Few technologies in history have matched the potential of artificial intelligence to transform the economic landscape in which they emerged. Deployed broadly and effectively, AI can make workers more efficient, enable more dynamic entrepreneurship, and lift economic growth, all of which could be necessary bulwarks against the powerful economic headwinds brought on by aging populations and mounting public debt. AI can also make possible revolutionary new advances in science and medicine and democratize access to knowledge once reserved for the select few.
But the acceleration of AI also carries risks. It could be applied in ways that displace workers, disrupt how students learn, create new bioweapons and cyberweapons, and strain current supplies of energy, water, and land. Compounding these challenges is a mismatch: the benefits of this new technology may take a long time to become visible and widespread. As with earlier transformations, such as the Industrial Revolution, people are likely to experience the disruptions that come from adopting AI before they feel its benefits. And whereas users directly feel the effects of tools like chatbots, some benefits that AI is indirectly making possible, including more accurate predictions of natural disasters and improved medical diagnoses, may not be perceived as owing to the technology at all.
Public opinion has already begun to turn against AI, especially in the United States. One of the main targets has been data centers, with polls showing that a significant majority of Americans oppose the construction of data centers in their communities. Opposition to data centers needed for the new AI economy now cuts across party lines and age groups. Numerous surveys have found that residents worry about data centers using too many resources, creating endless noise, and raising their utility bills, all while developers benefit from public subsidies at the communities’ expense.
Although some of these problems are local, they also reflect wider concerns about AI. Some leading figures in AI have no doubt contributed to this anxiety with dramatic warnings of a white-collar-job apocalypse and even the possibility that AI will lead to human extinction. These are real issues that must be understood and addressed, but the growing public pushback against AI also suggests a deeper distributional challenge: individuals, communities, and businesses are being asked to help shepherd a transformative and potentially unsettling new technology at a time when many citizens see its financial gains going to a few people whom they increasingly distrust. AI is now a major issue in American politics: candidates running in competitive elections, no matter which party they belong to, appear increasingly averse to being associated with support for AI.
These anxieties are understandable. AI adoption risks repeating the mistakes of globalization, which generated large and widely spread economic benefits but led to concentrated losses of jobs, growth, and industry in particular communities. Government support in response was insufficient or arrived too late. Although the economy gained overall from free trade, the political bargain failed to give affected workers and communities a credible stake in the gains, even as individuals benefited as consumers from lower prices. The result was a backlash that contributed to the protectionism and fragmentation that continue to plague today’s politics and society.
Turning AI’s capabilities into broadly shared and enduring prosperity requires more than investing in the technology itself. The central challenge of the AI era is coordination and cooperation across groups responsible for the many complementary investments that promote the adoption of AI, including technological diffusion, worker empowerment, firm reorganization, and infrastructure construction. Because these complementary investments must be made across different institutions and at different times, progress requires market mechanisms to encourage cooperation as well as coalitions among organizations, firms, workers, communities, and governments that can turn AI’s technological gains into widespread economic and social benefits. To borrow a concept from game theory, AI is a “cooperative game” in which the largest gains arise only when many groups coordinate, and those gains persist only when every group expects to do better as part of shared coalitions than outside them.
But coalition building requires trust, and public trust—particularly regarding technology and who benefits from it—has eroded. Grand promises of scientific and medical breakthroughs will not overcome deep-seated hostility. Durable cooperation requires giving affected groups a concrete stake in the opportunities and gains that AI generates rather than merely promising future benefits. Without that stake, the individuals and groups whose support is needed to make AI flourish could block its progress. Governments could impose stifling regulations on the industry or stop issuing data center permits, firms could decline to adopt new AI-driven systems, and workers could avoid sharing the tacit knowledge needed to make the technology useful. As the economist and Nobel Prize laureate Michael Spence once observed: “You cannot solve a cooperative game uncooperatively.” The way to ensure that society benefits from AI’s enormous potential while mitigating its risks is to make cooperation the clear choice.
A WHOLE NEW WORLD
AI is staking a claim to being the next general-purpose technology—one that applies across a wide array of activities, improves over time, and generates complementary innovations. Electricity was an early general-purpose technology. When inventors discovered how to commercially harness electricity in the late nineteenth century, the ability to use it to power many types of machines transformed industrial production. Computing, another general-purpose technology, used electricity to process information itself, revolutionizing sectors as varied as engineering and health care.
