Do we really want Facebook and Amazon to rule AI?

Do we really want Facebook and Amazon to rule AI?
© Greg Nash

A government commission is currently considering an innovation that could be as transformational for Artificial Intelligence (AI) as a hadron collider is for physics. It’s called a National Research Cloud, and right now the federal National AI Research Resource Task Force (NAIRR) is determining how we can develop such a cloud to broaden access to computing and data and spur basic and non-commercial AI research. At stake may be rates of investment in basic scientific research not seen since the days of the Cold War. 

The concept is straightforward. The federal government will provide access to computing and data resources needed for AI research that are becoming increasingly inaccessible to academic scholars. The best way to broaden access requires the federal government funding credits for researchers to access existing cloud computing power in the short term while it builds public cloud computing options for long-term use.

Yet critics are attacking this idea from both sides, putting at risk the potential for substantial innovation.

ADVERTISEMENT

Some Big Tech skeptics reject the very idea of a National Research Cloud, worrying it will increase the power of large technology companies, and about the harm AI poses to vulnerable communities. On the other side, free marketeers believe the government is taking too activist of a role and that we should rely exclusively on private technology companies — who offer a wide range of cloud services — to provide commercial cloud credits to academic researchers.

These objections are mistaken. The Stanford Institute for Human-Centered AI, where three of us hold appointments, led the call for the NAIRR in 2020. Over the past 10 months, we gathered an extensive research team to engage in an in-depth study of how to design the NAIRR and released our findings in a comprehensive report. Our research strongly counters these objections. 

To those concerned about the concentration of AI research power in private technology companies: We’re already there. Failing to create a National Research Cloud would only further consolidate private sector power in AI. 

A decade ago, Ph.D.’s with expertise in AI were equally likely to go into academia versus industry. Now they are twice as likely to join industry. This has led to a couple of problems. First, private sector research is subject to the direction, oversight and veto of tech companies, leading to research myopia. Facebook, for instance, dismissed its internal studies of how Instagram was toxic for teenage girls. Second, private sector research can be directed towards only a narrow field of applications. As data scientist Jeff Hammerbacher said, “The best minds of my generation are thinking about how to make people click ads.”

A National Research Cloud is a compelling way to address both of these problems, as it will broaden access to AI resources outside of the corporate context. It will expand who is able to develop, interrogate, or audit AI systems, going beyond narrow technical fields to include the physical sciences, social sciences and humanities. Liberating data in a privacy-preserving fashion about earth observation, labor markets and our justice system — today often accessible only to a select few — will also direct AI towards a more diverse set of pressing social problems. 

To those on the other side who fear government involvement more than private sector influence: As described, leaning on the private sector alone greatly hobbles innovation, but the public sector plays a critical role in spurring basic research and doing so cost-effectively. Consider the example of satellite imagery. Until 2008, the U.S. Geological Service charged some $600 per single satellite image. When it made the imagery free, it fueled the use of computer vision to study global warming, habitat modification, poverty and urban sprawl, generating an estimated $3-4 billion of annual benefits.

In the computing context, the federal government has deep experience building out computing facilities that operate at the cutting-edge, from Oak Ridge National Laboratory’s Summit system, which was the world’s most powerful supercomputer from 2018-2020, to the National Science Foundation’s investment in a high-performance computing network that contributed, along with the private sector, to major efforts such as the COVID-19 response. Many universities have also discovered that relying on commercial cloud services can be three to eight times as expensive as building their own systems. Such a public system can’t be built overnight, but it may be well worth the upfront effort and expense.  

Not only that, but America also desperately needs to build up a public sector workforce that is ready to use, monitor and regulate AI systems — and a National Research Cloud can make that possible. Legacy IT systems continue to plague the government. As the Government Accountability Office noted, the Department of Defense as of 2016 was “still using 8-inch floppy disks in a legacy system that coordinates the operational functions of the United States’ nuclear forces.” The NAIRR is an opportunity for the federal government to reset and rebuild and should not be viewed as a zero-sum game between techno-skeptics and markets-only crowd. 

On its current path, our AI future will rest increasingly in the hands of a few small, industry players. The National Research Cloud can correct this imbalance, broadening the set of voices that have access to one of the most important technological developments of our time. 

Daniel E. Ho, J.D., Ph.D. is the William Benjamin Scott and Luna M. Scott professor of Law and associate director of the Stanford Institute for Human-Centered Artificial Intelligence, Stanford University. Jennifer King, Ph.D. is a Privacy and Data Policy fellow at the Stanford Institute for Human-Centered Artificial Intelligence. Russell C. Wald is director of Policy at the Stanford Institute for Human-Centered Artificial Intelligence. Christopher Wan a J.D./M.B.A. candidate at Stanford University, is a coauthor on the report, “Building a National AI Research Resource,” and provided research and editorial assistance for this story.