Getting to all-of-the-above: A framework of solutions for AI’s coming impacts on work and workers

Introduction: From no policy ideas to many

On June 2, The Washington Post ran a story titled “Five ideas for how we survive the possible AI jobs apocalypse.” The headline underscored the public’s anxiety about AI’s implications for jobs and the apparent lack of a plan, at any level of government, for addressing that.

This was not the typical story about how AI could destroy jobs or, conversely, how hard it still is to discern exactly what the impacts of AI on work and workers could be. That challenge, along with the nation’s remarkably buoyant economy and low overall unemployment rate, helps explain why shaping AI’s economic impacts was barely on the policy agenda, especially at the federal level, until this year.

The Post article signals an important inflection in the national conversation about AI. It is among the first of its kind by a major news source to highlight the wide range of economic “solutions” that have garnered some attention and prominent supporters. While we still lack a plan, we suddenly have a kind of “ideas soup” that reflects many influences: different political agendas; hope versus pessimism about government’s ability to govern AI (and as part of that, help shape a positive future of work); wide-ranging expectations about what disruptions will be most harmful as AI transforms the economy; and, as the public worries most, the business interests of a tiny handful of Big Tech companies that currently dominate the development of AI and its infrastructure, including Anthropic, OpenAI, Google, Amazon, Apple, Meta, Microsoft, SpaceX, and chipmaker Nvidia.

The ideas highlighted in the Post article—some of which are championed by sitting elected officials or candidates for office in this year’s elections, or by influential scholars and business innovators—range from modernizing unemployment insurance and making it more generous (e.g., to replace the comparatively higher incomes earned by white collar workers whose jobs are increasingly vulnerable to AI’s advanced capabilities) to proposals to make sure the tax code does not inadvertently serve as an accelerant for automation. It also includes proposals to directly tax companies’ AI use (the tokens that indicate usage), and creating public wealth funds or broad-based stock ownership (i.e., direct “gain sharing” in AI-produced wealth by the public). This month, Anthropic became the first leading AI developer to publish an economic policy framework, organized around economic impact scenarios and multiple policy tools such as those above.

All of these proposals—and the need to dramatically dial up our readiness for AI’s impacts on the economy and society—are now becoming more visible, thanks in part to election-year stakes, but also to how rapidly the technology has continued to advance since the release of ChatGPT in 2022 launched the era of generative AI.

The public is increasingly anxious about what AI means for jobs. According to public opinion research by Pew Research Center, Gallup, and other sources, this concern has often outranked other public concerns about AI, even those about product safety and national security risks. Averaging the results, about three-quarters of surveyed adults fear job losses due to AI. In addition to age, gender, race, and other differences in views on AI (especially concerns about bias in hiring), there is some partisan sorting. Democrats are somewhat more likely than Republicans to use AI often and work in occupations that are more exposed to potential disruption, according to recent Brookings research. What’s more, sentiment about the economic stakes is much more pessimistic than public views of AI in areas such as scientific research, health care, and education.

The news headlines, from college graduates struggling to find jobs to the eye-popping spending in Big Tech, help explain that pessimism. In March, for example, Meta, the parent company of Facebook and other popular apps, announced the layoff of 700 staffers while creating a new compensation program that could offer just six of its top executives nearly $1 billion per year in additional compensation over the next five years. In May, the company laid off another 8,000 employees. The stated reason? A pivot to artificial intelligence. Meta’s announcements follow a series of others in the tech sector this year, variously described as signs of what’s to come and disingenuous “AI washing” to cover up over-hiring and signal ruthless resolve to Wall Street. Meanwhile, two of the largest AI developers, Anthropic and OpenAI, are preparing for what could be among the largest stock offerings in history, joining SpaceX in the tiny club of companies valued around $1 trillion dollars or more.

The questions the public has about AI and work specifically are at once broader and closer to home than wealth creation or executive compensation in Big Tech: What about my job? What about my kid’s economic future in the face of AI disruption and potential automation? Will my community be left behind by another sweeping force of economic change? Public opinion on AI overall has also turned starkly negative over the past year, as the news media has documented. And according to Stanford University’s annual AI Index Report. this negative sentiment is far stronger in the U.S. than in other countries.

But when it comes to the election-year effect, campaign platforms can often start and stop at broadly sketched proposals. Governing inevitably requires much more explaining, persuading, making detailed proposals, and weighing tradeoffs, as the Post story hints in a brief rundown of the perceived pros and cons of its five big ideas. The politics of advancing any of these ideas—and at scale—is far from straightforward.

Yet there’s a larger issue when it comes to responding wisely and fairly to the unknowns about what’s coming in the economy. Look more closely at the five ideas the Post article describes—a reasonable excerpt of a wide-ranging but hard-to-focus debate—and something else is apparent: The ideas address very different problems or “solves,” and on wildly different time scales. They range from well-established if often criticized programs, such as unemployment insurance, to radically new ways of treating wealth in the digital economy. In fact, some proposals would redistribute wealth and not address access to good work, at least not directly. Other ideas seek to discourage job automation, while others merely propose to replace income for a time—not the sense of purpose, dignity, and social connection that also makes work central to our lives and the fabric of society.

