CAREER Q&A

Three ways to build a strong AI-training pipeline

Nature talks to artificial-intelligence researcher Oren Etzioni about how to ensure the health of academic programmes in the field in the United States.
Roberta Kwok is a freelance writer in Kirkland, Washington.

Search for this author in:

Oren Etzioni speaks at a TED 2015 conference in Vancounver, Canada

Artificial-intelligence researcher Oren Etzioni has suggestions for keeping enough AI faculty members around to train the next generation.Credit: Bret Hartman/TED

Oren Etzioni is chief executive of the non-profit Allen Institute for Artificial Intelligence (AI2) in Seattle, Washington, and is on leave from the nearby University of Washington. He offers some recommendations for how to stem the outflow of artificial-intelligence (AI) researchers from academia to industry — a loss that is damaging academia’s ability to teach incoming undergraduates.

How common is it for AI researchers to leave academia for industry?

It is a very sizeable trend for fresh PhD graduates and faculty members. In machine learning, you see some significant departures.

Industry compensation packages are highly variable. But together, the salary, stock and bonuses offered to a senior professor can exceed US$1 million per year. If you can barely pay your mortgage and have two kids who want to go to college, it’s very hard to say no to such an offer.

Some people come back to academia, and some don’t. When they do, they often bring rich experience with them. I took a leave of absence in 1999 to become the chief technology officer at an Internet company. I returned a year later and introduced a course on Internet systems, which became very popular.

Is there now a shortage in academia — for instance, in postdoc applicants or faculty-member retention?

The biggest source of unfulfilled demand is at the undergraduate level. At many US universities, students who want to get a computer-science or AI degree often aren’t able to. For instance, the University of Washington’s computer-science programme is highly selective and takes into account grades and numerous other factors; it’s able to accommodate about one-third of the students who complete prerequisites and apply. And the biggest factor that drives availability is faculty members to teach courses. There’s absolutely a supply-and-demand mismatch.

One key variable I feel passionate about is immigration. At my university, more than 40% of enrolled PhD students are foreign nationals. So immigration is a huge pipeline. Yet the current US administration has taken steps to curtail people coming in and to reduce graduates’ ability to stay, which creates a disincentive to apply. Could we fill all the graduate-programme slots with US citizens? Absolutely. Would it radically change the quality of the programme if it lost more than 40% of its PhD students? Of course it would.

What should universities and policymakers do?

First, make it easier for AI researchers to get funding. The bane of all scientists’ existence is chasing large grants. There needs to be less overhead, such as lengthy proposals, meetings and status reports. Universities can offer resources to reduce the busywork on their end.

Second, increase the number of women, other people from under-represented groups and immigrants who are applying to PhD programmes. To address the demand from undergraduates, you need more professors; to get them, you need more people getting PhDs. If there are more PhD-programme applicants, you can keep the quality constant, or potentially improve it because you have more people to choose from.

Third, within reason, academic compensation has to be responsive to supply and demand. No one can remotely match million-dollar compensation packages, but universities should consider increasing pay a bit.

People should take a long-term view and think about a healthy ecosystem for education and research. At AI2, we have created a sustainable alternative. We have made offers to University of Washington faculty members to spend part of their time at the university and part at AI2. It allows them to continue teaching, mentoring and fulfilling service obligations. AI2 also provides substantial research funding, which reduces time spent chasing after and managing grants. As a non-profit, we can’t afford to match the highest industry compensation packages. But we give a healthy increase over academic salaries and offer other benefits — for example, we don’t internally review or restrict publications. So far, five University of Washington researchers have accepted our model.

It’s not ideal. There are downsides to having academic researchers share their time with another organization. But the financial incentives are so powerful that if you don’t do something, many more people are going to go to industry.

doi: 10.1038/d41586-019-01250-2

This interview has been edited for clarity and length.

Nature Briefing

Sign up for the daily Nature Briefing email newsletter

Stay up to date with what matters in science and why, handpicked from Nature and other publications worldwide.

Sign Up