The detector started it. The logs ended it
The public case against Santiago Schnell began with a number from a detector. In August, Semafor reported that Pangram had flagged his Washington Post column as at least 80 percent AI generated. The Dartmouth later reported that Pangram rated nine of Schnell's professional articles from that year as 96% "AI-written." That was the newspaper's reported detector result, not a finding by Dartmouth's advisory committee. Schnell admitted using AI, and said it did not substitute for his research, judgment or authorship.
Dartmouth's committee declined to build on any of that. Its report opens by stating it did not use Pangram or other AI tools to decide how much he had used AI, and did not use AI to draft the report either. Instead it did the slower work: it read the AI policies of the publications involved, at the dates that mattered, then compared them with his emails and with the records of what he actually asked Dartmouth's enterprise Claude to do. The findings did not depend on whether a detector could classify his prose correctly. The committee compared the available Claude logs and communications with the publishers' policies in force at the relevant dates. That allowed it to assess specific uses of AI against specific disclosure requirements, rather than infer authorship from a percentage score.
Two cases the record settles
The first is a research paper, "Reaction Acceleration," published in ChemSystemsChem on September 16, 2026. After peer reviewers demanded major revisions, the Claude logs show Schnell asking the model to substantially edit the manuscript, troubleshoot code, generate figures, act as a referee on the paper, and check a response to reviewers that ChatGPT had drafted. The publisher's rules, at submission, acceptance and publication, required disclosure of AI use of that kind. No disclosure appeared. Schnell told the president he had relied on pre-2025 author instructions. The committee checked the archived policy and found disclosure had been required since at least 2023.
His attempt to correct the record then became part of the record. On October 6 he emailed the journal disclosing ChatGPT's help with language, code review and consistency checks. He did not mention Claude, the tool the logs show doing the heavier work.
The second case needs no archive. His August 9 Washington Post op-ed, about universities fighting AI cheating, was submitted under a policy stating that writers agreed their content was not created or manipulated with artificial intelligence. He had already told Semafor he used ChatGPT to refine arguments in it and check grammar. The committee called that a clear violation of the paper's stated policy.
What the report does not settle
Read the limits before treating this as a verdict on all 30-plus pieces. The committee says a complete review was impossible in the time available, so it targeted writings where a violation could be readily ascertained: sole or lead authorship, a publisher policy barring AI or requiring disclosure, an available record, and a missing or incomplete disclosure. That selection method finds violations efficiently. It cannot tell you how typical they were of the rest of his output, and the committee does not claim otherwise. The committee also examined a co-authored paper in which Schnell ran code through Claude and ChatGPT. It could not reach a conclusion either way because the publisher's rules at the relevant dates did not clearly require disclosure of that use.
The evidence was also incomplete in a way that cuts in one direction only. Schnell handed over his university laptop but not the logs from the personal AI accounts he used for the same writing, and he did not meet the committee at the time it requested. The findings therefore rest on the institutional account, the one place a university can actually look.
And the resignation was not a sentence. The committee explicitly was not running a disciplinary process and recommended no disciplinary action. Beilock's announcement says the decision to leave was his. Inside Higher Ed adds the political fact that sits next to that statement: he was less than two weeks from a faculty no-confidence vote.
The rule other institutions can copy
Dartmouth calls itself the birthplace of AI, and this week it hosts a conference marking 70 years since the workshop that launched the field. Beilock's closing point was not that AI use is the offence. It was transparency: leaders should model the disclosure they expect from students and scholars.
That is the transferable part. A disclosure regime that depends on detectors will keep producing arguments about percentages, and detectors are contestable in exactly the cases that matter most. A regime that depends on records is different. If a university gives staff an enterprise AI account, the logs exist, the publisher policies are dated documents, and the question becomes a factual one: what did you ask the tool to do, and what did the venue require you to say about it? Schnell's case illustrates how an institution can assess AI use against documented publication policies and available account records without relying on a detector. Other institutions considering similar reviews will have to decide what records they can access, what policies apply, and how to handle gaps in the evidence.
Sources
- [1] Dartmouth College, Office of the President, "An Update on the Provost Review," October 28, 2026 — resignation, the committee's conclusion as quoted by the president, "This was his decision," interim provost Matthew Delmont through June 2028, AI at 70 conference contextRead source
- [2] Advisory Committee Report to President Sian Leah Beilock, October 27, 2026 (Dartmouth, PDF) — primary source for: no detector or AI used, method and scope limits (at least 30 writings, targeted review), the ChemSystemsChem and Washington Post findings, the disclosure-required-since-2023 finding, the October 6 email omitting Claude, the forthcoming-article logs, the inconclusive co-authored case, the missing personal-account logs, and the absence of any disciplinary recommendationRead source
- [3] Inside Higher Ed, "Dartmouth Provost Resigns Over AI Use," October 28, 2026 — the Pangram figures as reported by Semafor (at least 80 percent on the Post column) and The Dartmouth (96 percent across nine articles), Schnell's "didn't substitute" defence, and the pending faculty no-confidence voteRead source