As this report will explore, some promising ideas—especially for keeping people connected to good jobs over the next decade or two, as economic changes play out—are barely visible in the public conversation, yet may represent some of the most important steps we can take now. Tackling—at last—our national shortage of skilled nurses and workers in the skilled trades is one example. Another is developing economic scenarios and highlighting the policy choices they raise, as New York City’s comptroller did in a landmark recent report on AI and the city’s fiscal future.

How should public and private decisionmakers and the voting public think about what’s ahead, and what should we make of the ideas above or others? It’s time to organize and expand the fragmented pool of would-be solutions, not in hopes of an easy consensus, but because it’s hard to make progress when we are largely talking past each other about what we will need to solve for, let alone how, at what costs, born by whom, with what potential and limits, and so on.

This report offers a four-part framework of response strategies and illustrates what each looks like or could look like in the future (see Table 1). The four types of strategies are:

• Brakes on labor automation, even as AI adoption expands (i.e., not discouraging AI adoption broadly).
• Steers that seek to influence the choices of workers and employers toward less vulnerability and measures that improve the quality and availability of work, for noncollege workers as well as those with higher education degrees.
• Buffers that seek to reduce harms, whether from income loss or the profound sense of dislocation and resentment that can accompany prolonged joblessness, especially when it is concentrated in a community.
• Shifts, or deeper transformations in the role of paid work, length of the standard work week, or how economic security, gain sharing as new technology creates new wealth, and the value of human labor are recognized and rewarded.

The main message of this report is that we quickly need to shift thinking and action to a holistic, all-of-the-above mindset—beyond narrow ideas or false binaries, such as the one that asks us to choose either innovation or broadly inclusive opportunity and human dignity.

The framework offered in this report prioritizes buying time to adapt; equipping workers and future workers so they can be more adaptive; deciding what a future-facing safety net must be capable of; incentivizing both workers and employers in promising directions that reduce vulnerability, if only for a time; improving the measurement of AI’s effects on the economy and workers; and advancing broader conversations about the role of work, the ownership of AI-generated wealth, and the variety of ways we can generate economic security, enjoyment, and sense of purpose in changing economies.

The next section looks at the perils of solving problems that come with many uncertainties and a wide range of stakes and stakeholders, especially those that center on economic opportunity, hardship, and who pays the so-called “price of progress.” The report then outlines what we can do and why, and offers a call to action to leaders in the public, private, and civic sectors. The report concludes with a look at encouraging steps that state policymakers are taking in the absence of national policy, along with ideas for building on those initial steps through better planning and learning-while-doing.

Recognizing solution traps: Silver bullets, bias, and rationalizations

The U.S. and other nations have a well-documented tendency to look for silver bullets for complex problems (“let’s train everyone for the jobs of the future”)—and, worse yet, to sometimes shift the goal posts and decide that something unfortunate isn’t an urgent problem that needs solving after all. The first trap only requires narrowing our gaze and skipping over complexity, so everything looks like a familiar problem, whether to regulate, educate for, or compensate for. The second trap calls for rationalizing things that are at odds with our espoused values. Humans, as history instructs, are highly practiced at both.

Given our nation’s particular history of judging deservingness and allocating harm, our tolerance for high levels of inequality and hardship—often explained as the necessary price of progress, for the greater good—is hardly inspiring as we enter this next industrial revolution. From waging an immensely costly and mostly ineffective “war on drugs” to the endless variety of quick fixes prescribed—with special confidence on the campaign trail—for delivering on the American Dream, our history is littered with silver-bullet fantasies and false promises.

Less visibly, that history is also marked by powerful rationalizations that let serious problems, such as poverty in the midst of plenty, persist on a low boil. A classic insight about the history of public policymaking, and especially the need to shape major social transformations in positive ways, is a humbling one: that a condition, however unfortunate, is not actually a “problem” unless a critical mass of us (and those we put in office) believe it must be changed. Political scientist John Kingdon came to that conclusion four decades ago, after analyzing what does and doesn’t make it onto the national policy agenda and why. But arguably, that’s at least as important today in our more polarized age.

I along with Brookings colleagues Molly Kinder and Mark Muro highlighted this in a guest essay for Time magazine on the eve of the 2024 election. AI’s looming implications for the American worker were not on the policy agenda at all, and barely in the public conversation, despite the fact that both presidential candidates presented themselves as champions of workers. That absence, in itself, was a problem.

