Nine pieces of professional writing, one detector, a median score of 96%. That is what the student newspaper The Dartmouth found when it ran Pangram, a commercial AI-text classifier, against everything Santiago Schnell, Dartmouth’s provost, had published in 2026: opinion columns, a film review, two academic papers. Pangram scored five of his pre-2022 works, from before ChatGPT existed as a product, at 100% human-written. On October 28, Schnell resigned.
The proximate cause was an August 9 op-ed in The Washington Post, titled ‘Universities are fighting AI cheating. But there’s a deeper problem,’ which argued that a university degree should distinguish independent student work from AI-assisted work. Semafor reported that Pangram scored the piece 100% AI-written. Semafor’s sample was 310 guest columns published across The Wall Street Journal, The Washington Post and The New York Times over one month; using Pangram’s own threshold of 80%-plus AI content to call a piece AI-written, 10 of the 310 crossed it and another 40 scored as partly AI-generated. Schnell’s was one of the 10, and the only one by a university provost writing about AI cheating.
Schnell’s own account, given to Semafor at the time, was that he wrote the essay, then ‘used ChatGPT to refine a few arguments, check for grammatical errors, and submit it to the Washington Post,’ and that it ‘went through multiple rounds of edits’ there. The Post’s policy requires guest writers to confirm a submission was ‘not created or manipulated with artificial intelligence or editing software’; a spokesperson pointed to that policy but did not comment on the other 17 Post columns Pangram flagged as partly or wholly AI-written in the same sample. Schnell also argued the detector carries bias against writers like him: a native of Venezuela with a doctorate in mathematical biology, he said AI detectors ‘are biased against non-native English speakers, such as myself.’
That defense runs into a body of published testing on Pangram specifically. A September 2025 NBER working paper by Brian Jabarian and Alex Imas found near-zero false-positive and false-negative rates for Pangram across genres, lengths and models, including text run through humanizer software meant to evade detection. A University of Chicago Booth study found Pangram holds false-positive rates at or below 1% on academic writing while an open-source baseline detector flagged 30% to 69% of human-written text as AI. Pangram’s own published figure is a false-positive rate of 0.0041%, about 1 in 24,000, on its current model version. None of that rules out individual errors — Pangram’s co-founder has acknowledged specific false positives in other cases — but it is a different claim from Schnell’s general one about bias against non-native speakers, and the company’s benchmarks don’t show elevated error rates on that basis.
What the committee actually checked
Dartmouth’s investigation did not rest on detector scores. President Sian Leah Beilock appointed an independent committee, led by former provost and computer-science professor David Kotz, which by its own account did not use AI-detection software at all. Instead it obtained logs from Schnell’s account on Claude, Anthropic’s chatbot, which Dartmouth provides to faculty and staff as an enterprise tool, plus his university email. Schnell handed over his Dartmouth laptop but, per the committee’s report, did not provide logs from any personal LLM accounts he may have used, and the committee said he ‘failed to make himself available at the time requested’ for an interview.
The logs it did get showed, by the committee’s description, that Schnell used Claude to ‘substantially edit his research manuscript, troubleshoot errors in computer code, and generate figures’ for at least one journal paper that required disclosure of that level of assistance and did not receive it. Logged prompts included asking Claude whether it saw ‘any scientific issue with the paper’ and whether a ChatGPT-drafted reviewer response needed changes; for a separate philosophy-of-governance piece, he asked Claude to check the draft for ‘signatures of AI writing’ and then revised it using the tool’s suggestions. The committee’s eight-page report, delivered October 27, concluded that Schnell’s own description of his process — language refinement, copy editing, clarity — ‘understates the nature of the Provost’s actual AI use’ as shown in the logs, and found his conduct inconsistent with the university’s Code of Ethical Conduct and Academic Honor Principle. It identified AI use requiring disclosure in two further publications beyond the two he’d already flagged, and cleared a fourth, co-authored paper.
The paper about the problem
One of the pieces the committee reviewed was itself about this exact question. In a September 2026 paper titled Who Wrote This?, which Pangram scored 99% AI-written, Schnell argued that sorting text into human-written and AI-written is the wrong frame entirely, because all writing is becoming some mixture of the two, and what should matter is whether the named author takes full responsibility for the result. He laid out his process in detail: he asked GPT-5.6 Pro to connect AI-detection to Turing’s imitation game and to statistical decision theory, and the model proposed structures, titles and candidate literature and drafted sections. Schnell says he chose the problem, the journal, the direction of the argument, decided which results to keep — including a decision to retain a study by one researcher, Liang, while weighing it against counterevidence — checked citations, fixed errors, reorganized the argument, and had GPT-5.6 Pro and Claude review the finished draft for flaws reviewers might catch.
Calling [the paper] ‘coauthored’ would introduce a different confusion. Mathematical tools have long made substantive contributions without thereby becoming authors… The point is not that the software contributed nothing. It is that contribution and authorship are different roles.
The argument is not obviously wrong as a matter of philosophy. A calculator that does the arithmetic in a proof is not a coauthor of the proof. But Schnell’s own paper was a workflow diagram without a step for him writing, and he declined to share the underlying chat logs that would let anyone outside the process check how much of the judgment he describes actually happened before the AI output existed versus after. That is the gap the committee’s log review was built to close, and on at least one paper it says the logs show more AI involvement than Schnell disclosed.
A faculty revolt, a resignation, an unresolved standard
Student and faculty anger built through September and October: Dartmouth’s AAUP chapter circulated a letter calling for Schnell’s removal, and a faculty no-confidence vote was reportedly scheduled for roughly two weeks after the committee’s report landed. Schnell resigned before it happened. Beilock’s campus memo said he was leaving ‘out of respect for his colleagues and the institution,’ while reaffirming her own position, laid out in a September Atlantic essay, that universities ‘guarantee their own irrelevance’ if they don’t adopt AI tools. History professor Matthew Delmont became interim provost through June 2028. The resignation lands in the same year Dartmouth marked the 70th anniversary of the 1956 workshop where the term ‘artificial intelligence’ was coined — an irony the campus did not let pass quietly, according to The Dartmouth’s coverage of the debate the investigation set off.
Not everyone treats the outcome as proportionate. Writing in Semafor, Reed Albergotti — whose own reporting started the story — argued Dartmouth overreacted: his broader survey had already found AI-generated text in roughly 50 guest columns across major papers in a single month, making Schnell’s conduct ordinary rather than exceptional, and Schnell’s stated process (choosing ideas, doing the research, deciding structure, drafting, deciding final wording) describes using AI as an execution tool rather than outsourcing the thinking. A Dartmouth student columnist, Eli Moyse, took the opposite view in the same piece, calling the conduct ‘the very worst academic violation a faculty member can commit’ because Schnell repeatedly presented AI-assisted text as his own original work.
What’s missing from both the university’s policies and the public argument is a number: no one has specified what fraction of a text, or what kind of AI involvement, crosses from assistance that needs no disclosure to use substantial enough to require it. Schnell’s own distinction — that tools get credited only when they play ‘a critical role in the scientific process’ — is exactly as vague as the policies he’s accused of violating. Dartmouth’s committee resolved one case by reading chat logs line by line, which doesn’t scale to a university’s worth of op-eds and papers every year. The next test is whether Beilock’s promised standard of ‘transparency and human judgment’ comes with a threshold anyone could check without subpoenaing a Claude account.