Latest / MBA Training AI / Enterprise AI chat logs: how Dartmouth's provost was caught

Transcript

Enterprise AI chat logs: how Dartmouth's provost was caught · MBA Training AI

• 0:05
  Host: MBA Training. Today: Dartmouth provost resigns over AI use, and the chat logs did it.
• 0:12
  Host: He asked Claude whether his own article had signatures of AI writing.
• 0:16
  Expert: That's straight out of the committee report. He uploaded a manuscript and prompted Claude: "Are there many signatures of AI writing in this article?"
• 0:26
  Host: Oh, come on.
• 0:27
  Expert: And then, according to the report, he submitted a revised manuscript that incorporated the kinds of changes Claude suggested.
• 0:35
  Host: So the AI wrote it, and then the AI helped hide it.
• 0:40
  Expert: Careful. The report says the revisions followed Claude's suggestions. It doesn't say Claude wrote the article.
• 0:47
  Host: Fine, my imaginary lawyer thanks you. Back up for anyone who missed this.
• 0:53
  Expert: Santiago Schnell was provost of Dartmouth, meaning the chief academic officer, second in command at the college. He resigned Wednesday after a faculty committee found he'd used AI in published work without disclosing it where publishers required it, and that he wasn't forthcoming about it. The Boston Globe says he's leaving the provost job and his professorship too.
• 1:17
  Host: And it all started with an opinion column.
• 1:20
  Expert: In the Washington Post, in August. The topic was how universities should handle the ethics of AI.
• 1:27
  Host: Of course it was.
• 1:29
  Expert: Semafor ran it through Pangram, an AI detector, meaning a model trained to tell human prose from machine prose. Inside Higher Ed reports it flagged at least 80 percent of the column as AI-generated.
• 1:43
  Host: Wait, what? The man in charge of academic standards...
• 1:47
  Expert: ...outsourced the column about academic standards. Yes. And his defence was that AI never replaced his research, his judgment or his authorship.
• 1:57
  Host: Everybody says that. I say that about spellcheck.
• 2:00
  Expert: Then the student paper, The Dartmouth, ran a batch of his pieces from this year through the same tool. Inside Higher Ed says it came back 96 percent AI-written.
• 2:12
  Host: So the students caught the provost. I enjoy that more than I should.
• 2:16
  Expert: This is the part that matters for anyone with a corporate AI account. The committee didn't use detectors at all.
• 2:23
  Host: Then what did they use?
• 2:25
  Expert: His chat logs. It reviewed Schnell's logs from Dartmouth's enterprise AI tool, Claude, along with his emails. Enterprise here means the licensed version the college gives its staff.
• 2:37
  Host: Hang on, say that again.
• 2:39
  Expert: The evidence came from the work account. Not a detector's hunch. The receipts.
• 2:45
  Host: So everyone listening who has Copilot or ChatGPT Enterprise at work...
• 2:49
  Expert: ...has a written record of every draft they asked for help with, sitting on a server their employer controls. Yes.
• 2:57
  Host: That's a much bigger story than one provost.
• 3:01
  Expert: And some of the logs are rough. The Dartmouth quotes him telling Claude he'd had ChatGPT write a version of a response, then asking: "Based on this version would you change anything in your version?"
• 3:14
  Host: Two chatbots doing a group project, and the human is the stapler.
• 3:19
  Expert: Then in October Schnell disclosed his ChatGPT use in an email to ChemSystemsChem, but not his use of Claude.
• 3:28
  Host: That's owning up to the parking ticket while you're driving the getaway car.
• 3:32
  Expert: On that chemistry paper, the committee found he used Claude to substantially edit his research manuscript, troubleshoot errors in computer code, and generate figures, after peer reviewers (the outside experts who vet a paper) asked for major revisions. The journal required that kind of use to be disclosed. He didn't disclose it.
• 3:54
  Host: And his excuse?
• 3:56
  Expert: He told the president he'd relied on older author instructions. Valley News says the publisher has required disclosure since at least 2023. His own line, according to the report: "This wasn't the case, but a disclosure would have been the best approach."
• 4:12
  Host: The dog ate my author guidelines.
• 4:15
  Expert: More or less.
• 4:16
  Host: Here's where I land: he had to go. Not because he used AI. Because he preached the sermon and skipped church.
• 4:24
  Expert: I'll push back, because his side has a real argument. It also noted that it was not recommending any disciplinary action. That's the committee, per Valley News.