Thankfully, that has changed, with significantly more attention to what policymakers should start doing about an AI-affected economic future. But especially for that reason, understanding solution traps and rethinking what kinds of responses should count as solutions is urgent now, as the U.S. and other nations confront one of the most complex and rapidly advancing technologies ever invented—and consider what it means for the future of human work, purpose, and well-being. It’s not only critical that we recognize half-measures and empty gestures as such, or reject the simplistic idea, inconsistent with the real social and economic history of technological progress, that imposing any rules on AI development will imperil American innovation. It’s also critical that we think across very different kinds of stakes and time horizons.

A substantively and politically flexible framework for responding to AI’s economic promise and risks now arguably must be able to “supply” several things. For example, we need mechanisms that can function in steady-state fashion, such as a more flexible, always-accessible education and training system that adapts over time, along with fit-for-purpose measures to handle contingencies such as income replacement, which might kick in several times over the course of a worker’s career changes, or concentrate “adjustment assistance” on workers in hard-hit industries and regions. In theory, our economy has both steady-state and contingent policies and delivery systems in place now, side by side. But neither type functions adequately or serves everyone it should. Our systems could learn much from those in other economies. We are only beginning to explore what AI readiness should look like for these systems—in any country.

In terms of what will make particular policy choices good, we need mechanisms that not only promise measurable effectiveness, based on some agreed-upon outcomes we want to achieve, but also political durability, with the traits of robust political bargains that can outlast terms of office and overcome partisan and other differences. Social Security and Medicare came to embody such bargains, but encouragingly, so does the much newer cause of dramatically expanding funded apprenticeships nationwide, which is a win-win for employers, workers, and government. That cause is flourishing, in Democratic- and Republican-led states and swing states alike, and is championed by prominent voices in both major parties. The same can be said for the idea of expanding America’s “ownership economy,” in part by expanding who can be an owner and not just a wage earner.

Bolder ideas for responding to and shaping the economic impacts of AI have tailwinds—in our politics and culture—that our current debate over AI and jobs does not yet acknowledge.

The AI and jobs debate: A brief history

The aforementioned Washington Post story describing five big ideas for navigating an AI-affected economic future is a window on a much larger and messier public conversation. That conversation has evolved in fits and starts for a decade now—well before ChatGPT’s public release in November 2022 put us in a new era seemingly overnight.

The conversation still rests on a challenging foundation: There is no consensus on AI’s job effects so far (in part because disentangling them from other forces buffeting the economy is hard), let alone about the likeliest effects to come, from automation to “augmentation” of human labor. Privately, CEOs of large companies say they expect significant AI-related layoffs in the next few years. In surveys this year, 55% of business leaders and 68% of investors say they also expect less entry-level hiring—consistent with the public’s broad expectations. But to date, the uncertainty has left room for wide-ranging speculation and corresponding complaints that it is too early to take much action, or presumably discuss scenarios, such as what we could do if “X” came to pass and what it would take to be ready to do so.

Furthermore, for all its apparent superpowers, AI sometimes makes things up and is otherwise buggy. It also generates new and only partially understood risks. As was true in earlier waves of technological change, a routine and vital role for human workers (for now) is catching and troubleshooting errors the emergent technology creates, including the latest and most aggressively promoted.

In lieu of consensus—or any clear, linear arc of change—there are some clear schools of thought about AI and the future of work. Still vocal are the techno-optimists (and not just in Big Tech), reassuring us that economic fallout will largely resolve itself, or the fallout can surely be fixed by some kind of universal guaranteed income. Meanwhile, the monumental buildout of data centers—the physical backbone of an AI-supercharged society—means lots of engineering and construction jobs, for which we have an acute shortage of workers, not a surplus, for the foreseeable future.

Then there are the doomsayers, or, better said, there is doomsaying about large and sweeping risks to jobs as we know them and, conservatively estimated, tens of millions of incumbent workers, not to mention young people just starting out. Icons of American tech, who are bullish about unleashing the potential of new technologies, have issued stark warnings—of an anxious public, desperate for answers as AI accelerates—for much longer than today’s media narrative tends to acknowledge.

Consider the frenzy since ChatGPT’s first release. Then consider that it has been nearly a decade since Microsoft co-founder Bill Gates floated the idea of a “robot tax.” Its purpose, he suggested, would not be to discourage the productive application of machine intelligence, but to slow the pace of automation—buy society some time—while offering a major new revenue source to help workers retrain and otherwise adapt to broad economic disruption.

More recently, from the frontlines of the “frontier” AI model development implicated in that projected disruption, Anthropic CEO Dario Amodei has consistently sounded alarm bells. Earlier this year, in a remarkable statement by a business leader, let alone one in such a powerful growth industry, Amodei’s essay, “The Adolescence of Technology,” warned of a large and essentially unavoidable disruption of “most existing jobs.” Amodei assessed that threat, which he considers imminent in the next few years, alongside other AI-powered threats from cybercriminals, repressive governments, sociopaths, and other bad actors. (Sadly, for the rest of us, these bad actors tend to be early adopters of new super tools.)

