Why Harvard's Dean 'Encourages' Students to Use AI

The end of summer brings a fresh throng of Americans to college campuses. But the vibes are wobbly in higher education:

• In surveys, trust in universities has fallen to near-record lows.
• The value of a degree is in question, as tuition costs have increased in the last few decades while the college premium—the expected benefit from getting a diploma—has flatlined and possibly even declined in the last few years.
• The unemployment rate of young college graduates seems to be rising.
• The Trump administration has waged war against various schools and the general principle of on-campus diversity.
• Artificial intelligence has unleashed a torrent of cheating on campus and created chaos in the job application process, as both students and employers use large language models: the former, to write hundreds of AI-assisted applications; the latter, to screen those AI-written applications with AI-written filters, thus removing from the process of finding a first job those procedural frictions sometimes known as “people.”

As higher ed grasps for a reason for being in a new age, some people say college should lower its sights and aim to be little more than a country club with lecture halls; or a book club with dorms; or a multiyear, every-season-but-summer camp. But even that diminutive identity seems endangered. The Atlantic has reported that “Even College Students Have Stopped Hanging Out.”

Because there are thousands of colleges and universities across the country, it’s almost impossible to offer any diagnosis of higher ed that applies to every single institution. But the trend line is worrying, nonetheless. Costs are rising, perceived benefits are falling, trust is imploding, and nobody knows what to do about this machine intelligence that can write final exams with one dumb prompt.

Into this tempest, David Deming, an economist who also oversees undergraduate education at Harvard College, threw another lightning bolt. In an email to students, he proposed embracing a policy of so-called “AI encouragement.” The notion that Harvard—Harvard!—would encourage the use of AI sent shudders through the faculty and student body. It resounded across the country, with many professors and writers accusing America’s most famous university of surrendering to techno-determinism and sacrificing academic integrity in the process.

Today we hear directly from David Deming. He and I talk about the history of higher education in America and ask two big questions: What is college for, anyway? And how should professors across the country, at Harvard and everywhere else, teach students to become not only deep experts but also deep thinkers in an age when our computers frequently whisper that they’re happy to do our thinking for us?

The Three Eras of Higher Education

Derek Thompson: In a recent essay, you described three eras of undergraduate education in American history. What was the first era, and why did it end?

David Deming: In 1636 when Harvard was founded, the founding gift was 400 books from John Harvard to the library. That doesn’t sound like a lot, but it was about two-thirds the size of Emmanuel College at Cambridge, Harvard’s alma mater. Overnight, Harvard became a unique repository of knowledge. Our founding story was built around providing access to knowledge that was scarce.

The first printing press in the American colonies was housed in President Henry Dunster’s house in Harvard Yard. For a long time it was the only one in the colonies. When our library caught fire in the 1700s, Harvard sent students home to study in their local communities, because without access to those books there was no reason to be on campus. So at the founding of this university and many others, access to knowledge you couldn’t find anywhere else was the whole purpose.

That didn’t last. By the late 19th and early 20th century, knowledge had proliferated. Our library was no longer the only library, our printing press no longer the only press. The great university presidents of that era, in our case Eliot and Lowell, opened up the curriculum, created electives, and let people specialize.

The second age of education was the era of expertise. This is the modern university as we understand it: a flourishing of disciplines, philosophy, history, English, economics, biology, with faculty who are the world’s leading experts in a narrow domain curating students’ access to expertise. I believe we’re now near the end of that second era.

Thompson: A few years ago, I read about “the death of the Renaissance man.” The idea was that intellectuals used to learn a little physics, a little biology, a little literature. They were well-rounded without being deep experts in anything, because the scientific revolution was too young to allow for deep expertise. Do you think the age of experts, for all the wonderful things it produced, was also a step back—that something was lost as people became more narrowly focused rather than broadly well-rounded?

Deming: I agree with that. It was a step forward for scientific research, developing deep expertise and applying it across a complex society that requires it. We don’t just have family doctors anymore, we have radiologists, cardiologists, oncologists, and if you live in Boston, several kinds of interventional cardiologists. That proliferation of expertise is the key to a lot of human progress. But it had a trade-off: it fragmented us. The more specialized you are as an academic, the less you understand the intellectual life and culture of other disciplines.

In the early days of Harvard, everybody studied the same narrow classical curriculum, which produced Renaissance men and women because you had to be broad. Now you come to a place like this and hone in on something to become an expert. That’s great for society and for research, but it also pushes us away from each other, against our communal instincts and against the idea of a university. Expertise is still important, but the question is whether the university is the only place to get it. I think that’s much less true than it used to be. So the question now is what is the value of a university in a world where you have other ways to access expertise.

Thompson: So, the first era of education was about information. The second era was about expertise. What’s the third era?

Deming: We’re building it as we speak. I think it’s in some ways a return to the founding purpose of a college, the medieval collegium: a learning community of scholars and students building the kinds of skills, habits of mind, and virtues that can only be built in community. There’s an academic side to it and a social side to it.