But AI works through a more powerful input—intelligence—which makes it possible to interpret information, predict outcomes, diagnose problems, communicate with others, and make decisions. As a result, AI can raise productivity in existing sectors: a nurse can spend less time on paperwork and devote more time to patients, engineers can more quickly develop new designs, and small firms can scale up their analytics capabilities to compete with larger competitors. It can also increase efficiency so significantly that it makes possible wholly new products and services that were otherwise too costly or complicated to deliver.
AI also goes beyond existing general-purpose technologies by transforming the process of research and discovery. By inventing new ways of optimizing scientific processes, AI opens up previously unknown possibilities, similar to what microscopes and telescopes did in earlier eras. AlphaFold, a specialized AI model for which Demis Hassabis and John Jumper shared half the 2024 Nobel Prize in Chemistry, is a clear example. Determining a single protein’s three-dimensional structure from its amino acid sequence used to take months or even years; by using machine learning, AlphaFold has predicted the structures of more than 200 million proteins, fundamentally reshaping what biologists are able to investigate.
The potential of AI is thus a combination of the Industrial Revolution, whose general-purpose technologies transformed production, and the Enlightenment, which brought about inventions for making inventions—the new instruments and new paradigms that transformed the process of discovery. AI holds the promise of providing both the productivity boost and the tools to pave the way for discoveries that can address many of the problems weighing down aging, debt-ridden societies across the world.
GOING DEEP
The latest wave of AI has diffused at unprecedented speed. Google’s Gemini app and OpenAI’s ChatGPT each now reach more than one billion monthly users. Google’s AI & Economy ATLAS, a study of how people are using the company’s AI products that two of us (Imas and Manyika) worked on, analyzed 15 million interactions in more than 150 countries. It found that AI was used in tasks in 68 percent of occupations, and these occupations made up roughly 90 percent of all U.S. employment. Even automotive technicians and industrial mechanics used AI to interpret test results and diagnose electrical faults. Another study that two of us (Imas and Manyika) ran in collaboration with MIT FutureTech found that nearly half the scientists who were surveyed used some type of AI model every day, saving them time and enabling more interdisciplinary research.
But the gains of diffusion are neither guaranteed nor automatic. AI has been adopted broadly across and within sectors, but thus far its penetration remains fairly shallow. The ATLAS study found that only 21 percent of tasks in a typical occupation have been substantially affected by the technology. According to the Anthropic Economic Index, which measures how the AI model Claude is used in workplace tasks as varied as writing software, analyzing data, and managing administrative workflows, potential use far exceeds actual use. This suggests that access has spread much faster than adoption. Incorporating AI deeply enough to allow it to live up to its potential requires a bigger transformation. A firm does not change how it operates simply by buying access to an AI coding agent. It must also invest in organizational reforms and worker training and often needs to redesign its governance structure or its product lines.
AI adoption risks repeating the mistakes of globalization.
The banking sector, for example, has widely adopted AI but mostly used it to perform superficial tasks. As it is currently deployed, AI gives analysts a faster way to research companies, populate spreadsheets, and identify the sources of investment returns. Deep adoption, on the other hand, requires changing workflows and job responsibilities. A bank must decide what proprietary data to feed a model, how to protect client information, and how to verify the results of the AI models it uses. Junior employees may spend less time collecting information and more time interpreting it, forming ideas, and advising clients. But many tasks that AI can automate are also forms of training. Firms will therefore need to find ways for experienced employees to help younger colleagues acquire the skills the experienced employees once developed by producing the underlying analyses themselves.
The true economic promise of AI lies in deep and transformational adoption. This requires revamping organizational processes, preparing workers to capitalize on the new technology, and applying AI in ways that enhance productivity and innovation and improve consumer welfare. Innovations that increase output and can generate new products and services have not yet appeared at a meaningful scale. But there is enormous latent demand for many new products and services, including specialized services such as individualized tutoring, health care, and legal and financial assistance, and additional demand will emerge from completely new activities that technology will make possible. By fundamentally changing how work is done and what work can be done, AI can lower the cost of providing tailored products, making new opportunities economically viable. Serving demand that previously did not make economic sense adds new value to society rather than merely redistributing a fixed economic pie.
But enlarging the economic pie does not guarantee a smooth transition for every worker. As with globalization, the new demand that AI unlocks will not necessarily arise in the same jobs, sectors, or communities in which existing work is done. This potential mismatch risks major labor market disruption and political backlash. Public anxiety is already growing, especially among younger workers seeking entry-level positions that are the most likely to be automated by AI. One study by the Stanford economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen found that the employment of 22-to-25-year-olds in occupations most exposed to AI automation was 19 percent lower than it would have been if employment had grown at the same rate as occupations less exposed to AI disruption. Yet the overall effect of AI on jobs, including entry-level positions, remains uncertain. The Economist has estimated that AI has created around one million jobs in the United States but led to about 200,000 layoffs, and the gap between the unemployment rate among 20-to-24-year-olds and the overall unemployment rate is low compared with recent decades. Most likely, as one of us (Manyika) wrote with Spence in these pages, in 2023, some jobs will be lost, some will be gained, and many more will be permanently changed.
