A College Provost's Denial Email Also Scored 100% AI. That's the Real Story.

Quick answer

Dartmouth Provost Santiago Schnell spent 2026 warning universities about AI-assisted cheating. This month, student journalists at The Dartmouth ran his own published work through the AI detector Pangram and found a median score of 96% AI-written across nine pieces he published this year, against a clean 0% on everything he wrote before ChatGPT existed. Schnell says he wrote every argument himself and used AI only to tighten phrasing. Then a followup email he sent explaining exactly that also scored 100% AI. Faculty are now petitioning for his firing. Dartmouth's president has opened a formal review. The case that's spreading fastest isn't really about whether Schnell cheated. It's about what happens once a detection score becomes the only tool anyone reaches for, including the person trying to defend themselves with it.

What actually got tested, and what came back

The facts are specific enough to check yourself. The Dartmouth pulled five of Schnell's solo-authored academic papers from 2012 through 2019, before ChatGPT's November 2022 release, and ran them through Pangram. All five came back "100% human-written." Then reporters tested nine pieces he published this year: a Washington Post op-ed on AI and academic integrity, two math papers, and several shorter commentaries. The median score across those nine was 96% AI-written. Four more opinion pieces in the National Catholic Register and Public Discourse scored between 75% and 100%.

Schnell doesn't deny using AI. He says he develops his own arguments, does his own research, and uses ChatGPT for "language refinement, copyediting, improving clarity and organization." He calls that assistive, not authorship. That's a real distinction, and it's exactly the ambiguous middle ground we've written about before: most real-world AI use isn't fully human or fully AI, it's a mix, and mixed text is where today's detectors struggle most. A peer-reviewed study out of Vrije Universiteit Brussel found detectors correctly clear pure human writing almost every time, but can flag a lightly AI-polished human draft as fully AI-generated. Nothing about Schnell's case contradicts that research. It's a textbook example of it.

The part that makes this different from every other flagged-writer story

Reporters asked Schnell for a statement explaining his AI use. He sent one. They ran it through Pangram too. It came back 100% AI-written.

Schnell's explanation was that he wrote a draft himself, then used an AI tool to help with phrasing, took some of its suggested edits, reviewed the result, and signed it. That's the identical process he says produced his flagged op-ed. Which means the detector treated his account of what happened the same way it treated the thing he was giving an account of. If you're trying to prove your writing process by demonstrating it in real time, and the demonstration itself gets flagged, there's no clean way out of that loop using the same tool that put you in it.

This is the trap that a single-number detection score creates when it's the only evidence anyone brings to the table. We've made the same point about court cases before: a detector score is a strong reason to ask a question, never a substitute for asking it. Chris Callison-Burch, the University of Pennsylvania researcher who studies these tools, put it plainly in a September Communications of the ACM feature about a different case entirely: the problem isn't usually that the score was wrong, it's that nobody had a conversation before treating the score as a verdict. Schnell's case adds a wrinkle that court cases don't have. Here, the accused party tried to have that conversation, in writing, and the writing itself got flagged.

Why this escalated past a fact-check into a firing debate

None of this would be a governance crisis if Schnell weren't the person who published a Washington Post essay in August arguing that universities need to police the line between what students can do independently and what they lean on AI to do. Dartmouth's own chapter of the American Association of University Professors is now circulating a petition asking how students can be held to a standard the provost himself apparently didn't meet. A student op-ed writer called for his firing directly. Dartmouth President Sian Beilock announced an independent faculty-and-alumni committee to review Schnell's AI use "measured against the criteria of the publications to which he submitted his work," while acknowledging the broader problem goes beyond any one person.

That's a genuinely different kind of consequence than a professor rerunning a paper through a detector before grading it. Schnell isn't a student facing an academic-integrity charge. He's a university's chief academic officer, and the review process forming around him looks more like a governance inquiry than an honor-code case, weighing publication disclosure policies against actual practice rather than adjudicating one flagged essay in isolation.

What this means if you publish under your own name

Most people reading this aren't provosts, but plenty write under a real byline: a newsletter, a company blog, a LinkedIn post, an op-ed pitch. Schnell's situation is a preview of a problem that's about to get more common, not less. We've covered how newsrooms are already running detection passes on submitted opinion pieces, and the gap between different outlets' thresholds for what counts as acceptable AI assistance is wide and inconsistent. If your writing process involves any AI-assisted editing, a published disclosure policy at the outlet you're submitting to matters more than your intent. Two of the four journals that published Schnell's work this year require AI-use disclosure. He didn't disclose on two of them, and those two are the ones that scored highest.

The practical lesson isn't "stop using AI to edit." It's "know what the venue you're publishing in actually requires, and check your own draft against a detector before an editor or a curious reader does it for you." A single score won't settle the question of how much AI touched your text. It will tell you whether the finished piece reads the way heavily AI-polished text tends to read, which is useful information to have before publication, not after a student newspaper has it first.

Run your own draft through the check first

If you write under your own name for any outlet with an AI disclosure policy, or you're not sure how AI-assisted editing will read to a detector, run the draft through our free detector before you submit it. You'll get a sentence-level breakdown instead of one headline percentage, which is the difference between catching a problem early and finding out about it from a reporter.

FAQ

Did Pangram falsely flag the provost's writing? There's no evidence of that specifically. What the case shows is the well-documented gap between "AI-generated" and "AI-edited," a distinction current detectors are worse at than the simpler human-versus-fully-AI question. Schnell's pre-2022 work scored cleanly. His account of a human-drafted, AI-polished process is plausible and matches known detector behavior on that kind of text. Whether his account is accurate is a separate question from whether the detector works.

Why did his denial email also score as AI-written? Because he says he used the same drafting process to write it: his own thinking first, then AI help with phrasing before sending it. A detector measures patterns in finished text. It doesn't know whether that text is an op-ed or an explanation of an op-ed. If the underlying process is the same, the score can look the same too, which is exactly why a score alone can't resolve a dispute about process.

Should a single detection score decide whether someone gets disciplined at work? No. Every case that's reached a court or a formal review so far has turned on the process around the score, not the score's accuracy. Dartmouth's response reflects that: a review committee weighing disclosure policy and practice, not a single Pangram result treated as a verdict.