Who benefits from a national AI program?

Earlier this year, the National Science Foundation announced the pilot version of its response to Silicon Valley’s AI boom. It’s called the National Artificial Intelligence Research Resource, and it’s supposed to “democratize” access to AI by making gigantic and expensive AI models available to academic researchers.

But Sarah Myers West, co-executive director of the AI Now Institute, is skeptical of the initiative’s goals.  

Marketplace’s Lily Jamali spoke with Myers West about the NAIRR and her recent op-ed in Foreign Policy magazine. As she explains, the issue that some critics have with the NAIRR is that the government can’t launch its own AI program without partnerships that are potentially lucrative for Big Tech.

The following is an edited transcript of their conversation.

Sarah Myers West: The tech industry now is tremendously concentrated, especially in artificial intelligence. It’s concentrated in both the access to resources needed to build AI, but also all the incentive structures in research and in where these resources get deployed. All roads kind of lead back to a small handful of large tech firms.

Lily Jamali: Let’s talk more about that. The pilot version of NAIRR brings together a number of government agencies like the Department of Energy, the Department of Defense, as well as big tech companies like OpenAI, Anthropic and Nvidia. Can you explain the structure of this pilot project?

Myers West: Yes, so the pilot project contains a few different components. One is it provides access for researchers who submit applications to be able to use government supercomputers. It also includes the ability to apply for access to government data sets that are held by agencies like the Department of Energy, or the [National Geospatial-Intelligence Agency] in order to build models using these data sets. And then there’s this third component, which is kind of like this marketplace of offerings from a number of different types of entities, most of whom are large tech companies who have said, we’ll give [application programming interface] access for a certain number of researchers to be able to build off of our models for free. But this often comes with some strings attached, where they’ve defined what types of researchers or what types of research they would want these resources to be used for.

Jamali: There’s so much money at play here. You write that in the current iteration of NAIRR, some of the biggest tech companies stand to benefit from this flow of taxpayer money into this very concentrated market for AI. Can you first explain that argument?

Myers West: Yes. So, if you look to the original NAIRR proposal, the version that might be launched after the pilot, it’s structured as a set of, essentially, licensing contracts that would go to cloud computing firms. Now, there’s not that many cloud computing firms out there, and the largest by far are run by big tech companies. It’s Google, Amazon and Microsoft. And so, the eventual likelihood is that any investment of taxpayer money is ultimately going to flow to the companies that have essentially monopolized cloud computing and access to computational power. If you go one layer down, that’s Nvidia, the largest chipmaker.

If the idea is that artificial intelligence is meant to be at the forefront of innovation in this country, is our vision for the future going to be clouded by these companies’ profit motive? And how do we change that? I don’t think that we will get there if all of the incentive structures continue to be set by the same handful of large companies that are already the incumbents.

Jamali: When you hear the National Science Foundation say the NAIRR will democratize AI, what do you make of that?

Myers West: Well, I think when we think through what the end goal is here, we’re going to need a better litmus test than just democratizing the sector. Instead, I think we really need to be asking questions that probe who benefits from the investments in this industry and in AI. And is that going to benefit society at large? I think that there’s evidence, if you look back to the ways that AI is already in deployment and already affecting people’s lives, often it’s in ways that either enable austerity measures — they make things cheaper, they often result in the devaluation of work or depressing wages — or alternatively, they ramp up control, whether that’s in educational settings or in workplaces through increasing surveillance. Are those the ends that we want to be really investing in as a society? And, really, more fundamentally, if you are to democratize AI, does that reach the needs of the many? Or does that democratization happen within the hands of maybe just a slightly larger handful of companies?

Jamali: I’m going to characterize that answer as skeptical. What do you think might be a better way to actually democratize access to AI, then?

Myers West: Well I think, like I said, I’m not sure that democratizing access is necessarily the right end goal here. I think it’s creating space for a broader public conversation about what kinds of AI do we want to build. And what kinds of societal problems do we want to solve in which AI might play a role? Starting from the underlying question of like, what’s the challenge that we’re facing here? And is AI the right solution? Rather than starting from the presumption that AI is inherently going to work to the good of society. Because I think we have some examples where it certainly can, but we have just as many that it’s currently not.

Jamali: But it feels like the train has left the station on AI. Are you kind of saying that maybe AI is just not the answer? Because it feels like we can’t really stop the trajectory.

Myers West: It certainly feels that way and especially this year, after this tremendous amount of excitement but also anxiety about how AI is going to shape our future. But it’s important to keep in context, AI has an almost 70-year history, and what “AI” even means has changed over time. There’s still a lot of scope to reshape the incentive structures and the goalposts of this industry and the field at large. And I think starting from the perspective that there’s really nothing that is inevitable about AI, the future very much can be shaped and needs to be shaped by all of us who will be affected by it.

More on this

To be sure, the purpose of the NAIRR isn’t just to altruistically give researchers access to high-tech resources. As Sarah Myers West points out in Foreign Policy, for the U.S. government, it’s about global competition.

In 2021, the National Security Commission on Artificial Intelligence put out a report calling on the government to step up its AI investment or risk getting lapped by China within the decade. The report argued that the way to win is by having the government and private sector join forces. Sounds like pretty much exactly what’s happening with the NAIRR initiative.

This NAIRR pilot, however, is still just that. A pilot. It was established when President Joe Biden signed his executive order on AI last fall. But with a budget of $140 million, the pilot is operating on something of a shoestring by government standards.

As Myers West puts it, it’s a “small-scale” version of a much bigger plan that would cost $2.6 billion and require approval from Congress before it sees the light of day.