The Great AI Jobs Transition

"America needs more jobs, and it needs them now.” 

It was true in 1971 when President Richard Nixon signed the Emergency Employment Act into law. It’s also true today and will likely remain so well into the future unless the government intervenes. A slew of factors, including but not limited to geopolitical discord, trade wars, artificial intelligence, and a reckoning from COVID-era business decisions, have led to a “low hire, low fire” labor market. People are staying put (or attempting to). Firms are holding on to their staff (or augmenting and automating with AI).  

This dynamic has put a particular tight squeeze on the job prospects of young Americans. The current unemployment rate for Americans ages 22 to 27 is 7.8 percent, nearly twice the national unemployment rate of 4.2 percent. It’s time to review several pages of economic history and prepare for a world in which young Americans find themselves in even more dire straits. 

The AI policy analysis offered by many outlets, including Lawfare and AI Frontiers, has generally focused on frontier model risk, export controls, and compute governance — all very important topics. The labor market consequences of the same technology have received comparatively little sustained attention from the national security community, even though a generation of scarred young workers would weaken the very industrial and human-capital base that AI competition is meant to protect. Congress’s response to AI workforce disruption is, in this sense, a national security question hiding in plain sight. 

This essay first outlines divergent forecasts of job opportunities, then explores how the federal government responded to similar concerns in the 1970s, and concludes with a set of recommended policy experiments that require public-private collaboration and have the potential to set America’s youth up for success.  

Conflicting Forecasts of Future Employment for Young People 

Economists have yet to reach consensus on whether AI and related economic trends will imperil the economic prospects of young Americans, empower them to earn higher wages in safer, more meaningful jobs, or some mix of the two. The ongoing dissensus may be a product of well-informed stakeholders forecasting wildly different futures. Expert forecasters, including the likes of Bharat Changer of the Stanford Digital Economy Lab, suspect that unemployment may hit 13.5 percent for recent grads by 2035. They also foresee underemployment nearing 60 percent for the young cohort. Dario Amodei, CEO of Anthropic, has likewise painted a bleak picture for entry-level workers in the near- to intermediate-term.  

On the other hand, some researchers--in addition to the Bureau of Labor Statistics--contend that heightened youth unemployment is not a new or startling development and suspect that AI will lead to a net increase in jobs, as have prior general-purpose technologies. They, too, have plenty of evidence to bolster their case. Data from the Federal Reserve suggests that AI exposure is reshaping where young workers find jobs, though not eliminating entry-level positions in the aggregate. AI-based job creation tallied 120,000 in 2024 alone. Construction jobs, typically staffed by younger workers, have surged by 30 percent since 2022. Google’s Chief Economist, Fabien Corto Millet, predicts that latent markets—industries that have yet to be created—will also offer young people an avenue for good jobs.  

Congress may prefer a wait-and-see approach to these divergent futures. Yet, if and when any sort of significant shock occurs in youth employment, those policymakers may wish they had taken that possibility more seriously. Any prolonged setback in the economic prospects of younger Americans comes at a high cost to the individual and to society generally. What’s more, those costs will likely span decades, exemplified by the fact that economists commonly refer to youth unemployment as a “scarring” event. 

Response to Sustained Unemployment in the 1970s 

Policymakers in the 1970s responded to related concerns around unemployment with a mix of substantive and symbolic policies. The history is instructive for a specific reason. The federal responses of the 1970s offer a controlled comparison between two approaches to large-scale labor disruption — one that moved quickly and pushed design authority outward to states and localities, and one that moved slowly and tried to centralize design in Washington. The first worked modestly. The second did not. That contrast, more than any program detail, is what should shape the contemporary response to AI. 

Two specific policies illustrate the difference between substantive and symbolic policy and suggest what may be helpful in response to contemporary labor market disruptions. The first is the Emergency Employment Act of 1971 (EEA). President Nixon signed the EEA with the intent of creating 150,000 jobs. The first “general public employment legislation” since the New Deal, it aimed to match unemployed workers — specifically, veterans of the Vietnam War and young people — with new, meaningful jobs within state and local governments. Nixon, who had previously vetoed similar legislation, explained, “local programs [under the EEA] will be designed with a view toward career advancement and toward developing new nonsubsidized careers for the worker.” He distinguished these kinds of jobs from “dead‐end, W.P.A.‐type jobs.”  

The bill, in many ways, worked as intended. More than 140,000 jobs were filled. Its success may have been a product of its simplicity. Congress raced to pass the EEA in response to widespread economic frustration, leaving little time for special interests to insert carve-outs and procedural checks into the bill. States and local governments had few boxes to check when it came to allocating funds and designing new careers.  