• 4:35
  Host: Then why is he gone?
• 4:37
  Expert: Beilock, the president, says he chose to leave. And Inside Higher Ed points out the resignation came less than two weeks before a faculty no-confidence vote, which is a formal statement that the faculty no longer trust him.
• 4:51
  Host: So he jumped before he was pushed.
• 4:54
  Expert: Maybe. But look at the institution around him. Beilock has called for "more AI" at the university. The college runs an enterprise Claude deal. And the Globe reports faculty and students say that these public messages have come with murky policies.
• 5:11
  Host: Oh, come on. The Post's policy wasn't murky. The committee called breaking it a clear violation.
• 5:18
  Expert: For the Post, sure. But the rules change from venue to venue. One outlet bans AI-made submissions, a journal wants disclosure, and the report suggests the Catholic higher education journal does not have a required disclosure at all.
• 5:33
  Host: So the patchwork is the problem, not the guy.
• 5:36
  Expert: The patchwork is what made skipping disclosure feel like no big deal. He's still on the hook for reading the rules. Both are true.
• 5:45
  Host: I don't buy it. A provost oversees an honor code that makes students disclose their AI use. He knows what disclosure means.
• 5:53
  Expert: And most professionals right now are polishing emails, decks and reports with these tools under rules nobody has written down. In most companies, nobody has said where polishing stops and authorship starts.
• 6:07
  Host: Polishing is fine. Hiding it is the offence.
• 6:10
  Expert: Hiding from whom, under which rule, though?
• 6:14
  Host: We're not going to settle this.
• 6:16
  Expert: No, we're not.
• 6:17
  Host: Honestly? This one makes me a bit queasy. He's a scientist with a real career, and what people will remember is him asking a chatbot whether he sounds like a chatbot.
• 6:28
  Expert: Same. Nobody got hurt by a badly disclosed chemistry figure, and he still lost his job and his university.
• 6:36
  Host: Okay, back to it. Pangram started all this, and detectors have a terrible reputation.
• 6:42
  Expert: They do, and the committee stepped around that on purpose. The committee "did not use Pangram or other AI tools when determining whether and to what extent Provost Schnell used AI". It looked at what he actually did, not at what a model guessed about his prose. That's stronger evidence, and a more uncomfortable precedent.
• 7:04
  Host: Uncomfortable how.
• 7:06
  Expert: Picture the Monday meeting where human resources is looking into a disputed report, and someone realises the enterprise AI logs are just sitting there with timestamps. You can argue with a detector. It's a lot harder to argue with your own prompts.
• 7:22
  Host: Great, so my next performance review is a transcript of me asking for a "punchier intro."
• 7:28
  Expert: Which you ask for constantly.
• 7:31
  Host: Not on the record, I don't.
• 7:33
  Expert: The committee also admitted its limits. He'd published at least 30 pieces since joining Dartmouth, and given the short timeline, the committee focused on a handful of them. The Dartmouth reports the committee also found his cooperation fell short.
• 7:49
  Host: So they only read part of the record, and he still resigned.
• 7:53
  Expert: Part of it, but the part they read was bad enough.
• 7:56
  Host: What are you watching next?
• 7:58
  Expert: Publishers, first. If every journal and newspaper keeps writing its own disclosure rules, people will keep falling through the gaps. If they agree on a shared standard, cases like this get simpler. Then employers: whether enterprise AI logs become routine evidence in misconduct cases, and whether anyone tells staff that up front.
• 8:21
  Host: And Dartmouth itself.
• 8:22
  Expert: Beilock wrote that leaders must model the transparency and human judgment that we expect from our students and scholars. Now the college has to write rules that live up to that sentence.
• 8:34
  Host: My open question, then. If the logs are the evidence, the line on disclosure gets drawn by whoever reads the logs, not whoever writes the policy. Is that a fairer world or a creepier one?
• 8:46
  Expert: I don't know. Let's ask Claude whether this episode has signatures of AI writing.
• 8:52
  Host: It'll say no.
• 8:54
  Host: Sources: Valley News, Boston Globe and Inside Higher Ed. Every link is in the show notes.
• 9:01
  Host: Done. Want to know where you actually stand? Take the AI self-assessment at mba-training.com.