INVENTIVE INCENTIVES
Harnessing AI’s potential to increase prosperity without destabilizing society requires coordination. In addition to AI companies making new models, products, and tools, infrastructure developers must build the physical systems such as data centers that underpin AI. Businesses need to supply data and drive the organizational transformation to deploy models at their full potential—not just for efficiency but also to expand existing products and services or create new ones. Employers must decide how these technologies will be used, how workflows will be transformed, and how to enable workers to be more productive. Workers have to adopt the technology and invest in learning AI-relevant skills. Local communities must provide the land for data centers and consent to the technology’s expansion. And governments need to set rules, fund foundational research, and provide social insurance to support and manage transitions for those whose work or livelihoods are displaced by AI.
Each of these groups bears unique costs, captures different shares of the benefits, answers to its own constituencies, and responds to distinct incentives. This also means that unless they coordinate—whether through market mechanisms or active cooperation—these groups could slow or block AI development. Firms might underinvest in organizational change and use AI only to cut costs rather than to expand revenues, workers might resist or slow down adoption, communities might block infrastructure, and governments might restrict development. Each action might make sense in isolation, but together they lead to a suboptimal outcome for society.
This is a classic cooperative game scenario in game theory: these groups can accomplish much more when they work in concert than when they act individually and without coordination. A coalition with a common goal does not need to be a formal organization; it can be any grouping of actors who operate in ways that allow them to coordinate decisions and contribute resources to a shared project through contracts, policies, standards, investment commitments, or political bargains. Each possible coalition can generate more growth and economic value by promoting the adoption and diffusion of AI than any one actor can by operating separately.
In an ideal world, all these actors would form a single “grand coalition” to capture as much value from AI as possible. Their complementary actions and active coordination to promote AI would kick off a virtuous cycle in which wider adoption would spur advancements that would make AI more useful for society, unlock new demand and productivity growth, and encourage more people, communities, and companies to adopt the technology and support its expansion.
Such a grand coalition, however, is a benchmark rather than a blueprint. Progress is not all or nothing; smaller, more focused coalitions can still create economic and social gains that far outweigh what each actor could achieve alone. What is needed is a stable coalition—one in which no participant or group of participants can do better by leaving and doing something on their own. In the AI economy, exiting the coalition does not mean building a rival AI system. It simply means exercising a veto on increased AI use, which could take the form of imposing a local moratorium on data centers, resisting workplace adoption of AI, or blocking new regulations that encourage AI development. Because the withdrawal of any one participant would destroy more value than that participant could ever capture alone, it is critical that every essential group prefers being inside the coalition to being outside it.
A crucial question, then, is how to share the gains of working together so that cooperation remains stable over time and no coalition member wants to leave. In the case of AI, one factor that can help drive cooperation is that the technology creates conditions for what economists call a “convex game”: because the inputs to AI are complementary, each new participant makes everyone else’s contribution more valuable, causing the gains from joining the coalition to increase at an accelerating rate. Building better computing infrastructure, for instance, makes AI models and tools more capable; more capable models and tools increase the benefits for companies that want to invest in transforming their organizational processes and retraining workers, which increases productivity and spurs innovation; and wider adoption of AI in the workplace drives the need for additional computing power, which creates incentives for more infrastructure providers to build. Once a core group starts working together, the next actor that wants to join has even more to gain.
These characteristics of the AI economy also make it easier to share the gains of growth more fairly. A simple approach is to distribute the benefits based on how much each member contributes to a coalition. Under the framework of the Nobel Prize–winning economist Lloyd Shapley, each participant’s contribution can be measured by the average additional value that possible coalitions create when that participant is included. If a coalition wants to use AI to discover new treatments for chronic or rare diseases, for instance, it needs to include drug developers, AI labs, research hospitals, frontline physicians, patient groups, and the government. AI labs supply specialized models and computing power, research hospitals curate historical clinical data, physicians administer drug trials, and patient groups consent to share biological records (under privacy guardrails set by the government). These contributions are not measured in the same units, and although smaller coalitions can make partial progress, all these groups are needed for the drug discovery pipeline to fully succeed. A Shapley-style arrangement would share the gains based on how much different coalitions benefit from having each member inside them compared with what would result if the member were outside them. As Shapley’s research showed, in a convex game such as AI adoption, compensating each partner in this way produces a stable coalition that incentivizes everyone to join.
