“This Year I Believe We Must Act”: Deming’s Take on Harvard’s AI Policy

On Sept. 2, the first day of classes, Danoff Dean of Harvard College David J. Deming emailed the student body, sharing his thoughts and goals for various aspects of the first-year experience. He suggested a “barbell approach,” meaning allowing AI in some areas and prohibiting it in others. His suggestion included replacing take-home exams and problem sets with in-class work while allowing AI in writing-intensive classes.

Rather than quell student concerns, this email seems to have raised more questions about what Harvard’s AI policy will look like and offers little clarity on how it would be implemented.

In his email, Deming wrote, “We would all benefit from getting out of the AI-detection business because it damages trust between faculty and students.” Deming also cited the personal blog of Colorado State University political science professor Kyle Saunders. In the entry Deming pointed to, Saunders explained his philosophy of encouraging AI in writing courses, emphasizing that student AI use cannot be effectively eliminated from assignments outside the classroom.

Saunders further noted that AI must be “‘encouraged’ and not just ‘allowed’” in papers to reduce the uncertainty students’ worry that their work will be looked down on for using AI, even though it is technically permitted. Adopting Saunders’ model would encourage students to use AI to whatever extent they deem appropriate, with the expectation that they fully disclose it and remain accountable for accurate, thoughtful work.

This idea is far from new. Increasingly, AI is used as a tool that can encourage creativity in students and produce more effective writing. One review of public data found that AI feedback produced a positive moderate effect on university students’ writing. However, concerns remain about whether students develop critical thinking and writing skills when they use AI in coursework.

The practical effects of such policies in classrooms largely depend on how students choose to use AI—is this software being used to deepen understanding or to outsource the work of learning? According to the “Harvard Undergraduate Survey on Generative AI,” completed in 2024 by Shikoh Hirabayashi ’24, Rishab Jain ’25, Nikola Jurković ’25, and Gabriel Wu ’25, roughly one-third of Harvard students worry about their peers using AI to gain an unfair advantage.

“All of my classes have a strict no AI policy, and that’s personally my preference because I don’t think it is conducive to learning. And I think people rely on it too much,” Gabriel Sloan ’30 said in an interview with the “Harvard Independent.”

The taboo nature of AI use in humanities and social sciences, which are generally writing-heavy, presents a particular challenge to creating a pro-AI academic environment. However, in STEM courses, AI use would be further restricted rather than encouraged.

“[Take-home exams and problem sets] create a difficult dilemma for students who don’t want to cheat but also fear falling behind to their classmates,” Deming wrote. He suggested leaning toward blue-book exams and cutting down take-home work as a solution.

While switching to blue-book exams would prevent use of AI in assessments, there might be downsides, too. “I think that we must be very clear about what is lost when instructors can no longer assign work to be done outside the classroom, which is simply anywhere between hundreds to thousands of hours of reading, writing, thinking; in other words, of learning,” said Dr. Anastasia Berg ’09, a University of California, Irvine philosophy professor and author of the article “The Ivy League Surrender to AI Is Insane,” in an interview with the “Harvard Independent.”

“We must also not forget that all the time we now have to use for in-class writing will come at the expense of pedagogical activities that we found until very recently to be essential for education: lectures, discussions, student questions,” she continued.

Beyond the educational perspective, Berg pointed out how encouraging AI could affect students’ cognitive abilities. “Analyzing a challenging text, locating its central claims and the strategies employed to support them, coming up with objections, refining one’s employment of concepts—these are all fundamental philosophical skills.”

The cognitive angle matters because college students are still developing those skills. “Many people who are happy to incorporate AI in education are extrapolating from the impression that they see themselves use AI effectively to the idea that it would be a good idea to let people work with it as early as possible,” Berg noted. “I think that to the extent that such people are actually using it effectively, it is because they’ve had decades of developing their cognitive skills without AI.”

The greater question remains: just what is effective AI use, and how do we ensure students develop the cognitive skills they need? Moreover, is it up to the College or University to decide this as a whole, or should professors have a greater say over what happens in their courses?

For Grace Klausner ’30, a college-wide AI policy is not a major concern: “I am relatively neutral with the AI policy because most classes have laid it out relatively clearly.” Students like Klausner have already experienced AI policy and its enforcement, just at an individualized level.

Institutionalizing a standard AI policy could further restrict professors’ choices in course policies. The divisive grade-cap policy, set to be implemented at the start of next year, will limit the number of students receiving A grades in each class and has already constrained professors’ control over their course policies and syllabi. While an AI policy may remain a suggestion, any formal policy would affect classroom day-to-day teaching.

Delfin Urundul ’30 (durundul@college.harvard.edu) and Star Binyamin ’30 (starbinyamin@college.harvard.edu) are comping the “Harvard Independent.”