How the Tug-of-War Between Industry and Academia is Shaping AI Research

fter nearly a decade of teaching and researching computer science at Harvard, Boaz Barak began to encounter questions he believed he could better answer from outside academia. So, in January 2024, he stepped away from his tenured professorship to join OpenAI.

The catalyst? A growing concern about safety mechanisms for artificial intelligence — or the lack thereof.

Barak finds the mismatch between the technological improvement of artificial intelligence and society’s preparation for it “scary.” While he feels like there are plenty of people focused on AI’s expansion and integration into the economy, there may not be “enough people that are focused on making sure that it’s safe and making sure that it goes well for humanity.”

Working at OpenAI, he believed, was a chance to further his knowledge in a hands-on way he couldn’t do as a full-time professor. As a member of OpenAI’s technical staff, Barak had the opportunity to work directly on safety and alignment for “models that are used by many, many millions of people.”

Barak is far from alone in his pursuit to maximize the impact of his work. The AI hype and fear has pushed both students and professional academics to weigh the merits of two paths: pursuing research in academia, or moving to industry.

Rachit Bansal worked at Google DeepMind before pursuing a Ph.D. in computer science at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. The choice between academia and private sector research is one that he and his colleagues are “constantly thinking about.” David Alvarez-Melis, an assistant professor of computer science at Harvard, says 90 percent of the Ph.D. students he advises are seriously considering both. And in terms of practicality, the private sector has a number of draws.

“It’s always on your mind just because it’s so financially advantageous to go into industry right now,” says Timothy Ngotiaoco, a senior machine learning research engineer at the Kempner Institute. “People are really racing to build the best model possible, and they will pay a ton to get researchers.”

T. Anderson “Andy” Keller, who is finishing up the third year of his postdoctoral fellowship, says it can be hard to balance your research interests — which are not always represented in tech companies’ projects — with the realities of “what your financial situation requires.”

“If you’re offered many hundreds of thousands of dollars a year or something like that, it can be very hard to turn that down,” Keller says.

In addition to the monetary draw, many researchers need to factor in whether the computing power available in academia will serve their research interests. Some research questions, such as the evaluation and benchmarking of large language models, require large amounts of computational resources or closed-source data that universities simply do not have.

Such resources can be measured in graphics processing units, the specialized circuits used to train and run many AI models. While Elon Musk’s xAI projects that it will reach 1 million GPUs by the end of 2026 (estimated to cost tens of billions of dollars), Harvard researchers have access to just one-thousandth of that. The Kempner Institute, for example, is home to one of the largest academic clusters in the world and provides 1,144 GPUs.

“If you were Google or Meta or OpenAI, you’re probably doing a single training run on like, 20,000 GPUs or something like that,” says Ngotiaoco. “That’s almost 50 times more than what I work with.”

Not everyone sees academia’s limited resources as a hindrance. Alvarez-Melis, the computer science professor, says the discrepancy in computing power forces his students to be “more creative and smarter in the questions we ask,” leading to new opportunities and discoveries.

“Often, I think that some of the most interesting developments in not only machine learning but in science come from thinking about constraints and working around constraints,” he says.

While academic and industry research sometimes overlap, Alvarez-Melis says that certain topics have a “more natural home in academia” because they require fewer resources or wouldn’t fit into a company’s objectives. As a result, many still gravitate towards academia for the intellectual freedom it provides. Multiple researchers say they enjoy the flexibility to explore the topics they are most interested in, rather than having a company’s priorities govern the work they do.

For instance, Kyran Romero, a post-baccalaureate fellow at the Kempner Institute, chose to pursue academic research because it gave him the flexibility to work on technologies that address resource gaps in indigenous communities.

“For something as niche as indigenous community technology,” Romero says, “there isn’t really a lot of market share or value incentive for companies to devote teams of research to that.”

But the hesitation to join major tech companies is also rooted in ethical concerns. Some of these corporations now work closely with the U.S. government, influencing national security in a way that gives some researchers pause.

For example, in 2018, more than 3,000 Google employees protested the company’s involvement in Project Maven, a Pentagon project seeking to deploy AI tools to automate warfare. Though Google did not renew its contract to work on Project Maven, the company, along with multiple other tech giants, has maintained ties with the federal government and its military.