Amodei’s rogue’s gallery and threat inventory is an expertly explained and varied one, which is reason alone to give his essay a close reading and wide discussion, regardless of one’s initial assumptions. But his remedies for large-scale job disruption and other fallout from rapidly evolving AI—such as even greater concentration of wealth and power—are, for now, only a sketch. For example, he posits that we may be able to “steer” companies to emphasize innovation—which he defines, refreshingly, as doing more with roughly the same number of human workers—more than cost cutting (doing the same things with fewer workers). And he acknowledges that we are likely to see lots of both.

For those of us who are not AI model builders or major AI investors, the most important takeaway for now may be a simple one: Leading developers of one of the most complex and widely deployable general purpose technologies ever imagined are telling us to do much more than we are currently doing to make ready—at least—for the coming economic disruption.

As Kinder, Muro, and I emphasized in a 2024 Brookings report on AI and the future of work, which drew on OpenAI data and analyzed over 1,000 occupations in the U.S. labor market, most workers lack union representation or other meaningful sources of collective voice to help shape their economic future. In the private sector, just under 6% of workers are unionized, according to federal data. What’s more, we found that the exposure of one’s job to ever-more-capable AI is inversely correlated with the odds of belonging to a union and having the collective bargaining power that confers. The most exposed are the least collectively organized (specifically, unionized). We termed this “the great mismatch.”

Meanwhile, researchers at Brookings, Opportunity@Work, and other organizations have documented the serious risks AI deployment poses to the career ladder and career transitions, including risks to “gateway” jobs that have traditionally led to higher-wage occupations that do not require a four-year degree. Analysts have also underscored the importance of thinking about the very uneven AI readiness of our country’s regional economies (i.e., local labor markets), not just readiness by categories of workers (including adaptive capacity in the face of job disruption) and industry sector. The profound economic, social, and political impacts of deindustrialization, after all, were highly concentrated in the regions most reliant on heavy industry—and the highly unionized, middle-class jobs manufacturing offered to generations of families and their communities.

Finally, a flurry of analyses—for example, a tracker launched by The Budget Lab at Yale University—have warned that we lack reliable indicators of AI-driven job losses. So many factors, from tariffs and oil prices to consumer confidence and pandemic-era over-hiring, are influencing employers’ current decisions about layoffs and hiring. A larger question is whether certain kinds of jobs are being offered as often as they used to be (“missing jobs”) and whether AI is affecting less visible markers of job opportunity, such as the volume of entry-level hiring, especially for young workers entering the labor market.

A different kind of analysis and policy imagination, in the vein of solutions research, has also expanded rapidly over the past two years, as highlighted in the next section.

A framework to guide discussion and action: What we need to solve for, why, and how

As news headlines about AI-related job threats constantly remind us, prediction has a utility all its own, especially for grabbing public attention. But it has well-documented limits too. When uncertainty and complexity are both significant, building scenarios and asking what-if questions can help decisionmakers consider everything from early “no-regret bets” to medium- and longer-term contingency plans.

In that vein, between extremes of doomerism and optimism, several core premises about AI, work, and economic security are gaining credibility:

1. AI’s potential for large-scale job disruption is real and broad, given the demonstrated range and sophistication of its capabilities and the fact that they continue to advance. For now, that disruption has been mostly held in check by the organizational “friction” seen in earlier waves of technology adoption—not by intentional pro-worker design features of the AI tools, the direct costs of acquiring the tech, formal labor regulation, or worker bargaining power. The public sector is an important outlier, given how differently talent decisions work and what constrains them, but not a total exception. As Anthropic emphasizes in its recent report about the latest step change in AI capabilities, societies need viable ways to buy time.
2. Employers in every sector have a broad, not narrow, spectrum of choices about how to adopt and use AI in workflows and about the role and voice of workers in shaping those changes. This “effects-are-not-predetermined” premise reflects the discretion available to managers, owners, and shareholders (in the case of business and traded companies, specifically); the range of the technology’s capabilities, with a capacity to surprise even its creators; and a well-documented, global history of worker-informed, iterative technology adoption, especially in manufacturing but in services too—a practice that Thomas Kochan of the Massachusetts Institute of Technology (MIT) Sloan School of Management termed “giving wisdom to the machine.” Likewise, in a report on building pro-worker AI for The Hamilton Project at Brookings, MIT economists Daron Acemoglu, David Autor, and Simon Johnson made the case for a range of rules and incentives, from procurement and taxation to mobilizing public and private investment, to directly encourage (steer) employers toward pro-worker “innovation,” as Amodei defined it. So in addition to brakes, societies need to find ways to constructively steer the choices that both workers and employers can make.
3. Job disruptions and related effects are likely to unfold unevenly across geographies, occupations, and industry sectors, and across shorter and longer time horizons. Many so-called “general purpose” technologies, such as electric power, took decades to diffuse through the economy. AI is already moving much faster than that but affecting AI adopters and uses in very different ways, in part because of what Wharton School professor Ethan Mollick and collaborators termed the “jagged edge” or “jagged frontier” of the technology’s capabilities. This has many potential implications; for example, as previewed above, that economic regions, not just workers and organizations, could benefit from “AI readiness” planning and other support. The already-significant inequality between U.S. regions has been sharply amplified in recent decades by technological change. In addition, when it comes to economic security adjacent to work, the U.S. may need means of supporting worker transitions that are much more effective than similarly motivated programs launched decades ago as part of trade adjustment assistance to aid “displaced workers.” To address displacement, such buffers need to be more robust, flexible, and effectively implemented.
4. Livelihood risks—together with the potential for AI-powered benefits—are significant and sweeping enough to warrant a rethink of the role and footprint of human labor, wealth or “gain” sharing, and the social contract. This includes, but need not be limited to, everything from the length of the work week and the nature of employment relationships—for now, a long-established and rather basic binary in federal labor law, which separates salaried from contracted workers—to safety net policies broader than displacement assistance, innovative forms of public sector or publicly guaranteed employment, guaranteed income, public wealth funds, and more. Four-day work week demonstration studies, for example, have been gaining steam and attention from business in several advanced economies for over a decade now. But the lessons are almost never connected to discussions of our economic future with AI. These kinds of deeper shifts—which go beyond brakes, steers, and buffers—are, at best, at the margins of the current public conversation in the U.S. But in the current political climate, such proposals could move quickly toward the mainstream of discussable policy ideas.