If you want to learn how to work well in a team, how to lead or be led, how to give and receive criticism, those are things you can’t learn from behind a computer screen or on your own. You can’t learn virtues like integrity or honesty unless they’re tested in community, tested against whether they’re costly to you. You can’t learn about justice or fairness unless you have to adjudicate competing legitimate claims to resources within a community.

So I think this is a thing our college can still do uniquely: bring some of the most talented young people in the world here for four years to learn together and from experts and faculty. That won’t change. But I think we need to double down and reemphasize the communal aspects of learning, the skills and habits of mind and virtues we can’t get from behind a computer screen or a YouTube video.

AI and the Third Era

Thompson: Your story so far is technological. The proliferation of knowledge forced the turn from the first era to the second. Is there another technology forcing the turn from the second era to the third?

Deming: This technology is artificial intelligence. AI is controversial, both a source of hope and of real concern. I want to be clear this isn’t AI boosterism, it’s not “the university needs to embrace AI and use it all the time.” It’s an acknowledgment that when models trained on the collective expertise of humanity can provide personalized learning and personalized access to expertise, sufficiently motivated people can get expertise on any subject. In my own field, I’ve seen a frontier model write an economic model better than I can, better than most economists can, and explain things to an eight-year-old, or an eighteen-year-old, better than I can.

It’s not a right-now thing. But looking down the pipeline, I think it’s clear we will no longer have the monopoly on the expertise we once had. That doesn’t mean expertise doesn’t matter. But we can’t rely on that monopoly to sustain the university in this third age. AI will commodify expertise and increase access to it, and that will be a good thing for humanity, the same way the expansion of knowledge was good in the first two eras. If you’re a young person in a poor country without access to a place like Harvard, AI is a tremendous boon, because you now have access to expertise you never would have had before. But you still don’t have access to learning in community with other talented, motivated people. You still need to come to a place like this.

Thompson: The idea that AI is forcing us to change what college is and is for is a contested one. There are people who think AI is a scam, a fad, a bubble. They’d respond to what you just said by pointing out that Wikipedia has existed for decades and hasn’t suddenly made the world smarter. If the only two things you knew about the world were that Wikipedia exists and PISA test scores, you’d think Wikipedia was mechanically driving down the world’s ability to read, do science, and do math. Fifteen years ago there was similar excitement around massively open online courses. Why isn’t AI just MOOC 2.0?

Deming: It might be. I’m not making a prediction, I’m just looking down the pipeline like everybody else.

A MOOC is essentially a lecture, and that’s still true today: You can go on YouTube and access the best lecture on any topic, probably better than what you’d get in your college course, as far as a lecture goes. But that’s not the primary purpose of an education. Part of it is meeting your individualized learning needs, not just delivering content but helping you understand it in a way that speaks to you, diagnosing the gaps in your knowledge, and providing an experience that helps people learn together.

Learning is communal partly because it requires a bit of suffering to be successful. Most of the important things I’ve learned, I didn’t understand the first time, and I needed a push to focus on them. Some people are autodidacts who can learn on their own, but most people don’t buy textbooks and master subjects at home, even though they’d like to, because it’s hard. The community of learning together, which is what a college education provides, is a way to tie ourselves to the mast and do hard things together. I think the MOOC revolution misunderstood that education isn’t just the delivery of content, it’s the ability to learn together and to customize learning for the learner. That’s what AI can do that MOOCs couldn’t.

AI will meet you where you are. You can tell an AI tutor “I don’t understand this, explain it again,” and there’s a lot of research showing that when people take advantage of AI’s personalization features, with guardrails, meaning you do the initial writing or take a baseline test yourself and then use AI to refine and get feedback, it can meaningfully boost learning gains. That’s what’s different about AI versus MOOCs. But AI still doesn’t do the communal part, the learning together, the binding ourselves to the mast. That’s still the most human aspect of our education, the part AI can’t fully cover.

Harvard’s Controversial ‘AI Encouragement’ Policy

Thompson: There’s a live and ferocious debate about what to do about AI in college, and different schools are reaching very different conclusions. Caltech faculty voted to strongly encourage in-person, seated exams so students can’t rely on take-home exams that could be faked with ChatGPT. Faculty in the social science core at the University of Chicago announced an AI ban and a prohibition on all classroom technology starting this fall. Two weeks ago, you sent an email to Harvard students on the first day of class endorsing a policy you called “AI encouragement.” What does that mean?

Deming: We have to honestly acknowledge that everybody, not just our students, is using AI all the time for everything, and design our education around that reality. What I said in my email was that I favor a barbell strategy. Either you have an enforceable way to ban it, which I’d support in many cases, or you don’t. I actually think using AI to do the writing and thinking for you is bad, and as a 47-year-old, it’s not that hard for me to put the AI down and write on my own, because I grew up in a pre-AI era. But for students coming to college today, most of their high school writing experience already included this tool, and many of them need help building guardrails around it.