“For a lot of these AI frontier labs to remain at the frontier, they often engage in certain contracts with maybe the Department of War,” says Ada Fang, a Kempner Institute graduate fellow who is also pursuing a PhD in chemistry. She says there are “definitely a lot of ethical issues surrounding what the frontier AI models are being used for.”

Fang, whose work mainly focuses on using AI for biology, says her main objective is to progress modern medicine with innovations. But she realizes these could also be used to develop bio-weapons. She hopes that people within industry and academia remain “aligned in the sense that this is not something that we should ever pursue.”

For most Faculty of Arts and Sciences positions, however, policy mandates that “no more than 20 percent of one’s professional effort may be directed to professional activities outside Harvard.” But within Harvard’s School of Engineering and Applied Sciences department, many researchers are finding ways to pursue both industry and academia.

A dozen of the 43 SEAS tenure-track professors affiliated with computer science have held industry positions within their past five years at Harvard. Faculty say SEAS Dean David C. Parkes has actively encouraged this trend, offering part-time arrangements for departing professors to remain tied to teaching while pursuing industry connections. These part time jobs have been given the name “Catalyst professors,” and entail five-year, often untenured appointments that allow for mentorship and research while remaining more flexible than a standard tenure track role.

Barak, for one, took three semesters off before returning to Harvard as a part-time professor, and now teaches only two computer science classes — a course on AI safety and a seminar on effective research practices and academic culture — while remaining on OpenAI’s technical staff. With these dual roles, he feels that he can both directly influence the trajectory of AI while also passing technical lessons onto his students.

As a Catalyst professor, Barak says he can balance his love for teaching with his work in AI safety. He doesn’t believe every faculty member should be part-time — students would lose the “focus and commitment” of a full-time faculty member — but if no one was a part-time professor, he says students would lose an important opportunity to learn from someone actually working in the industry. He thinks his current AI safety course would be more difficult to teach without practical experience.

“I think that I’m doing the most good where I am right now,” he says.

But Kempner research fellow Naomi Saphra, an incoming assistant professor at Boston University, says her colleagues moving to industry is a concerning trend.

“That really is bad for the students,” Saphra says. “Students now have to start thinking about the attrition rate of the faculty at their department when making these decisions about where to go for a PhD”

Moreover, Saphra critiques companies for rarely allowing researchers to publish their work because the competitive AI market discourages the sharing of ideas.

Researchers themselves are aware of this reality. “Most of the universities will want to keep the intellectual property or make it publicly available, whereas companies obviously want to keep it for themselves,” Keller says.

Concerns regarding proprietary information and competitive advantages can also make intellectual exchange between industry and academia challenging. Romero recalls a researcher from Google DeepMind who came to Kempner as a guest speaker but would only speak in generalities about his work. “The second we started to want to get into the meat or nitty gritty, they were kind of like, ‘oh, it’s trade secrets, I can’t really expose more,’” he says.

Nevertheless, tech companies and frontier AI labs remain promising environments for researchers looking to make contributions to open science. The 2017 research paper “Attention Is All You Need,” that laid the foundation for most of today’s large language models, was produced by a group of Google researchers. At NeurIPS 2025, a premier conference for machine learning research, Google DeepMind and Meta were the top contributors from the United States in terms of the number of accepted papers. Stanford University, Carnegie Mellon University, and Microsoft followed.

Barak says previous breakthroughs in science often happened in academia or government labs. But “here, it really is the case that a lot of it is happening in industry.”

Some problems are only revealed when creating models at a large scale, Fang says, skewing the possibility of new discoveries towards those with industry resources. Still, Fang argues that those discoveries can later be explored by academics who “wouldn’t have otherwise thought about” large-scale challenges.

As AI continues to evolve, many researchers remain hopeful that the two fields can work in tandem to define its future. “I think that it is important that these two fields, these two domains, stay intertwined,” Fang says.

—Staff writer Megan L. Blonigen can be reached at [email protected] and on Signal at megblon.50. Follow her on X at @MeganBlonigen