Academic researchers know that artificial intelligence (AI) technologies have the potential to revolutionize the technological landscape of nearly every industry. Although researchers are trained to apply these innovations in ethical and fair ways, they have limited access to the expensive and powerful technologies required for AI research compared to commercial technology companies.
This disparity has academics and government-funded researchers worried that developments arising from the AI gold rush could leave marginalized people behind.
For example, a radiologist could use a generative AI agent to read x-ray images, theoretically leading to more accurate diagnoses and better health outcomes. But if that AI agent is only trained on data from hospitals in wealthy areas, it may not be able to recognize signs and symptoms that are more common in lower-income areas.
Wealthy populations “may have a fundamentally different distribution of symptoms that doesn’t necessarily match up with, say, people who have a harder time getting regular doctor’s care,” said Bronson Messer, science director at the U.S. Department of Energy’s Oak Ridge Leadership Computing Facility in Tennessee, which is home to Summit, one of the nation’s most powerful publicly funded available supercomputers, which some academics use for AI research.
“There is a lingering concern that the data being used to train generative AI may have inherent biases that are nearly impossible to discern until after the fact, because generative AI agents can only interpret what they are given.”
Resource disparities
Eliminating these biases is one of the overarching goals of the National Artificial Intelligence Research Resource Pilot (NAIRR), which the National Science Foundation (NSF) helped launch in January.
“This is a hot topic, and the U.S. academic community is in the best position to bring it to light,” said Messer, a member of the NAIRR Allocation Working Group. “I don’t want to leave this to Meta or Google. This is an issue that should be discussed in the public literature.”
The NAIRR pilot is the result of President Joe Biden’s executive order on the safe, secure, and trustworthy development and use of AI, and NSF is leading the pilot in collaboration with the Department of Energy, according to a news release.
Through the two-year pilot program, 77 projects, the majority of which are affiliated with universities, have been allocated computing and data resources and services, including remote access to Summit and other publicly funded supercomputers. The two-year NAIRR pilot program prioritizes projects focused on using AI to address “societal challenges” in areas such as agriculture and healthcare.
University researchers have long been at the forefront of innovation in these and other fields, but access to the increasingly complex infrastructure needed for AI-driven research and development (known as AI computing) is expensive and highly concentrated in private technology companies like Open AI and Meta in specific regions like the Bay Area, New York City, and Seattle.
Compared to tech companies, as a recent Brookings Institution paper put it, even the nation’s best-funded research universities lack anywhere near the computational power needed to “query, fine-tune, and train large-scale language models (LLMs) to drive unique advances,” and smaller institutions, most of which are located far from major tech hubs, have even fewer resources and expertise to conduct AI research.
To put it in perspective, Reuters reported in April that Meta planned to amass about 600,000 cutting-edge geographic processing units (GPUs) — computer chips that power AI applications — by the end of the year. But Summit, at Oak Ridge National Laboratory in Tennessee, has about 27,000 GPUs.
“A fundamental democratic issue”
The potential impact of disparate AI research resources is “a fundamental question for democracy,” said Mark Muro, a senior fellow at Brookings Metro who specializes in the interplay between technology, people and places. “If AI is going to be a big driver of productivity, and that’s true, then it’s a problem if only a handful of places are really benefiting from that economic impact.”
The same is true for choosing what research questions to investigate and where and how to pursue those research questions.
“It may end up being a limited research agenda selected only by large technology companies,” Muro said.
And those companies may not have the interest or financial incentive to tackle local problems, like a particular health crisis or forest fire management. “Those problems could really be activated by an AI solution, but you might not have someone researching them there. But there are times when place-based solutions are really important.”
Like the upper echelons of the academic scientific research community, the tech industry is dominated by white men, raising concerns about racial and gender bias.
“If we don’t increase the representation of people in AI research, including more minority-serving institutions and HBCUs, we’re only going to exacerbate the problem of a lack of diversity in the field,” said Jennifer Wang, a computer science student at Brown University who co-authored a paper with Muro on disparities in AI research published by the Brookings Institution earlier this month.
“Currently, a lot of AI research is focused on developing better, more performant models, with little attention paid to biases within these models,” Wang said. “Because these models are not built with specific populations in mind, little consideration is given to capturing linguistic nuances and cultural context.”
Democratizing AI research is one of the main goals of the NAIRR pilot, which is scheduled to run for two years, but the directors of both NSF and the Office of Science and Technology Policy have expressed hope that NAIRR will be sufficiently funded to continue beyond that.
Several well-known research universities have received funding, including Brown University, Harvard University, and Stanford University, but lesser-known research institutions, including the University of Memphis, Florida State University, and Iowa State University, are also participating in the NAIRR pilot.
The Iowa State project aims to use the Frontera supercomputer housed at the Texas Advanced Computing Center at the University of Texas at Austin to develop “large-scale vision-based artificial intelligence tools to identify agricultural pests and ultimately recommend control measures,” according to a university news release.
Equitable Access
But without the support of the NAIRR pilots, the rocket may never have launched.
“Academic research is often in its early stages and has the flexibility to focus on societally important problems that industry may not currently be interested in, ensuring that important issues like agricultural resilience receive the attention they deserve,” Bhaskar Ganapatisubramanian, a professor of engineering who is leading the project and director of Iowa State University’s AI Institute for Resilient Agriculture, said in an email. “This allows academic research to prioritize the public interest over commercial gain and to focus on ethical considerations and long-term societal impacts.”
The NSF expects “the next set of projects will be announced soon, and roughly monthly thereafter,” as long as resources are available, Katie Antipas, director of NSF’s Advanced Cyberinfrastructure Office, said in an email.
But given that Congress cut NSF’s fiscal year 2024 budget by 8 percent, it’s unclear whether NAIRR will become a permanent home for the academic research enterprise.
“We see great potential in the two-year NAIRR pilot,” Julia Jester, vice president for government relations and public policy at the Association of American Universities, said in an email, “but NSF and other agencies will need much more resources to achieve the project’s goals of broadly improving access to the AI infrastructure needed to advance research and train the next generation of researchers.”
Suresh Venkatasubramanian, a professor of computer and data science at Brown University and director of the school’s Center for Technology Responsibility, has just launched the NAIRR pilot project to develop tools to increase transparency of the data used in law school training.
As AI technology begins to permeate research, business, medicine, and every profession, it is important to fully understand its impact.
“It’s really important that higher education institutions across the board embrace advances in AI, learn about AI, use AI in practice, and reimagine how we use AI beyond what we’ve imagined, and beyond the solutions that people at tech companies are providing,” Venkatasubramanian said. “And that can’t happen without equitable access to core computing units.”
