing apart from a full model draft. That is the gap this story sits in.

The older comparison is still the cleanest evidence in the piece. Schnell's pre-ChatGPT writing did not trip Pangram. His 2026 writing did, at a very high median. Something about the newer prose changed. The detector is allowed to notice that. The notice does not name the something. It could be heavy generation. It could be the sort of line-level polish that makes careful academic English look like a model, especially once an editor has been asked to smooth it.

Schnell also pointed at a bias that researchers have documented for years. Stanford's 2023 study found seven detectors were near-perfect on essays by U.S.-born eighth-graders and still flagged 61% of TOEFL essays by non-native English writers, with 97% of those essays flagged by at least one tool. Formal, predictable English is exactly the texture a perplexity-style score treats as machine-like. A provost's op-ed is not a TOEFL essay, and the bias does not erase a 96% result. It does mean "near-zero error" was never a promise about every polished paragraph.

What a high badge still cannot do

A percentage is a share of the text that still matches the detector's picture of generated prose. It is not a probability that the person cheated, and it is not portable across products. Pangram's 96% will not reappear as 96% on GPTZero, on ZeroGPT, or on DetectGPT, because each model is trained and thresholded on its own. How AI writing detectors score a paper is the longer version of that disagreement.

Campus tools are a third family. Can Turnitin be wrong covers the same mistake in a course setting: a surprising indicator gets treated as a finding, and the finding gets treated as proof. Turnitin's classifier is not Pangram. A kind number on a public site does not cancel a harsh one in a learning-management system, and the reverse is also true.

If the worry is that a person wrote the work and the score still climbed, that pattern already has a name. A high AI score can land on writing the student actually did. The useful response is the same one a provost would want: keep the drafts, say what the software was asked to do, and be ready to talk through the lines instead of arguing with a single badge.

How to read a story like this without copying the panic

Start with the method. Which files, which dates, which product, which version. The Dartmouth did that part well: before and after ChatGPT, one named detector, a median instead of a single dramatic screenshot. What it cannot supply, and what no detector can supply, is the missing middle, the difference between "I wrote this and then asked for cleaner sentences" and "I asked for the piece."

Then read the author's account as a claim, not as an excuse and not as a confession. Assistive editing is now ordinary in newsrooms, labs, and classrooms. Some policies allow it if the writer checks the facts and keeps the argument. Some policies ban it. A detector does not know the policy. A 96% score does not either. The score only says the surface looks generated, which is what you would expect if a model had been allowed to revise the surface.

If you want to see that surface on your own draft, paste it into DetectGPT and read the highlighted lines rather than the headline number. Fifty words is the floor, 3,000 is the cap, and there is no account. The result is a style estimate. It will not tell you whether a provost, a student, or you opened a chatbot, and it should not be quoted as if it did.

FAQs

Did an AI detector flag Dartmouth's provost?

The Dartmouth reported on October 14, 2026, that Pangram 4.0 scored nine of Provost Santiago Schnell's 2026 pieces at a median of about 96% AI-written. His pre-ChatGPT samples scored as fully human.

Does that mean he did not write them?

No. He says he used AI for refinement, copyediting, clarity, and organization, and that it did not replace his judgment or authorship. The detector cannot see that distinction. It only scores the finished prose.

What is Pangram?

A commercial AI-writing detector. It is not DetectGPT, not GPTZero, and not Turnitin. A percentage on Pangram will not match a percentage on another product.

Why do detectors advertise near-zero error rates and still flag people?

Those rates come from specific test sets. Polished, formal, or non-native English can look like model prose, and light AI editing can look like a full AI draft. A tiny average error rate is not a verdict on one essay.

Are AI detectors biased against non-native English writers?

A 2023 Stanford study found seven detectors flagged 61% of TOEFL essays by non-native writers, and at least one detector flagged 97% of them. That does not prove any single later score is wrong. It does mean tidy English is a known false-alarm pattern.

Will DetectGPT match Pangram's number?

No. DetectGPT is a separate style check. Use it to see which lines look machine-made. Do not treat it as a reprint of Pangram, Turnitin, or any other report.

What should a writer do if a score like this lands on their draft?

Keep dated drafts, be specific about what a tool was allowed to change, and revise the lines you could not defend out loud. The percentage is a map of the prose, not a finding about the person.