Prices, contracts, and incentives are indispensable for building a stable coalition, and market design and competition provide the architecture that enables coordination. But markets alone will not be enough. In the case of AI, the various costs and benefits will not unfold on the same timelines for all participants; some of them will benefit long before others catch up. Public institutions are necessary. Only policymakers, for instance, can create the common rules that will allow frontier labs in the coalition to coordinate on making AI safer, and only policymakers can provide enough support for workers to adapt to the AI economy. The bitter experiences that many workers and communities felt during globalization have made coalition building harder yet even more necessary.
COALITION OF THE WILLING
A workable coalition does not need to bring every participant to the same negotiating table. A national bargain can establish common terms and align incentives on issues such as AI safety standards, workforce transition support, and infrastructure permitting, and smaller agreements at workplaces and in local communities can bring needed groups on board to cooperate. At both levels, the principle is the same: each group must receive a credible stake in the upside of the economic gains that the coalition creates through the overall growth of AI.
At the national level, cooperation should focus on the groups whose participation will determine whether AI technology can scale enough to generate massive economic gains. This includes the federal government, leading labs, major infrastructure builders and providers, organized labor, and large firms deploying AI. Big business matters because its use of technology and its investments in the workforce create demand that fuels the entire supply chain. Labor organizations will want to negotiate favorable terms for workforce transition, such as expanded unemployment and wage insurance, as well as for retraining; a lack of participation from organized labor will deprive the coalition of the support needed to facilitate adoption, especially in sectors such as construction in which unions have more power. For critical infrastructure, including power grids, telecommunications networks, and payment systems, the priority needs to be cybersecurity and reliability: before they integrate AI deeply into core operations, infrastructure providers need to believe that more capable models will not expose them to catastrophic risk.
The government can establish a predictable, risk-based regulatory framework to set clear standards in potentially dangerous areas such as biosecurity and to help coordinate the testing of advanced AI systems. A national framework would preempt a fragmented patchwork of idiosyncratic state-by-state bans on low-risk commercial uses of AI. Policymakers can also support AI development by funding research on foundational questions in artificial intelligence and increasing immigration and visa pathways for highly skilled workers to attract top global talent to U.S. labs. AI model and application developers, meanwhile, can participate in safety testing, and companies building out AI capacity can make available more computing power to small businesses, nonprofits, and local governments. Infrastructure providers can construct new data centers and show their support for the communities they work in by implementing ratepayer protections and paying local taxes, and large companies can invest in worker training, organizational transformation, and new business models. Finally, organized labor can supply the skilled employees needed to build the physical infrastructure underlying AI adoption, help redesign organizational workflows within companies, and lend its political weight to promoting AI nationally.
Each group benefits more than it is expected to contribute. AI model and application developers and infrastructure providers gain a predictable development path and a skilled workforce. Large firms have access to the support they need to more deeply adopt AI and improve their productivity. Workers gain new opportunities and job security during the economic transition. Washington also stands to benefit as the economy grows and the United States maintains its global technological leadership.
Promises of scientific breakthroughs will not overcome deep-seated hostility.
AI model developers and other companies in the AI ecosystem are already working together to diffuse AI more widely. Leading developers, including Anthropic, Google, and OpenAI, along with partner organizations, have embedded engineers inside large organizations to identify valuable uses of AI, redesign workflows, and integrate new technology systems into production. These efforts recognize that model access is not enough: adoption requires technical assistance, organizational knowledge, and sustained engagement. A coalition can extend this capacity to smaller firms, schools, clinics, and municipalities. For AI model and application developers, supporting diffusion expands the market and provides insurance against potential political backlash. Only if the gains of AI adoption are shared with a broad swath of organizations and the general public will citizens support the expansion of the technology in society.
Frameworks for coalition building with labor groups are also emerging. Rather than working with individual firms to sign bespoke agreements, technology developers, industry associations, and policymakers are supporting worker cooperation through existing unions, apprenticeships, and retraining programs. The International Brotherhood of Electrical Workers, for example, is mobilizing its members to oppose data center moratoriums while calling for better labor agreements. Google is working with the IBEW’s partner organizations to train 100,000 electrical workers and 30,000 apprentices, giving workers access to well-paid jobs and providing developers with a source of skilled labor and the political support needed to build.
