On August 9, 2026, Dartmouth College provost Santiago Schnell wrote a Washington Post op-ed titled “Universities are fighting AI cheating. But there’s a deeper problem.” This week he discovered firsthand how deep the problem was: he resigned after an advisory committee found that he had failed to disclose significant use of artificial intelligence in an academic paper and in the same op-ed, even though publishers required such disclosure.

The case quickly became a lesson in how not to build AI governance at a university. The committee used the term “insufficient cooperation”: Schnell handed over his Dartmouth-issued laptop, but not logs from his personal LLM accounts related to the writings under review; he also declined to meet with the committee. The concrete evidence came from the Claude enterprise account Schnell used through the college. There, the LLM appears as a tool for revising papers, finding and fixing bugs in code, and editing or generating figures.

The first structural implication concerns the forensic value of enterprise accounts. An organization that provides a corporate license has access to administrative logs and can reconstruct interactions. That is precisely what allowed the committee to proceed. But the investigation stopped at the boundary of personal devices and personal accounts. If a researcher uses an LLM with private credentials, the institution sees nothing unless disclosure policies are enforced with technical tools. That is the core problem: governance based on goodwill does not hold when the person who must comply is also the one preaching it.

There is also a second-order effect on publishing policy and the broader market. Publishers that already require AI disclosure may tighten the rules, moving from self-certification to checks on logs and histories. Universities, for their part, may push for mandatory institutional accounts and periodic audits. But that path risks producing shadow AI: if oversight becomes too invasive, anyone who wants to avoid it can simply use personal tools or services not tracked by the employer. The result is an incentive to formalize only visible use, leaving the rest in the dark.

For those evaluating self-hosted or on-premise deployments, the case signals a subtle but important difference. In a cloud enterprise context, the vendor retains the logs and can make them available to the organization, as happened at Dartmouth. In a self-hosted environment, the organization can manage logs locally with greater data sovereignty, but it still must enforce audit policies and cannot prevent people from using external services with personal accounts. An audit trail is not an automatic effect of either cloud or on-premise: it is a control choice. For those assessing these trade-offs, AI-RADAR offers analytical frameworks at /llm-onpremise to compare constraints and incentives.

The heaviest detail, in the end, is not that Schnell used Claude to fix bugs or refine figures. It is that the committee had to reconstruct everything from a corporate account while the person involved did not cooperate. If the future of AI governance in institutions depends on access to logs, the Dartmouth case raises an uncomfortable question: how many uses remain invisible simply because they happen outside official channels?