AI Tutoring Hype Collides With Trial Showing Lower Grades

A trust problem that starts at the top

This autumn, the academic AI debate has stopped being about whether students cheat. The question now is whether anyone in higher education, from provosts to deans to the people promoting edtech, can be trusted to describe what AI does to learning. Three developments landed within weeks of each other. A viral post exaggerated a Harvard tutoring study. A senior Ivy League administrator was caught using AI to write warnings about AI. And a large randomized trial found that giving students an AI tutor was associated with lower grades. Read together, they point to an expensive gap between the claims and the evidence. Universities, families and school districts are all paying for that gap.

The most widely shared case of a top academic stumbling over AI did not happen at Harvard. Dartmouth College opened an investigation into its provost, Santiago Schnell, after the student newspaper found apparent AI use in work published under his name.8 The Dartmouth reported on September 21 that Schnell seemed to have leaned heavily on AI for all nine articles he had published this year. One of them was a Washington Post op-ed about students outsourcing their writing to AI.4 The AI detector Pangram had already classified that essay, titled "Universities are fighting AI cheating. But there's a deeper problem," as 100% AI-written.8

Schnell said he drafted the piece himself and then used ChatGPT to sharpen some arguments and check grammar. He later apologized and said he would disclose AI use in his public writing from now on.8 Students were not persuaded. One Dartmouth senior called the episode a "completely cartoonish" display of hypocrisy, and students more broadly saw a double standard: they are penalized for outsourcing work while institutional leaders are allowed to do the same.8

The Atlantic places Schnell in a longer list of similar cases. It includes an academic who defended an AI-assisted op-ed as written "with" AI rather than "by" it, and a Stanford misinformation expert whose court declaration contained citations ChatGPT had made up.418 The magazine's point is that none of these people gave up the tools after being caught.18 That is the real story. The people in charge of academic integrity are using the same tools they tell students to avoid. They separate their "ideas" from their "writing" in a way that sounds a lot like the excuses students give.4

Harvard's own AI fault lines

Harvard's problems are about policy, not one person's mistake, and they are deep. College Dean David J. Deming recommended that students' use of AI in writing-intensive classes be accepted and even encouraged. Historian Jill Lepore called that "the most outrageous instance" of an administration that is "oblivious" to the risks.17 She contrasted it with an August MIT report that warned many uses of AI "deprive students of the opportunity to learn."17 She also cited a survey finding that more than a third of Harvard undergraduates polled had used AI in ways that broke course rules.17

The Atlantic reported that a Harvard dean emailed students arguing that the college should get out of "the AI-detection business."12 English professor Derek Miller, who chairs the faculty committee on information technology, has called for Harvard to "reboot" its approach. He noted that colleagues are going back to handwritten work, blue books and oral exams, and argued that the administration too often treats a chatbot summary as equal to the slow work of reading and writing.2

Campus coverage shows a decentralized system that is struggling. The College's only standing AI rule is procedural: instructors have to tell students what their course policy is. The AI faculty committee chair said he does not expect a "global edict."15 Course syllabi range from total bans to active use.15 Faculty say AI cheating is "virtually impossible to prove," so plagiarism cases before the Honor Council have stayed at about 50 a year even as AI use has spread.15 One sophomore admitted she remembers little from last semester's classes after using ChatGPT summaries to study.15 Harvard's Faculty of Arts and Sciences is also changing vendors. It is adding Anthropic's Claude and ending blanket ChatGPT Edu access after June 2026.11

The study that went viral, and what it actually showed

The "what it means for your money" angle comes from hype about a Harvard study, not from any scandal. A widely shared post on X said a Harvard experiment had "deleted every reason universities exist." It argued that because an AI tutor beat a $60,000-a-year classroom, "the math of higher education breaks permanently."9

The study behind that claim is real and worth taking seriously. Kestin and colleagues ran a randomized trial in a Harvard physics course. They found students learned more than twice as much in less time with a custom AI tutor than in an in-class active-learning lesson, and that the students felt more engaged.29 Secondary write-ups give an effect size of 0.73 to 1.3 standard deviations and a median time on task of 49 minutes for the AI group versus 60 for the classroom.27 Reports do not agree on the details, though. A parent-oriented summary describes 194 students who each tried both formats over two weeks and says the authors saw AI as a complement to teaching, not a replacement.25 The paper's own figures list 142 students in the AI group and 174 in the classroom group.29

The viral post's argument breaks down on scope. The tutor was purpose-built with strict pedagogical guardrails. The claim that it replaces a whole university rests on one physics unit tested on immediate learning gains.929 A new working paper makes the same point directly: whether the Harvard result applies beyond "a single course at a highly selective university remains uncertain."28

The counter-evidence is large and inconvenient

That working paper may be the most important AI-education study of the autumn. Researchers randomized access to a course-integrated generative AI tutor for 2,379 undergraduates and 30 instructors across several subjects at a large U.S. public university.28 Among sections of the same course, students with access to the tutor ended up with final grades 0.37 standard deviations lower, about four points on a 100-point scale. Their participation on the learning platform fell by 0.90 standard deviations.28 First-generation students lost more than twice as much on final grades as their continuing-generation peers.28 The decline appeared even though only about 15% of students offered the tutor actually used it. The authors take this to mean that simply being assigned the tool may have pulled students away from other course activities.28

The paper places itself in a growing body of research. That research finds general-purpose AI can raise performance on assisted tasks while hurting later independent work, unless tutoring safeguards are built in.28 Lepore cites related findings: a Penn study in which high-school math students with ChatGPT did worse once access was taken away, and a Georgia Tech field study in which ChatGPT users learned less than Google users.17

There is good news for AI tutoring, but it is conditional. Stanford researchers describe two K-12 trials in which AI embedded in live math tutoring improved results. In one, students working with Google's LearnLM had a 66% success rate on harder follow-up topics, compared with 61% for human-only tutoring. In the other, a tutor-assist tool helped inexperienced tutors the most.21 Reviews of trial results find that gains cluster in structured subjects like algebra and grammar. AI tutors tend to tie or slightly trail well-staffed small-group human tutoring.24

Why the money question is real

Across this coverage, the pattern is consistent. Carefully designed AI tutors, kept within narrow and structured tasks, can help. Unrestricted chatbots and tools added on top of existing courses can quietly damage the learning they are supposed to support.2821 That has direct financial consequences. Districts and universities are signing AI contracts, and some vendor-friendly marketing advertises figures like "54% higher test scores" with little sign of independent evaluation behind them. For families paying tuition, the risk is not that AI makes a degree worthless overnight, as the viral post claimed.9 The risk is paying for a credential whose assessments can no longer show what a student actually knows.

Detection does not fix this. An MIT working group advised against relying on AI detectors, warning they could create an "atmosphere of distrust between instructors and students." Even the detector companies say their scores should start a conversation, not settle it.12 Professors at several universities are going back to oral and in-person exams.6 One biology professor said he has "pretty much thrown in the towel" on traditional writing assignments for lower-level classes.6

The verdict

The Dartmouth episode and Harvard's policy fight are symptoms of a single problem. Institutions are adopting AI faster than they are testing it, and the leaders responsible for integrity are not following their own rules. The research does not justify panic or triumph. It supports a narrower rule: buy and use purpose-built tutors with guardrails, measure what students can do without AI, and treat any claim of revolutionary learning gains as marketing until a large independent trial says otherwise. Right now, the largest such trial points in the opposite direction.28