At the local level, new data centers may impose concentrated costs on communities even when they create national benefits for AI users. Developers should therefore negotiate agreements that protect ratepayers and provide tangible benefits for communities through increased local revenue, better infrastructure, and more employment opportunities. Templates for how to do this are taking shape in many U.S. states. Leading AI companies and “hyperscalers,” including Amazon, Google, and Microsoft, have signed on to the Ratepayer Protection Pledge, a voluntary commitment to pay the full cost of new power generation and grid upgrades in areas where they are building data centers so that household electricity bills do not rise. These companies are also funding local workforce programs and water-restoration projects and contributing property tax revenue that will allow communities to pay for new schools and roads.
The economic benefits that AI brings will increase the stability of emerging coalitions as long as they are shared. The distribution of benefits does not need to be equal, but for groups providing complementary inputs, benefits must be apportioned according to what each group contributes and also be adjusted as the technology evolves. In practice, determining each group’s contribution is not an exact science but an ongoing structured negotiation: tools such as utility tariffs, labor agreements, and enabling policies can allocate benefits in ways that reflect the role of different groups in supporting AI expansion and adoption. The share of the upside that each coalition member receives must be large enough to make participation more attractive than obstruction. Giving key players a real stake in the economic surplus that is generated is how those who might have been potential impediments to AI become participants in, and perhaps even advocates for, the expanded use of this new technology.
Where AI substitutes for existing tasks rather than complementing them, policymakers can also learn from what went wrong with globalization. Public programs to help displaced workers can be effective, but they must be accessible enough to provide support when technological disruption occurs. Targeted parts of the federal government’s Trade Adjustment Assistance program to compensate and retrain workers affected by globalization succeeded where they were put in place—research by the Federal Reserve Bank of New York found that eligibility for wage insurance under the program increased reemployment, shortened the amount of time out of work, and raised cumulative earnings—but it reached only a small portion of workers. Other research found that most support was provided through disability benefits that required workers to permanently exit the labor force. With AI, policymakers will need to set up expanded wage and unemployment insurance programs and support employer-led retraining programs that help workers transition into in-demand roles before displacement occurs rather than improvise after workers begin to suffer.
POSITIVE REINFORCEMENT
A coalition does not need to be built all at once. Cooperation should start in areas in which complementary benefits are clearest and participation constraints are easiest to identify, including promoting AI adoption and worker training in particular sectors such as health care or skilled trades, and encouraging support for new data centers. At the national level, safety cooperation can grow out of partnerships between coalitions and the government. Because AI’s development trajectory remains uncertain, cooperation agreements should center on investments that could pay out across multiple plausible scenarios and can be revised as the trajectory of AI becomes clearer.
International cooperation can grow in the same way. Although AI may well eventually require cooperation among adversaries and allies alike, Washington can start by forming a U.S.-led coalition built around trusted partners willing to accept common safety and security standards. The United States could, for example, set up a coalition among trusted partners that agrees on principles and frameworks in exchange for access to technology and cooperation. This coalition will naturally widen as its terms prove workable and more countries see the benefits of participating. It could eventually include China, which, given its advanced AI capabilities, is critical to addressing AI safety issues; this will require navigating complex questions of alignment and verification, however, including how to set up trusted validation bodies and mechanisms.
There is reason to be cautiously optimistic about cooperation in the age of AI.
The Basel Committee on Banking Supervision provides an instructive example of international cooperation. It began in 1974 as a small forum among national regulators in 12 advanced industrialized countries and over time developed common standards for how to supervise banks around the world without empowering an imposing supranational authority, expanding its scope as those standards proved valuable. The Basel Committee was the seed for what is now the dominant international architecture that identifies and minimizes the risks of financial instability, including setting standards for how much capital banks need to hold and how to classify the riskiness of securities held on banks’ balance sheets. A future coalition for AI standards could grow through the same combination of clear entry conditions and credible commitments that have made the Basel Committee and the International Organization for Standardization, which publishes voluntary technical standards, successful.
Harnessing the potential of AI will not be easy. Critics will immediately jump to the conclusion that there are many reasons why such coalitions will not work in practice. Labs might decide that they can earn more on their own, incumbent firms might try to convert cooperation into tools to protect their market share, and governments might promise support for workers and communities and fail to deliver, as was the case with globalization.
But there is reason to be cautiously optimistic about cooperation in the age of AI. Because the technology is still relatively early in its development, it is possible to identify the groups needed to initiate cooperation, and labs, governments, communities, employers, and labor organizations have already shown interest in different parts of the bargain. Ultimately, society does not need to choose between innovation and shared prosperity. The two can be complements: cooperation will drive growth, and sharing the gains of such growth will sustain the cooperation needed to keep it going.
Loading…
