Dartmouth's provost warned us about AI cheating—but his own AI use cost him his job

Dartmouth College Provost Santiago Schnell resigned this week after an advisory committee concluded he had not disclosed significant use of AI in an academic paper and in a Washington Post opinion piece, even though the publishers involved required such disclosure. The resignation follows an op-ed Schnell published on August 9, 2026, titled "Universities are fighting AI cheating. But there's a deeper problem." The committee's report describes his cooperation as insufficient: he handed over his Dartmouth laptop but declined to provide logs from personal large language model accounts tied to the writings in question and did not meet with the committee.
Investigators did obtain one substantial record: the Claude enterprise account Schnell used through Dartmouth. Those logs show him deploying the model across several kinds of work, including revising papers, locating and fixing bugs in code, and editing or generating figures. In one exchange he asked Claude to act as a referee for a paper and its supplementary materials and repository before submission to ChemSystemsChem, framing the request around expertise in nonlinear chemical dynamics and mathematical chemical kinetics. In another he asked the model to make corrections to TeX files and a repository, generate new figures where needed, and report back on what changed.
Schnell has consistently characterized his use as editorial rather than generative at the level of whole documents. He told Semafor he had used chatbots to refine a few arguments and check grammar, and told a journal editor he used AI as an editorial tool to sharpen sentences, clarity and the expression of arguments. He said he had not disclosed the use to an academic journal because he did not understand that this kind of AI-assisted editing required disclosure.
That account sits awkwardly against parts of the record. The committee wrote that Schnell's public statements may understate the nature of his actual AI use and that he regularly went beyond his own public descriptions. One exchange cited in the report is likely to draw particular attention: Schnell uploaded a manuscript of a forthcoming article, "Academic Freedom and Catholic Mission: The Procedural Overlap Zone in University Governance," and asked Claude whether there were many signatures of AI writing in it, then later submitted a revised manuscript incorporating changes the model suggested. Read narrowly, that is a style check; read less charitably, it is an attempt to scrub detectable traces. The report does not resolve which interpretation is correct.
Schnell has also written explicitly about how he works with models. In a September 2026 paper on AI detection called "Who Wrote This?", he describes asking OpenAI's GPT-5.6 Pro whether an AI-written essay could ever be identified with certainty, and says the model proposed structures and titles, identified candidate literature and generated portions of early drafts. He says he chose the problem, journal, audience and argument, checked citations, corrected errors and rewrote most sections, with occasional language editing from GPT-5.6 Sol, and had GPT-5.6 Pro and Anthropic's Claude Fable 5.1 review the manuscript before submission. He argues against treating AI as a co-author, drawing an analogy to mathematical software that contributes substantively without becoming an author. The paper is, in effect, a disclosure document; the committee's finding is that comparable disclosure was missing elsewhere.
Dartmouth's response has centered on transparency rather than retreat. President Sian Leah Beilock, who wrote a September Atlantic piece titled "Why I Want More AI at Dartmouth," said in announcing the resignation that leaders must model the transparency and human judgment expected of students and scholars. The college frames itself as the birthplace of the field, citing 1956 as the year the term was coined and the foundational question posed. It is now hosting a conference called "What Must Remain Human?" with a student hackathon sponsored by Anthropic and speakers including Mira Murati and Yann LeCun.
The surrounding student picture is mixed. Survey data presented at the conference indicates 61 percent of college students think AI helps explain difficult concepts and half believe they do better work after talking through problems with AI tools, while 70 percent say take-home work has become unfair because of AI and 39 percent feel pressured to use it because peers do. Among Dartmouth students, 92 percent worry AI can harm writing skills, and many believe their peers use far more AI than they actually do.
For developers, the case is a practical lesson in provenance. Enterprise accounts produce audit trails that personal accounts do not, and an institution's willingness to inspect them turns vague assurances into a documented record. This suggests teams building writing or research tools should expect disclosure requirements to attach to editing, figure generation and code repair, not just to drafted prose — the categories that most people instinctively treat as ordinary tooling. A likely trade-off is that fine-grained disclosure of every prompt is impractical, which pushes institutions toward structured declarations like the one Schnell wrote for his detection paper. The unresolved question is where the line sits between an editorial aid and undisclosed authorship; the committee's report indicates it could not determine the full scope of his use, which is itself a caution about relying on voluntary accounts.
Why it matters: Academic and corporate policies are converging on disclosure as the enforcement mechanism for AI use, and this resignation shows that the standard can reach senior administrators, not just students. Developers building assistants for writing, code and figures should assume prompts, uploads and revisions may be audited through managed accounts, and product teams should design lightweight ways to record and declare assistance. My inference is that institutions will lean harder on enterprise tooling with logging rather than on honor-system statements, since the Dartmouth case turned on exactly that gap.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
People who love AI don't always love telling others they used it.