While government action alone is likely inadequate to meet any of these work and economic security challenges tied to the diffusion of AI, government has multiple, indispensable roles to play—making greater and more adaptive “state capacity” (the conditions for government effectiveness) essential. Government is implicated via myriad roles, including: regulator and standard setter; research and development investor; large-scale buyer and user of tech and related services; keeper of a social contract and funder of a safety net to help workers, employers, and communities manage financial risks, especially income loss; insurer, lender, and financial guarantor for lenders, operating companies large and small, college students and trainees, and others; critical source of data on industries, employment, and more; and attention-focusing “bully pulpit.”

In spite of the persistent unknowns about AI’s full range of uses and impacts, it’s becoming clear that we’ll need solutions that complement each other and have distinct objectives and potential. Grounded in the premises outlined above, what follows is a framework of such solutions, organized into four main types: brakes, steers, buffers, and shifts.

Brakes on automation mainly buy us time to adjust to disruption without undermining technology development itself and without necessarily discouraging AI adoption. That distinction is important now, especially for getting past the false choice between innovation and broadly shared gains.

Steers for workers, on the other hand, could nudge large numbers of workers, including but not limited to young people starting their careers, toward less vulnerable occupations. This includes well-paid ones, such as skilled nursing, which are nonetheless already being transformed by AI. Steers for employers would encourage AI uses that improve work for workers and help them create more value for employers, customers, and society.

Buffers will seek to reduce harms and hardship, in ways that unemployment insurance and other tools were designed to do for the 20th century industrial economy—but potentially multiple times over the course of a given worker’s career and retirement.

Meanwhile, more deeply transformative shifts could restructure our relationship to work and economic value—and even to each other, for example by creating public goods through AI-fueled profits and intellectual property, which we can access regardless of occupation or specific experiences of job displacement. Reimagined public wealth, a shorter work week, and a different kind of safety net could fall into this category.

The framework underscores why popular, even intuitive responses, such as encouraging broad-based AI literacy building in education and training, are at once very important and quite limited. They proceed from familiar models and fields, such as workforce training and income replacement, or—less visible and broadly supported—how trade unions think about and bargain over the use of new technologies in the workplace. But as an example of the shortcomings of familiar tools, training or retraining alone does not produce good jobs, at least not directly and readily. Deindustrialization highlighted that lesson, but so did the more recent rise of heavily indebted trainees without a job. To date, the most familiar tools have also generated little dialogue, let alone coalition building to drive action on public policy. Nor are they fueling realism about what particular solutions or levers can be reasonably expected to achieve in the face of something so transformative and complex.

But the framework also highlights other lessons, including the importance and current limits of public sector capacity (aka “state capacity”). For example, discussions in the U.S. about enhancing AI workforce readiness, which this report categorizes conservatively as a buffer, begin on a rickety and rather rigid foundation: a sprawl of education and training opportunities that manages to not only underperform but show great inertia, resisting transformative change.

Conversely, as a brake on labor automation (and as recently enacted state laws on AI use underscore below), existing policy models and practices for regulating algorithmic decisionmaking by employers—for example, through required disclosures or limits on automated scheduling decisions for workers—offer evidence, legal precedents, and other useful foundations for tackling related challenges AI poses. In other words, some policies, especially those of relatively recent vintage for regulating employer use of algorithmic or other machine decisionmaking that affects workers, offer readily applicable building blocks. Meanwhile, more established policy tools and practices, such as workforce development or 20th-century-style labor regulation, may need bolder rethinking and legislative reform.

