‘The bulk of the benefit and the threat is actually humans’

Science & Tech

‘The bulk of the benefit and the threat is actually humans.’

Isaac Kohane cautions against both doomer and utopian visions of AI: ‘We have agency here.’

Sy Boles

Harvard Staff Writer

October 23, 2026

8 min read

A series of faculty interviews contemplating a new age of disruption.

As the inaugural chair of the Harvard Medical School’s Department of Biomedical Informatics, Isaac Kohane thinks deeply about how artificial intelligence tools can help and hurt in the field of medicine. The same set of questions led him to found the Human Values Project, an international research initiative studying the values embedded in clinical AI and developing tools to make those values visible.

We talked to Kohane about AI in healthcare and education, and as a driver of ethical challenges for universities and policymakers.

This conversation has been edited for clarity and length.

Gazette: What is the role of universities in AI development?

Isaac Kohane: A fair assessment of the last 15 years is that most of the advancements in AI have come out of industry. Currently, it does seem that the highest-performing models are indeed the largest models, and that creates an advantage for organizations that are fully aligned with the business of building these models so that their business succeeds.

In that particular comparison, universities don’t look relevant, but I think that’s an illusion. Universities have a role to play, and it’s in two domains: What are the right applications for AI, and what are the right ethical frameworks in AI governance?

“Universities have a role to play, and it’s in two domains: What are the right applications for AI, and what are the right ethical frameworks in AI governance?”Isaac Kohane

There’s been a bizarre assumption that the leaders of these AI companies can either have or buy ethical guidance to do what they’re doing throughout the world. That seems to me to be prima facie incorrect, because ethical frameworks change from location to location.

Can you give me an example?

The Human Values Project asked a whole bunch of doctors around the world some thorny questions. For example, you’re presented with a woman who has severe anorexia. Her electrolytes look bad. She’s vomiting. You don’t hear bowel sounds. Her heart rate and ECG look normal. She’s refusing to eat; she says, “I’ll eat tomorrow.” If you ask a bunch of doctors if they should put in a feeding tube against her will or not, shockingly, doctors split 50-50. Perhaps less shockingly, so do the AI models.

Within that, there’s a lot of heterogeneity. It turns out that American doctors, compared to peers in East Asia or South America or in Central Europe, care much more about patient autonomy. It gets beaten into us, whether it’s because of our education or a fear of malpractice lawsuits. The AI models are much less pro-autonomy. On average, we found that these AI models agree more often with doctors around the world, and less with doctors in the U.S.

But the way enterprises work, they don’t work with lots and lots of AI models and go with the average. Typically, they pick one. And what we find across not just that example but across dozens of vignettes is that these AI models have their own ethical north stars. They’re consistent. Some AI models are more pro-autonomy; others are more pro-beneficence; others more pro-justice. But humans also differ. Some people want to live the longest possible life, and others just want to make it to their daughter’s wedding.

These questions are what big parts of the humanities have discussed, what the history of science has a lot to say about. This involves governance and government and law. So I think universities have an enormous role to play here. This can’t be left to a few companies who don’t have that rich tradition of thought. We have riches in that sense, especially if we embrace the full diversity of universities, which we should and will in the future, if we want AI to reflect the diversity of the human condition.

But it does mean that we, as universities, have to get out of the ivory tower a bit and engage more with the public.

What do you make of AI in medical education?

Right now, nobody has any magic recipe about the best way to use AI in education. My own kids, who are in college, tell me that if they use AI to help them solve problems, they’re not as ready come exam time. On the other hand, if they’re stuck, AI is very useful to them. So clearly there’s going to be a method for the madness, and it’s going to differ by different tasks and different specialties.

One of our biggest tasks as human beings is to take squishy little kids and turn them into full, responsible adults. There are unthinking ways to do that, which we’ve already seen with social media, and there are ways to do that with AI that are predictably not going to be so good.

Paradoxically, I think this is going to put education back in the driver’s seat. The reality is, the advent of the internet and video conferencing made students essentially defect from personalized engagement. The challenge for us is to ask, “What’s the added value that we can give that they would not get otherwise to be successful in life?” I don’t claim that we have the answers, but who else, other than universities, has the resources and the people who care about this — who can actually focus on it?

I’m struck by what you said about different specialties requiring different methods. It seems to me that the university is also a place where we can put our heads together and hammer out the distinctions of appropriate uses for AI in law versus medicine versus literature.

This is a platitude, but it’s true: Like any tool, AI can make things worse and it can make things better. It can make us more siloed, but it can also make us look outside our disciplines. For example, you could imagine, as part of the requirements for any ethical task force or education class, you could ask AI, “Of all the faculty at Harvard, who else should be at the table?”

And there are places at Harvard where we’re already incredibly cross-barrier. But if we want to be insular, AI is not going to stop us.

Do you think medicine as a field has been too quick to adopt AI tools?

I wrote an editorial for a perspective piece in the New England Journal of Medicine last year called “Compared with What?”, where I argued that we should not be so sanctimonious or fastidious when we say we should not use AI for healthcare because, compared to the best possible doctor, it’s inferior. Because, I wrote, most people don’t have the best possible doctor. On average, people have the average doctor.

The average doctor, we’re fallible. In fact, we’re not even always available. Our own residents in training at our hospitals don’t have their own primary doctors, and even under a socialized medicine system like in the U.K., you can be in a lot of pain and still not be able to see your primary care doctor for six months. Compared to that, you might be better off having a conversation with one or more AIs.

I think a lot about what I call the “bothness” of AI: peoples’ concerns about cognitive offloading, impacts on the job market, and environmental impacts, among other worries, and simultaneously the incredible capacities of these tools to advance human knowledge, discover new treatments for disease, and so on. How do you make sense of that bothness?

I think the bothness of AI is the same as the bothness of humanity. The biggest benefits and the biggest threats of AI, it’s humans making that happen. In the Ukrainian-Russian War, we’re seeing drones becoming increasingly autonomous … but it’s not AI deciding by itself to go and kill some humans: People are telling it to do that. And when I decide to use AI to treat myself, to ask whatever questions I forgot to ask my doctor or to get a second opinion, I’m the one doing the asking.

That’s not to deny the possibility of fully autonomous AI. But I think the bulk of the benefit and the threat is actually humans. We have to be responsible to each other. We have agency here. The doomer feeling out there is from giving up agency, and I think that’s a real mistake, because the bad actors out there aren’t going to give up theirs. The solution is to read carefully. Don’t get swept up in either the doomer stuff or the utopian stuff. Our society needs conversations about this, not fearmongering, and I think universities are a great place for those conversations to happen.