Sections

Commentary

The case for promoting the geographic and social diffusion of AI development

AMES, IA, USA - NOVEMBER 1, 2022: Student Innovation Center on the campus of Iowa State University.
Photo credit: Ken Wolter / Shutterstock

Last fall, we argued that the AI sector—especially large language model (LLM) activity—is being driven by concentrated research, modeling, and design work occurring in just a handful of “superstar” tech centers along the coasts: the Bay Area, Seattle, New York, and Boston.

In that piece, we suggested that such concentration is a negative force in shaping AI development and, more broadly, U.S. economic growth.

Still, it is legitimate to ask why such geographic concentration is a problem, and why broader diffusion of this activity into more U.S. regions would be beneficial. Answering this question is particularly important today, as new global and national AI safety institutes seek to identify and address the technology’s key risks while mostly leaving aside the need to more broadly distribute AI development.

To be sure, many scholars have made the case for concentrating innovation into core hubs. Research dating back to Alfred Marshall in the 1890s has indicated that sector growth may be more efficient and rapid when it occurs within concentrated geographic areas of innovation. More recently, Harvard Business School Professor Michael E. Porter developed what he calls the “diamond of national advantage,” which articulates how dense clusters of economic activity maximize the performance of industries. And today, our work and the work of others highlight the super-concentration of core AI development.

Figure 1

Despite the benefits of clustering, though, strong arguments suggest the value of more widely diffused AI development. Productivity, technology diversity, and socioeconomic equity and opportunity all come into play.

First, widening AI activity could capture currently unrealized opportunities for productivity growth across major industries. American industries will clearly benefit if the forecasted efficiency gains that AI may deliver could be deployed more broadly into more industries and regions. Conversely, it will be important going forward to convert sector- and region-specific knowledge and use cases into unique local and national breakthroughs. A case in point is Pittsburgh-based U.S. Steel’s collaboration with Google Cloud on its MineMind equipment maintenance system, developed using generative AI informed by data and insights from the company’s Minnesota ore operations. The initial phase of MineMind, beginning in September, will impact 60 haul trucks, potentially reducing the time to complete work by 20%. This can have an economic multiplier effect, spurring new job growth and economic activity.

A second reason to widen the geographic spread of AI development reflects the need to incorporate a broader range of perspectives in work to develop “safe” and “aligned” artificial intelligence. When AI development and underlying data are concentrated in a few high-income urban centers dominated by a homogenous group of individuals, there is a risk of creating systems that are misaligned with the values and needs of more diverse populations and geographies. As ProPublica has shown, AI models developed in Ohio offered flawed prediction rates of recidivism for offenders—disproportionately penalizing Black Americans in New York, Wisconsin, California, and Florida. Meanwhile, developments in the medical care sector have demonstrated that geographically narrow AI training datasets can exacerbate biases and limit the diffusion of AI’s benefits. Such geographic biases make the models perform worse and limit their generalizability, in part because geography co-varies with many important characteristics and variables in predictive modeling. In a review of radiology and biomedical research, one study found that 71% of patient data used to train deep-learning diagnostic models came from California, Massachusetts, or New York. Since demographic and health care characteristics vary considerably across geography, the regional concentration of training data can lead to worse outcomes for underrepresented populations and an exacerbation of health care disparities. For example, AI-driven medical interventions may be tailored to the profile, genetics, and needs of individuals in rich coastal cities, and provide lower-quality care for those outside of them.

Finally, a third benefit of closing the nation’s emerging AI gaps is the possibility of narrowing the nation’s stark opportunity gaps across places. To the extent that AI’s development repeats the familiar “superstar city” dynamics of previous digital adoption waves, it will exacerbate the tendency for some communities to fall behind and devolve into what economic geographer Andrés Rodríguez-Pose calls “development traps.” Conversely, to the extent that AI is source of productivity gains and growth rather than just the automation of work, the widening of its geography could give regions that have previously fallen behind access to the potentially transformative effects that come from leveraging new activities and industries.

The potential for ‘AI divides’ and the benefits of narrowing them argue for action

The unbalanced early buildout of AI appears likely to create divides that entail long-term problems. So, the nation has every reason to counter such divides.

Indeed, the federal government is beginning to do just that. To start, the Biden administration’s broad economic development push has frequently targeted strong flows of “place-based” investments toward “left-behind places”—in several instances, places advancing AI-related economic transformation strategies.

The Economic Development Administration’s (EDA) Build Back Better Regional Challenge (BBBRC) is an example. Through one of its place-based challenge grants, the program is funneling $65 million into an ambitious push to diffuse cutting-edge AI solutions into Georgia’s manufacturing sector. The emerging Georgia AI Manufacturing (GA-AIM) coalition is led by Georgia Tech, and involves an ambitious statewide investment surge to support AI talent development and worker training, particularly for underserved communities and businesses. Similarily, the BBBRC is channeling another $63 million into the Southwestern Pennsylvania New Economy Collaborative, which supports local AI and robotics workforce training and development, especially in urban areas and rural portions of the region.

More recently, the EDA’s much-watched Regional Technology and Innovation Hubs (Tech Hubs) program made designations or awards to regional initiatives focused at least in part on AI in places such as Wisconsin, Birmingham, Ala., and Minnesota.

Beyond these efforts that incorporate AI development into broader industrial policy, there also exist specific AI diffusion initiatives that provide compelling models for narrowing the nation’s AI gaps. Along these lines, several place-focused AI research programs provide examples of how federal initiatives can counter excessive AI unevenness and broaden the technology’s development.

A case in point is the the National Science Foundation’s (NSF) National AI Research Institutes program, which is based at universities across 37 states and invests almost $500 million over five years. These institutes provide and exemplify how targeted investment can help regions leverage existing competitive advantages. The AI Institute for Resilient Agriculture is based out of Iowa, and leverages machine learning to strengthen the state’s large agriculture sector. Meanwhile, the National AI Research Institute for Climate-Land Interactions, Mitigation, Adaptation, Tradeoffs, and Economy (AI-CLIMATE) is based out of the University of Minnesota Twin Cities and aims to leverage AI to lift rural economies through smart agriculture and forestry.

To promote further inclusion in these awards, earlier this year the NSF announced it was launching the ExpandAI program, which specifically targets institute awards focused on capacity-building and partnerships at minority-serving institutions. 

Other federal initiatives are on the horizon, most notably the NSF’s National AI Research Resource (NAIRR), which aims to democratize access to an array of essential computational and data resources, testbeds, software, testing tools, and user-support services. Now being developed as a pilot program with contributions from 12 other agencies and 26 corporations, NAIRR is focused on providing AI computing power, access to data sets, and training resources to new people and places. What is needed now is generous funding to make NAIRR a robust force for AI inclusion.

In sum, there are now both strong arguments for the benefit of widely diffused AI development and promising experiments at how to achieve just that. It seems like exactly the moment to begin laying the groundwork for a broader, more inclusive AI economy.

Authors