Dartmouth Just Authorized an AI Detector for Grading. Its Own Rules Show Why That's Harder Than It Sounds

Quick answer

On September 29, 2026, Dartmouth's interim dean of faculty sent every undergraduate instructor a seven-point policy authorizing one AI detector, Pangram, for reviewing student writing. It is one of the first explicit, campus-wide "here is the approved tool and here is how you may use it" policies to surface from a major university, rather than a blanket ban or a quiet, ungoverned free-for-all. The policy itself is a useful document regardless of which school you're at: it names a specific legal risk (FERPA), requires de-identifying student work before it goes anywhere, bars a detector score from being the sole basis for a grade penalty, and leaves the final call to a human committee. It arrived three weeks after Dartmouth's own provost became a national story for the opposite problem: his writing scored as AI-generated on the same kind of tool the policy now governs.

What the policy actually says

Interim dean of faculty John Carey's September 29 email, reported first by The Dartmouth, attached a document with seven numbered guidelines. The core mechanics:

Faculty may submit student writing to AI detection software, but Pangram is the only tool authorized, "because it has shown the most reliability of available detectors." Before anything gets submitted, instructors must strip the work of every personally identifying detail: names, family or address information, Social Security or student ID numbers, and even metadata buried in a Word or PDF file's author or commenting fields. The policy is explicit that this covers content, not just names, barring submission of any personal narrative or account that could let a reasonable person identify the student even after the obvious fields are scrubbed.

Students have to be told in writing, before any of this happens, whether their instructor may run their work through a detector and whether the result could affect their grade. And a flagged score alone changes nothing: faculty "may not use the results of an assessment from AI detection software as a basis for penalizing or otherwise affecting a student's grade without first communicating with the student about the quality and integrity of their work." When a case does escalate to Dartmouth's Committee on Standards, the committee is instructed to weigh "a constellation of data" beyond the detector's number, not the number by itself. The college is still deciding whether to buy an institutional subscription; for now, individual faculty members are using the tool on their own.

Why this needed its own policy: FERPA is not optional

The reason a university writes seven guidelines instead of just naming a tool and moving on is the same reason most institutions still don't have anything like this. The Family Educational Rights and Privacy Act of 1974 treats student coursework as a protected education record once a school holds it. Routing that record to an outside AI detector is a data disclosure, and FERPA puts real conditions on disclosure.

UC Santa Cruz got there first. Years before Dartmouth's policy, its campus FAQ on generative AI laid out three narrow conditions for running an AI detector on student work: the tool keeps data inside systems the university controls, the tool is contracted through the campus purchasing office with student-data protections built in, or the student agrees up front. Absent one of those three, the guidance is blunt: "Student work should never be submitted to web-based detection tools, such as GPTZero and other tools, unless one of the following three criteria is satisfied." A 2025 National Education Association policy brief on federal AI regulation treats that UC Santa Cruz language as the clearest statement anywhere in the sector of a risk most schools still haven't written down: "using a program to detect AI usage may require students' work to be processed through an outside third party, which may be a violation of FERPA."

Dartmouth's de-identification rule answers that exposure directly. Pull the names, IDs, and file metadata out before anything leaves campus, and the FERPA disclosure problem shrinks fast. Any school without a written policy can read this template in five minutes and start using it this week, regardless of which detector it picks.

The irony sitting right next to the policy

These guidelines didn't appear out of nowhere. Student journalists at The Dartmouth broke the story on September 21: Provost Santiago Schnell's 2026 academic papers and opinion columns scored a median 96% AI-written on Pangram, against a clean 0% on everything he published before ChatGPT existed, and we covered what happened next in our earlier piece on his denial email also scoring 100% AI. Schnell maintains he wrote every argument himself and used AI only to polish the language, a distinction a style-pattern detector genuinely cannot separate from full AI drafting, and the story ran in The Chronicle of Higher Education, The Boston Globe, and The New York Times before Dartmouth's president opened a formal review. Read guideline six against that backdrop and it lands differently, because the same administration now sitting in the middle of a live, high-profile dispute over what a Pangram score proves wrote the rule insisting a score never stands alone. Call it the "constellation of data" clause if you want the official term; we've made the identical argument from the opposite direction before, that a single AI-detector score should never be the whole case, and sophomore or provost, the rule doesn't change.

What this means if your school doesn't have a policy yet

Most schools fall into one of two camps. One bans detector use outright, citing the same false-positive worries fueling this whole debate. The other has written nothing down at all, so individual faculty make their own FERPA calls, upload by upload, with no institutional backing if a case goes sideways. Dartmouth chose a third path. Name one approved tool. Require de-identification before submission. Put disclosure to students in writing, ahead of time. Never let a score alone justify a grade penalty. Send disputed cases to a body that looks at more than a single number. A department doesn't need an enterprise contract to start any of this; it can run the same five steps this semester with whatever detector it already has, and settle the vendor question later. The discipline is the real export here, not the brand on the license.

Carey was explicit that none of it is mandatory. He told The Dartmouth that faculty "are not required" to use Pangram, full stop. The rules exist for the instructors who opt in anyway, so they know what comes first: tell students in writing, strip identifying details before anything gets submitted, and don't treat one flagged score as a verdict. His email also connects the timing directly to the Schnell story, calling it what "became a central topic of conversation on our campus" in the weeks before the guidelines went out. A live example of a score's limits sat right next to the new rules for using one.

A detector's job in this workflow is narrow: flag writing that merits a conversation. Nothing more. TheChecker.AI's free detector gives schools that same first signal, down to the sentence level, meant as a starting point for a conversation, not its conclusion. Pair it with a written policy like Dartmouth's, and the signal stops being the whole decision on its own.