University AI researchers are being squeezed out of their own field

Key points

• Academic researchers cannot access the internal workings of frontier AI models like ChatGPT or Claude, which are kept private by their corporate owners.
• Universities cannot afford the specialised computer chips needed to train or run state-of-the-art AI, according to researchers at the 2025 Schmidt Sciences AI2050 convening in Mountain View, California.
• A Johns Hopkins study found that AI language models give less detailed answers to questions phrased in ways more commonly associated with women than men.
• Google DeepMind's AlphaFold team, which won a Nobel Prize for predicting how proteins fold, was disbanded last month.
• Some researchers worry that AI tools solving advanced mathematics could threaten the future of human mathematicians.

Picture being a marine biologist in a world where the entire ocean sits behind a corporate paywall. That is roughly how Nika Haghtalab, a computer science professor at UC Berkeley, describes academic AI research right now. She told MIT Technology Review that studying today's AI is like being a biologist locked out of CRISPR, the gene-editing technology that transformed medicine, because private companies owned it completely.

The comparison landed hard at a recent gathering in Mountain View, California, where the Schmidt Sciences AI2050 programme brought together some of the most respected AI researchers working outside the big tech labs. The programme, funded by Eric and Wendy Schmidt, supports academics whose work touches artificial intelligence.

Why can't universities just do the research themselves?

Money is the short answer. Training or running a frontier AI model requires enormous numbers of GPUs, the specialised computer chips that do the heavy number-crunching AI needs, and universities simply cannot afford enough of them. The AI2050 programme gives fellows some funding to buy chips, which researchers called a genuine lifeline, but federal science funding cuts are squeezing budgets further.

Even researchers who do not train their own models face costs. Repeatedly asking OpenAI's or Anthropic's systems questions in order to study them properly adds up quickly, and the companies do not open their internal designs to outside scrutiny at all.

So academics are finding different territory. Anjalie Field, a computer science professor at Johns Hopkins, put it plainly: "I try not to work on problems that I think are gonna be solved by a tech company." Her recent study found that AI language models give less detailed, less sophisticated answers to questions written in styles more commonly used by women. That kind of research, which can reflect poorly on a company's product, is unlikely to emerge from inside one.

What about AI research that has nothing to do with chatbots?

A large slice of academic AI has never touched large language models at all. These researchers build focused, specialised tools: models that predict extreme weather, simulate how drugs interact with the body, or analyse climate data. They are not chasing the same goals as OpenAI.

But they have their own problem. When most people hear "AI" they picture energy-hungry chatbots, and that image makes it harder for climate or health researchers to argue for their own work's value.

The field's talent pipeline is also shifting. Several well-known academics have taken leave from universities to join frontier labs, and many AI2050 fellows now hold both an academic post and an industry role at the same time.

There is one brighter note. The very lack of resources pushes academic labs to find ways to make AI models smaller, cheaper, and faster to run. Tim Dettmers, a computer scientist at Carnegie Mellon, argues that AI tools could make human scientists dramatically more productive rather than replacing them. Constraints, he suggests, breed creativity.

If the next real breakthrough in AI comes from a cash-strapped university lab rather than a trillion-dollar company, the people in that Mountain View hotel room will not be surprised.