Dartmouth Just Authorized an AI Detector. Its Own Rules Say the Score Is Not the Case.

The AI detector that flagged Dartmouth’s provost is now the only one Dartmouth faculty may run on student work. Its own guidelines say the score alone is never enough for a finding.

Read that twice, because most of the coverage this week read it once.

On Tuesday, September 29, the interim dean of the faculty of Arts and Sciences, John Carey, sent Dartmouth’s undergraduate faculty a document of “guidelines on use of AI detection in teaching.” The student newspaper, The Dartmouth, obtained it and reported its contents on October 1. The headline version is simple. Professors may now submit student writing to an AI detector, and Pangram is the one detector authorized, “because it has shown the most reliability of available detectors.”

The headline version is also the least useful part. Most of the seven guidelines describe what the professor has to do, not what the detector does. And one of them quietly changes what a detector score is worth.

How Dartmouth got here in five weeks

The order of events matters, so here it is plainly.

On August 9, Provost Santiago Schnell published a guest column in The Washington Post arguing that “a degree should distinguish what students can do independently from what they can accomplish with AI.” On August 26, a Semafor analysis ran recent guest columns in three national newspapers through Pangram and identified Schnell’s column as “100% AI-written.” On September 21, The Dartmouth published its own investigation. Across nine pieces Schnell published in 2026, Pangram returned a median “AI-written” score of 96 percent, while the work it tested from before 2022 came back “100% human-written.”

Schnell’s response is worth reading closely, because your students are going to say a version of it to you. He told The Dartmouth that he uses AI for “language refinement, copyediting, improving clarity and organization,” that he develops the ideas and decides what appears in the final text, and that he takes responsibility for it. A few days later he wrote in the same paper, “In retrospect, I now realize I should always explicitly disclose use of AI in my public writing.”

Dartmouth’s chapter of the American Association of University Professors circulated a petition asking for his removal as provost. On September 25, President Sian Leah Beilock announced an independent advisory committee to review his AI use, “measured against the criteria of the publications to which he submitted his work.” She also wrote that “we need role-appropriate rules and policies for this technology.”

Four days later, the faculty got the detector guidelines.

What Dartmouth’s AI detector guidelines actually ask of you

Set the permission aside for a moment and look at the obligations, because they land on the instructor.

You tell students first. Faculty “must inform” students of the course AI policy, including whether the professor may use AI detectors and whether the “results of such assessments could affect grades.”

You protect the student’s identity. Student work has to be de-identified before it goes to a detector, in line with the Family Educational Rights and Privacy Act.

You talk to the student before the score touches a grade. Faculty must communicate with students before using detection results in “penalizing or otherwise affecting a student’s grade.”

Then comes guideline six. Before the Committee on Standards reaches a finding of responsibility in a suspected AI case, it will look at a “constellation” of “information beyond the result of an AI detection tool.”

That is the sentence that matters. Dartmouth authorized one detector and, in the same document, told its faculty that the score cannot carry a case by itself.

Carey also told The Dartmouth that faculty “are not required” to use detectors at all. The point, he said, was clarity about what to tell students, how to go about it, and “how to respect and protect data privacy.”

The constellation has to come from somewhere

Here is the question guideline six leaves on your desk. If the detector result is one point of light, where do the other points come from?

Not from the detector. Running the same paper through it again tells you nothing new. Not from instinct either, because “it does not sound like her” is exactly the kind of evidence a standards committee should be uncomfortable with.

The other points come from the assignment, and only if the assignment was built to produce them before the paper was due. A dated outline. A source log. A short checkpoint written in class. A five minute conversation where the student walks you through one decision. A record of what changed between the draft and the final.

Now look at what Dartmouth’s review of its own provost is doing. The committee is not stopping at the Pangram number. It is measuring his use against the rules each publication set, which is a question about stated standards and about process. That is a constellation. The provost’s case is being judged the way guideline six says a student’s case should be judged.

Most assignments being handed in this week cannot support that standard. They produce one artifact, one time, with nothing around it. If a detector flags it, the instructor holds a number and an impression, and the student holds the same defense a provost just used.

The rule your students sign now has a mirror

Dartmouth’s coursework guidelines start from a clear default. Students may not use generative AI tools for submitted coursework unless expressly permitted, and under the honor policy any use in submitted work requires acknowledgment.

The faculty petition put the problem in one line: “How do we ask our students and colleagues to follow the rules when the academic leader of Dartmouth is allowed to break them?”

You do not need an opinion about the provost to take the lesson. Students are watching what happens when an adult with power gets flagged. If the answer is a process, with notice, a conversation, evidence beyond the number, and a judgment against a stated rule, then that process is what they will expect when it is their paper. Dartmouth just wrote that process down for its faculty. Most of us have not written it down for our own courses.

This is not an argument for or against detectors. A detector can be one input, and the guidelines say exactly that. The argument is that whatever you decide about the tool, the rest of the case has to be designed into the assignment and announced in the syllabus before the first paper arrives.

What to do before your next paper is due

Three things, and none of them require a committee.

One. Write your detector line. One sentence in the syllabus that says whether you will use a detector, what a result can and cannot do to a grade, and that you will talk with the student before anything changes. If Dartmouth has it right, students are owed that sentence before the first submission, not after the first accusation.

Two. Pick your three points of light. For your next major assignment, name three pieces of evidence it will collect besides the final product. Keep them small. A source log and a five minute walk through of one decision cost you almost nothing to collect and are very hard to manufacture after the fact.

Three. Write the conversation now. Write the three questions you would ask a student in that conversation while you are calm and nobody has been accused. Questions written in advance are fairer than questions written in suspicion.

Do those three things for one assignment and you hold something a detector cannot give you: a case that stands up whether the score comes back high or low.

The detector can tell you a paper reads like a model. Only the assignment you designed can tell you who did the thinking.

Your next step

If you want the fastest version of this, take one assignment through The Constellation File, by hand, in about 30 minutes.

When you finish you will hold one syllabus line, one assignment with three points of light, and three conversation questions. Then the next thought arrives on its own. That is one assignment. Your course has more of them, and so does the next section.

That is the job the Assessment Builder does. It takes an outcome you already wrote and builds the assessment with its evidence checkpoints and the rubric line that grades them, so the constellation exists before the paper arrives, for every assignment in the course instead of one.

If you would rather think it through with other instructors first, the Harvest Kernel Learning Community is free, and this question is the conversation there this week.

Sources

• Jackson Hyde, “Dean of faculty authorizes professors to use AI detector Pangram to review student work,” The Dartmouth, October 1, 2026. thedartmouth.com
• Alex Klee and Isabela Pierry, “Op-eds, academic articles by Provost Santiago Schnell flagged as ‘AI-written’ by ‘near-zero error’ AI detector Pangram,” The Dartmouth, September 21, 2026. thedartmouth.com
• Santiago Schnell, letter to the editor, The Dartmouth, September 2026. thedartmouth.com
• Sian Leah Beilock, “A Review and a Community Conversation on AI,” Office of the President, Dartmouth, September 25, 2026. president.dartmouth.edu
• Kathryn Palmer, “Dartmouth Provost’s ‘AI Dependency’ Sparks Backlash,” Inside Higher Ed, September 29, 2026. insidehighered.com
• “AI writing has already begun to appear on the opinion pages,” Semafor, August 26, 2026. semafor.com
• Dartmouth Policy Portal, “Guidelines on using Generative Artificial Intelligence (GenAI) for Coursework.” policies.dartmouth.edu