Take education and training, which are commonly thought to be central to how we must respond to an AI-affected economy. Compared to other advanced economies, the U.S. spends little on training as a share of GDP (especially through our public workforce system) and gets a poor return on that investment. The system is not demonstrably inclusive, not responsive to changes in employer demand and skill requirements, and rarely outcome-driven. It is also highly skewed, in terms of public and private subsidy, toward making expensive higher education more affordable, mainly for teenagers and young adults with limited work experience. Applied skills training, such as through affordable apprenticeships for young students or older incumbent workers, receives a small fraction of the subsidy—and far less, in per capita terms, than other economies invest, as analyses have consistently shown. These shortcomings reflect inconsistent policy priorities and delivery failures, both of which implicate state capacity. Our systems for learning tend not to learn effectively or continuously. Choices to make major education and training reforms, for example, and achievement of at least very good implementation of those reforms, often lag our knowledge base on what works by years, sometimes decades.

It’s time to stop lamenting that so much of our workforce investing system operates on “train-and-pray” wishfulness—the phrase former Labor Secretary Tom Perez helped popularize. In a recent New York Times op-ed, former Rhode Island Governor and Commerce Secretary Gina Raimondo focused on this aspect of the AI challenge, and what stepping up boldly must include if the U.S. economy is to generate broad-based opportunity and the dignity that work confers in the decades ahead. It is an argument joining the competitiveness of firms and the nation to worker success and a much more outcome-driven role for government and public spending.

The compelling potential and the heavy lifts in that all center on a “buffer” response that already receives far more attention and support (including bipartisan political energy, as skilling is the centerpiece of the Trump administration’s policy agenda for an AI-affected job market) than contentious responses that could be much more transformative over time, whether initially as brakes, steers, or deeper shifts. Examples include:

• Expanding union membership and securing collective bargaining agreements that meaningfully govern AI automation, in addition to applying “worker-first” principles for AI use beyond unionized workplaces. As leading researchers of business innovation have underscored, the broader premise is that skilling and rapid AI awareness building should be complemented by steps to give workers more voice and leverage in shaping the use of AI in work.
• Building on human-in-command rules, for now state by state and mainly as safety-driven technology policy that doubles as a brake on job automation, to advance “pro-worker AI” steers more broadly and ambitiously.
• Using the immense buying power of government to encourage business models and practices enabled by pro-worker AI, both protecting access to livelihoods over time and enhancing the experience and value of work. Such policies would complement growing efforts by nonprofit, business-performance groups, such as JUST Capital, CECP, Council for Inclusive Capitalism, and the Good Jobs Institute, that are tackling AI use and its impacts on both profitability and workers.
• Mobilizing and resourcing the coordinated public and private commitments required to recruit and equip millions of workers into well-paid though not “future-proof” careers—such as skilled nursing and the building trades—that face chronic labor shortages and are less vulnerable to AI automation. We should not underestimate what it will take to meet these big talent pipeline challenges. For one, young people have long been taught to stigmatize work in the trades and associate good careers with white collar professional jobs—some of the very jobs that AI is poised to disrupt most—and to assume, inaccurately, that college degrees are a minimum requirement for those jobs. But based on recent surveys of Gen Z, growing anxiety about AI and frustrating job searches may be shifting those views.
• Rewiring our safety net and skilling system, perhaps inspired by Denmark’s widely admired “Golden Triangle” flexicurity model, which combines labor market flexibility (relatively easy hiring and firing), generous unemployment assistance, and active labor market policies. In this vein, Anthropic has just proposed some form of wage insurance.
• As previewed earlier, rethinking the length and demands of the work week, drawing on encouraging experiments in a variety of national settings, to enhance productivity as well as worker satisfaction and well-being.
• As OpenAI described in a white paper that did not endorse specific proposals, creating some form of wealth fund to offer the mass public an equity stake in AI companies. Anthropic’s new framework makes the case for “universal pre-distributive capital accounts” for every American. Elon Musk has floated similar proposals, essentially for large-scale, government-enabled income or wealth transfers.
• Closer to long-standing debates about tax fairness and growth impacts, reforming the tax code to put the value of labor for business on a more level playing field with that of invested capital, as economists Acemoglu, Autor, and Johnson have advocated as an additional lever for “building pro-worker AI.” As Sarita Gupta of the Ford Foundation put it in a recent Time magazine essay, “Our tax code shouldn’t favor buying software over hiring people.”

