Deming Says He Hopes to ‘Push Harder’ on AI, Acknowledges Widespread Student Use

Harvard College Dean David J. Deming acknowledged that far more students are likely using artificial intelligence than the College’s disciplinary system has been able to detect, saying he hopes to “push harder” on the issue in the years ahead.

Speaking at a fireside chat for Academic Integrity Week on Thursday and at an open forum earlier in the week, Deming described AI as an urgent challenge facing the College — but stopped short of proposing a standardized policy to govern its use, pointing to the difficulty of writing rules that work across courses as different as Computer Science 50 and an upper-level English seminar.

“The biggest challenge, from my perspective, with AI is that it blurs the boundaries of what we would call cheating or academic integrity in ways that actually make it really hard to write policy around,” Deming said Thursday.

Deming said the clearest cases of AI misuse — students submitting AI-generated work with hallucinated citations or accidentally including the chatbot prompt in their submission — tend to be the ones that reach the Honor Council, the student and faculty body tasked with reviewing potential violations of academic integrity policies.

“But for every one of those, there are many cases where students are using AI in a more subtle way,” he said.

The Honor Council does not publish statistics on how many of its cases have involved the technology. When asked Thursday whether he could share the proportion of cases classified as AI-related, Deming said he did not know the figure.

Assistant Dean of Harvard College Dwight Fee, who moderated the fireside chat and has served on the Honor Council, said the body requires clear evidence of AI misuse before pursuing a case.

“If we have to get into guessing, it just doesn’t become a case,” Fee said.

Asked to estimate how widespread AI misuse is among students, Deming cited a survey by the Harvard AI safety group that found 88 percent of students were using AI at least weekly. He said he expected the share to be higher now.

“We should all expect that it’s being used much more than we are detecting,” he said at the forum. “Just as a matter of statistical reasoning.”

But Deming said the problem extends well beyond the students who leverage AI to cheat. He described a “more insidious” dynamic in which students use the technology to take cognitive shortcuts that undermine the learning process — comparing it to copying a classmate’s problem set answer and believing they understand it.

“In order to learn, learning has to be effortful. You have to struggle with it a bit to encode it into your memory,” Deming said. “The struggle is the point. The struggle is what locks it in.”

He said the challenge of AI in the classroom is fundamentally different from traditional cheating because the technology makes it possible to outsource thinking in ways that are difficult to detect and do not fall neatly into existing policy categories.

“What about if I write the essay myself, and then I give it to ChatGPT to edit — is that cheating?” Deming said Thursday.

“One kind of moral guide is, ‘Would I be willing to admit that I did it?’ And if you would, it’s probably okay. If you wouldn’t, it probably isn’t,” he added.

Deming said the College is unlikely to adopt a single AI use policy, instead leaving faculty broad discretion over how the technology is handled in their courses. The approach has led to variation across syllabi, with some courses banning AI entirely, and others encouraging its use.

“The way you’d use AI in CS50 versus an upper-level English seminar — why would you want the same guardrails and guidelines for those two classes?” he said. “Faculty know what they want, and so I think about it more as we’re trying to help people who want the help.”

“I don’t want to say we’ll never have a policy, but I think that seems like a difficult nut to crack because the courses are so different and the goals are so different,” he added.

Deming said his office has worked with Dean of Undergraduate Education Amanda Claybaugh, Bok Center Director Karen L. Thornber, and former Dean of Science Christopher W. Stubbs to develop flexible guidelines for faculty in different situations — from those who want to ban AI entirely to those using it as a pedagogical tool.

“We’re trying to develop a strategy that is flexible enough to accommodate all those cases and then provide resources to people who are in those different buckets,” Deming said.

He also said he and Claybaugh convened a focus group of students to hear how they experience AI in the classroom, including examples of when the technology has been used successfully and when it has not.

Deming noted that in conversations with students over the past year, he has been struck by how many said they are actively trying to avoid AI.

“I’ve talked to a surprising number of students who not only are skeptical of it but actively try to avoid the technology,” he said. “Who I think hunger for an intimate environment — classroom environment — with no screens, no AI, no distractions, just full engagement in the moment.”

When asked what advice he would give students navigating varying course-specific AI policies, Deming offered a personal framework: he asks himself how much effort he would have put into a task before AI existed, and uses the technology only if it allows him to go deeper on a task and his effort remains the same.

“If I’m using AI to take shortcuts, and only I know whether I’m taking shortcuts or not, then it’s a bad use of the technology,” he said. “If I’m using it to go further or deeper or broader, then it’s a good use.”

Harvard has moved to integrate AI into some of its own operations this school year. Harvard Library is incorporating AI into its search systems, and the College plans to introduce a new AI chatbot advising tool for freshmen this summer.

Deming made AI in the classroom one of his top priorities when he took office last July. He said this year’s work — focused on surveying the landscape and supporting faculty — was only a first step.

“I hope in future years, I’ll try to push harder,” he said. “I think this is a really urgent issue, and I think that AI really makes me — makes all of us — need to think hard about what are we actually doing here? What is the purpose of an education?”

—Staff writer Sebastian B. Connolly can be reached at [email protected] or on Signal @sbc.23. Follow him on X @SebastianC4784.

—Staff writer Summer E. Rose can be reached at [email protected] or on Signal @ser.85. Follow her on X @summerellenrose.