Academic AI research challenges were on display at a recent Schmidt Sciences AI2050 gathering in Mountain View, California, where researchers described a field in which the biggest language models are expensive to run and their workings are largely controlled by private labs. University AI researchers make up most of the AI2050 group, according to MIT Technology Review's account of the meeting.
The immediate problem is access. MIT Technology Review reported that universities often cannot afford the GPUs needed to train and run frontier models. Anthropic and OpenAI also do not disclose the design and training details of Claude and ChatGPT to outside researchers. That leaves academics able to test how a model behaves from the outside, while unable to conduct detailed research on how it was built or direct its training.
Why are academic AI researchers struggling to study frontier models?
Frontier models are costly to develop and operate, and commercial developers keep key technical details private. A large language model generates responses from learned statistical patterns, rather than retrieving a prepared answer, as explained in this guide to how LLMs work. For outside researchers, even the narrower task of repeatedly sending prompts to commercial systems to evaluate them rigorously can become prohibitively expensive, MIT Technology Review reported.
Nika Haghtalab, a UC Berkeley computer science professor, compared the arrangement to scientists trying to study a crucial biotechnology tool controlled exclusively by companies. The comparison captures the research constraint, not a claim that outside study is impossible: researchers can measure outputs, but not inspect the underlying design and training process in detail.
AI2050 provides one route around the resource problem. The Schmidt Sciences initiative supports researchers working on questions meant to help make AI beneficial to society, and its senior and early-career fellowships back three-year projects, according to AI2050. MIT Technology Review reported that fellows can use some program funding to buy GPUs, though researchers at the gathering still described funding as a pressure.
Where are university labs focusing instead?
Some researchers are choosing questions that do not fit a commercial product roadmap. Johns Hopkins computer science professor Anjalie Field told MIT Technology Review she avoids work she expects technology companies to solve. Her study found that language models gave less sophisticated answers to prompt styles more commonly used by women. The reported result is an example of research on model behavior and disparities, rather than evidence about every model or prompt.
Academic AI is also broader than the contest to build the next general-purpose chatbot. Many researchers build specialized systems that analyze data, make predictions or simulate physical processes. Participants working on climate-related applications said public discussion that treats AI as synonymous with energy-intensive language models can make it harder to make the case for their work, MIT Technology Review reported.
The account also described some prominent academics taking university leave to join frontier labs, while many AI2050 fellows hold industry jobs alongside academic posts. Those examples show career choices researchers are confronting; they do not establish how common such moves are across academia.
Concerns about AI's effect on pure mathematics have added to the unease, though the evidence here is competing views rather than a settled outcome. Carnegie Mellon computer scientist Tim Dettmers argued that AI systems for science could make researchers more productive. MIT Technology Review also noted the case for academic work on smaller, more efficient models and new architectures when labs cannot match corporate resources. For empirical fields, collecting real-world data may remain a slower constraint on automation.
A separate 2025 report from the American Association of University Professors points to a governance problem inside higher education. Its two-week survey of 500 AAUP members from nearly 200 campuses found calls for AI training, shared oversight, equity, transparency, opt-out options and worker protections. That survey covers academic professions and campus AI tools broadly, not only researchers working on frontier models, but it found many respondents had little input into how institutions procure and deploy AI systems.
This story draws on original reporting from MIT Technology Review.