It’s not sufficient to say “don’t use AI, it’s banned” and then not enforce it, hand out a take-home paper, and expect compliance. Some students will comply, but they’ll feel bad about it, because they’ll look around and think, “If I behave with integrity and follow the course policy, I’m disadvantaged relative to classmates who secretly use AI and get a leg up.” We don’t want rules that reward the wrong behavior and punish the right one.

So my statement wasn’t AI encouragement in the sense of “AI is awesome, use it all the time.” It was: if you’re not going to have an enforceable ban, you may as well encourage it, because everyone’s going to use it anyway, and you should build resilience to that in your class. There are a lot of ways to use AI without having it do the writing for you. I never use it to write for me, but I use it for feedback. I’ll write a column and say, “Interrogate this, is it clear, give me feedback.” AI can help refine your arguments and do fact-checking, which is different from doing the thinking for you. So what I was encouraging faculty to do was embrace AI as a tool, build guardrails that make sense in their own classroom, and acknowledge that students are using it.

What happens in these conversations is that we fall into “is AI good or bad?” and if you say you’re encouraging it, people assume you’re in the good camp. But in a position like mine, you can’t really afford to take a side. You have to ask what’s the best way to design guidelines that are resilient to the fact that students have this tool, while preserving what’s really important about teaching and learning.

Thompson: Let’s be specific. Let’s say you’re teaching a class on the economic history of higher education. How exactly would you both encourage AI use and ensure people are learning rather than passing off Claude and ChatGPT prompts as their own thinking on take-home tests and finals? Walk me through how a policy of “AI encouragement” would work without creating the cognitive atrophy that unscrupulous AI use in classes is inevitably going to produce.

Deming: It’s a great question. I’m teaching a class in the spring called Economic Possibilities for Artificial Intelligence, so this is very real for me.

I’m experimenting like everybody else, but here’s what I plan to do. I’ll have students write a rough draft of their thoughts after feedback and discussion in class, in controlled conditions—like in a blue book exam, or on a computer cut off from the internet. The point is to get students to commit to an idea in a controlled setting where I know the initial inspiration is their own, and then they bring it back, and in a supervised way we help them use AI to get feedback, do additional research, turn in a second and third draft, and make an oral presentation about something new they added, so I can verify they did the work themselves instead of having AI think for them.

One way to think about it: We need to get much more deeply into the process of writing and idea generation, to understand how students get from point A to point B, rather than just saying “go forth and come back with a twenty-page paper.” That’s costlier, it may mean smaller classes or more expertise in the room, I’m not sure, but I think it’s what we have to do more of.

I also plan to have a no-device classroom, an intensive conversation where people do the readings, come in, and possibly write a short reaction paper at the end of class, again in controlled conditions, so I can see how the discussion changed their point of view. That’s my own thinking. I’m going to try it this year, and if you have me back, maybe I’ll have an update on whether it worked.

Thompson: My wife just finished her PhD in clinical psychology and had to defend her dissertation. She has to stand there like a lawyer for her own position, defending her conclusions against adversarial questions.

What’s the case against making undergraduate education—especially in classes like economic history, history, philosophy, and writing, where the essay has long been the coin of the realm—into dissertation defenses, where the grade is the ability to defend your written ideas? The defense is its own proof: If a student doesn’t understand what they wrote, they get a D, which is pretty indicative they used AI to write it. But if they worked with a language model to build a brilliant argument they have absolute mastery over, isn’t that exactly what the era of expertise is supposed to be about, using whatever tools are at your disposal? So what’s the case against going all in on dissertation defenses?

Deming: I love dissertation defenses. I’ve done many of them myself, in exactly the way you described your wife doing. Before AI, I’d have students write a research statement about where their dissertation is going, with citations to the literature, and then in the defense I’d keep asking them questions about papers they considered important to their work until they couldn’t answer anymore. I used to tell students the point is to find the limits of your knowledge, because that’s what being an expert really means: there’s a frontier out there, and you have to get to it before you can see what the most important unanswered questions are. My job is to get you there, and an oral defense is a great way to do that.

If students know they’ll have to defend what they write, they have the incentive to use AI in the right ways rather than the wrong ones, because they’ll be accountable for their ideas. The historical problem is that it’s very hard to do an oral defense in a fifty- or hundred-person class, it’s expensive and intensive. Maybe there’s a way to use technology to do it better; some people have tried building AI oral examiners, and I haven’t seen a great example I’d endorse yet. But the point is we need to acknowledge this is happening, use these tools to make learning deeper and more fulfilling, but keep it human. Standing in front of the people who’ve advised you for years and defending your ideas is still a great way to learn.

The Negative Reaction to ‘AI Encouragement’

Thompson: What was the reaction to your AI encouragement policy from faculty and students?

Deming: People had concerns.