Having a broad repertoire—designed to include complementary elements—can help policymakers, their constituents, advocates, and government’s delivery partners understand where the bigger versus easier “lifts” may lie and what part each player would need to play. The public debate lacks any such shared vocabulary, let alone tested repertoire, for now. But the OpenAI white paper and Anthropic economic policy framework begin to signal a shift—an inflection point. Big Tech has become much more eager to be associated with solutions.

Encouraging signs in state policymaking

With political Washington mostly at an impasse over what to do about AI’s implications for jobs, our laboratories of democracy are at work: At the state level, policymakers are taking action well beyond AI-related skilling.

As Mishal Khan and Annette Bernhardt, researchers at the University of California, Berkeley’s Labor Center, documented in a 2025 landscape analysis of work and tech policy, state governments have enacted a limited number of what this report has termed “brakes,” primarily in the labor-intensive sectors of health care, education, and creative production. To re-emphasize, “brakes” encompasses strategies that aim primarily to slow and scrutinize automation, not AI adoption generally. Some of these guardrail policies are clearly motivated by consumer safety, not just concern for workers.

For example, in 2025, Illinois enacted laws prohibiting the use of AI in place of community college faculty and mental health professionals, and replica bills have been introduced in Florida, New York, and Pennsylvania. Concerned about transparency for consumers, California has outlawed “advertising products using terms reserved for licensed health care professionals,” while Oregon did so for nurses. California has also defined “community college faculty,” in law, as referring only to humans. Similar, pending bills in New York, Maine, Texas, Connecticut, and other states would prohibit AI use in place of a range of human professionals, from media workers and teachers to commercial truck drivers.

More striking is the fact that a politically diverse range of states has enacted some form of the landmark “human-in-the-loop” law that California passed in 2024. These laws require human judgment where algorithms are used to shape sensitive decisions, such as determining a health insurance benefit or law enforcement finding. Proposed legislative protections tied to public rights and public benefit determinations by public agencies have expanded rapidly as well. In principle, such protective brakes might evolve into more ambitious steers; i.e., to design and implement broadly pro-worker, AI-powered workflows. Some of the building blocks for doing so are available but not yet widely known; for example, the EPOCH framework created by Isabella Loaiza and Roberto Rigobon of the MIT Sloan School of Management to more rigorously specify human capabilities that complement AI’s shortcomings.

In other examples of brakes, a range of state bills would require employers to do automation impact assessments prior to deploying new digital technologies and changing core job functions or prohibit employers from requiring workers to train AI on certain worker-produced work products, such as likeness, voice, art, and music. Such policies buttress the few collective bargaining agreements, most famously that won by Hollywood screenwriters after a lengthy strike in 2023, that protect creative human output in the face of expected growth in AI use by media and entertainment companies.

As nascent “steers” to influence work design and employment choices by employers, workers, or both, New York lawmakers have introduced several bills that would either tax or withhold subsidies from companies that opt to replace human workers.

In a range of states (and at the federal level in the bipartisan AI-Related Job Impacts Clarity Act), other bills focus on required disclosure of AI-related layoffs. These policy proposals typically seek to modernize laws already on the books, but they underscore something fundamentally important: It is hard to shape something, such as AI-related job effects, that we cannot track in credible ways. And all of our available data sources—traditional employer and worker surveys, closely watched layoff data and corporate statements about them, and the AI usage data made public by OpenAI, Anthropic, and other AI product developers—have major limitations. Stanford University’s recently announced AI Economic Indicators project is one notable effort to do better, using a set of “economic dashboards.” But for now, all such trackers rely on indirect indicators such as layoffs in AI-exposed occupations—not direct evidence that AI was a primary cause.

New Jersey, meanwhile, is showing itself a leader in “right-to-retraining” (a form of “buffer” strategy), at least by the measure of bills introduced. A federal version, though not currently advancing in Congress, was introduced before the pandemic.

Beyond issue-by-issue policymaking, more holistic planning—and making space to consider alternate futures with generative scenario thinking—is sorely needed. In this vein, New York Governor Kathy Hochul recently announced the FutureWorks Commission, made up of business and labor leaders and policy experts, and is charged with “developing recommendations on ways New York can protect the economic security of workers while harnessing the economic benefits of AI.” Other states should launch similar efforts. With 36 gubernatorial races this year, many will be able to do so after the elections in November.

Philanthropy and the broader nonprofit sector are preparing to play a larger role in shaping the economic future. For example, The Windfall Trust, a global nonprofit, has begun to help decisionmakers in government employ scenario planning to think about how best to respond and just how varied and bold our responses may need to be in the face of AI opportunities and disruptions.

Local leaders are also beginning to focus more on shaping the future with AI. Bloomberg Philanthropies and the Bloomberg Center for Government Excellence at Johns Hopkins University recently announced a first-of-its-kind international Mayors AI Forum, to debate and demonstrate an array of responsible AI uses with broad public benefits, including economic ones. Likewise, a group of philanthropic donors has launched Humanity AI, a coalition emphasizing that “our future with AI can and will be what we make it,” with priorities that range from the quality and experience of work to the effectiveness of education and health care, progress of scientific research, and more.

These encouraging efforts can benefit from the range-expanding, clarifying, and more holistic framework of action outlined in this report.

Leveling up on the job and related impacts of AI

The U.S. has lacked a proactive strategy for shaping AI’s impact on work and workers and making ourselves more ready for disruptions that lie ahead—or even a framework for such a strategy. As recently as the 2024 presidential election, candidates for higher office had little to say about it. AI’s likely economic impacts—as distinct from product safety, privacy, and other concerns—were essentially nowhere on the national policy agenda.

But over the past year and a half, encouragingly, policymakers at the state level have shown a willingness to debate new approaches and also to legislate—imposing AI-focused guardrails and new requirements for employers. In many cases, safety and transparency to the public have been the main policy goals, but sometimes basic digital rights and the livelihoods of workers have also been explicit goals. These enacted and proposed laws hint at the range of building blocks already available and worth expanding and learning from.

The rapid evolution of machine intelligence poses more than one kind of economic challenge, and relevant predictions—for example, specifically about career ladders and new occupations, let alone new industries—come with huge uncertainties. State and local policymakers are acting where their federal counterparts will not for now.

Taking a more holistic, adaptive approach has benefits beyond government and its voter and taxpayer constituents. Such an approach has much to offer the market players as well, given the large investments by Big Tech and the many industries keen to use AI. It offers a path away from extreme stances that do little to inform the public or catalyze public problem-solving: cavalier denial of economic security risks on one hand and warnings of imminent job apocalypse on the other.

But to date, elected officials and their constituents, along with regulated industries and public agencies, have engaged with only a fraction of the most immediate and sensitive AI uses that threaten human work, especially in recently enacted, judgment-focused human-in-the-loop laws. Beyond healthy caution and focus, policymakers, interest groups and the public at-large need a basic vocabulary for judging what job-relevant policy strategies and employer practices can and cannot hope to achieve, especially when it comes to ensuring human livelihoods on a large scale, as distinct from safeguarding high-stakes judgments—for example, about access to public benefits, child welfare, or parole—which have always been made by relatively few humans.

Mindsets matter, and so do imagination and clarity. We will need a wide-ranging, all-of-the-above mindset to understand and shape AI’s economic impacts, especially on the future of work. And given one, we should be honest with ourselves about what each kind of “solution” can realistically hope to achieve.

As a starting point, this report has offered a solutions framework and vocabulary that prioritize buying time to adapt, equipping (and not just informing) people so they can be more adaptive, steering both workers and employers in promising directions that reduce vulnerability (if only for a time), and opening up broader conversations about the role of work and ownership of wealth, including the variety of ways we might generate both meaning and economic security in changing societies, through and beyond paid work.

Notably, almost none of the response strategies mentioned to date, even the biggest “shifts”—from the four-day work week to some form of job or income guarantee—is a new idea. But as the long evolution and international variety of labor and safety net institutions suggest, such ideas might be repurposed and garner new support, or inspire better ideas, in a new and uncertain period of economic transformation. A close look at the politics of these options is beyond the scope of this framework report, but they offer myriad opportunities to overcome positional and partisan impasse. For now, the caricatured version of AI’s economic stakes offers a false and dangerous binary choice between innovation and an inclusive future.

Finally (and key to acting on those lessons), there is the central, underlying issue of state capacity, in the broad sense of effectiveness at all levels of government. Its shortcomings show in the delivery of many relevant legacy policies, such as workforce development and providing a safety net (let alone doing so in new ways), but also in detecting and deterring AI-powered employment discrimination, investing in novel work practices and standards, and other functions that government can uniquely perform or scale up.

To be clear, the coming AI-related economic disruption raises big policy design questions, not just delivery ones, as well as questions about our values. Furthermore, implementing new technology protections well will take time under any scenario. But the delivery failures, which the public experiences directly, are many, varied, and relentlessly amplified in our politics. They contribute to making too much of the public feel unseen and unserved, feeding back into a politics of grievance and mistrust—and the appeal of easy answers and blunt policy instruments. That is a far cry from a problem-solving government that is flexible, adaptive, and accountable for results.

The most sweeping economic transformation since the Industrial Revolution is now underway, deploying machines at once more powerful and less predictable than the steam engine, assembly line, or electric grid. Big Tech is at an impasse, but also perhaps a crossroads, with much of the public and many elected leaders: The public’s most consistent concern about this new industrial revolution is automation and degraded job opportunity, and as new state laws are suggesting, those risks turn out to be connected to safety and other risks already absorbing so much attention.

There are more building blocks already at hand than most of the public or elected officials realize, notwithstanding the unknowns. Brakes, buffers, steers, and shifts can all matter. It’s time to have a much clearer conversation about why each will be needed and what each might accomplish—and get on with the learning while doing.