Yet the dearth of details may also have capped the EEA’s impact. Absent more exacting requirements, some governments hired for roles more akin to the “dead-end” jobs that Nixon hoped to avoid. Similarly, rather than reaching the unemployed most in need of assistance, jobs supported by the EEA tended to go to workers with high qualifications. Still, the Act managed to assist a substantial number of individuals and to launch a number of job opportunities with pathways to durable employment. 

The tangible results achieved by the EEA are especially noteworthy in hindsight, given the Carter Administration’s failures to address unemployment in the late 1970s. President Carter and several members of Congress acknowledged the need for bold action on the economy. The Humprey-Hawkins bill was introduced to jumpstart the economy. As initially drafted, the bill would require the President to develop a plan to achieve full employment and share it with Congress on an annual basis; local committees would assist by compiling and sharing the job needs of their employers; the expectation was that it would be the latest version of a “New Deal-style federal job creation program.”  

However, unlike the EEA, back-and-forth negotiations between competing interest groups watered down what could have been a substantive effort. As enacted, the bill was little more than an agreement that the federal government should aspire to create jobs. The editorial board of the New York Times summarized: 

It is a gesture from a Government overly enchanted with symbols, a promise from a Government that should mind its promises, a flag for politicians to wave when next they need to demonstrate concern for the unemployed. It will not, however, solve the nation’s unemployment problem. 

The task for contemporary officials is to avoid repeating the past by advancing symbolic legislation at a time when a substantive response is demanded. More specifically, policymakers ought to figure out how to apply the EEA’s playbook — quick enactment, minimal federal prescription, and design authority pushed to actors with proximity to the problem — to modern conditions. 

Contemporary Efforts to Navigate the AI Transition 

As of early May 2026, it seems unlikely that Congress can take the expedited approach that led to the EEA’s moderate success. Too many stakeholder groups have already staked out firm views on AI policy to suggest that anything can navigate the House and Senate without being pockmarked with policy favors for specific groups. That’s precisely why Congress ought to get creative about how it proceeds.  

The EEA’s lesson applies with unusual force here. If Congress cannot move quickly and cannot legislate without the kind of interest-group accretion that hollowed out Humphrey-Hawkins, then the federal role must shift. Speed and decentralization in 2026 cannot come from Congress itself. They must come from the private actors already running workforce experiments — with Congress playing the narrower, slower, but still essential role of identifying which experiments work and codifying the standards that allow them to scale. 

Policymakers keen to respond to this AI economic moment need not develop bespoke solutions; they can simply amplify private efforts already underway. The AI labs that teed up this transition period have initiated promising programs that ought to be coordinated, standardized, and codified. This approach combines private-sector ingenuity with the might of the federal government.  

First, there’s a severe need for rapid reskilling programs that can direct people into high-demand sectors. Meta and CBRE are set to launch a free, four-week program to accomplish just that. Their LevelUp initiative is open to folks with all sorts of backgrounds and credentials (or lack thereof) who aspire to become fiber technicians. Participants will learn how to install fiber-optic cables, network equipment, and other data center components. Graduates will presumably be well placed to earn one of the 349,000 construction jobs that have yet to be filled.  

The government should closely watch this program and treat it as a private-sector test pilot that, if successful, deserves to be scaled and spread. Notably, this may come at a fairly low public expense. Clearly, there’s already a private demand for such programming, so the question is instead how the government can merely complement what’s already underway. For instance, this may look like grants to successful applicants to the program so that they can cover expenses while undergoing training. Such financial support may be especially necessary for younger workers, given their limited savings. Alternatively, a support role for the federal government may involve making government space available for related training programs, reducing the total training program expenditure, and increasing the likelihood that it can be offered to more individuals. Finally, the government can lend its research expertise to Meta and CBRE (and other private stakeholders that follow) to help select future destinations and training programs. This complementary assistance could become a standard add-on to qualifying reskilling programs that, like LevelUp, are free, responsive to high-demand, growing sectors, and, critically, have demonstrated track records of successful job placements. 

Second, there’s a tremendous shortage of robust statistics on AI adoption, which hinders policymakers’ ability to design responsive policies. Traditional surveys are unlikely to uncover who is using AI, when, why, and for what purposes. For one thing, it’s quite hard to precisely define what “AI adoption” even means. Technically, both someone using Claude to take on extensive, hours-long coding tasks and someone using Claude to figure out how to make it look like you’re working while on a trip are AI users. A basic assessment of AI use at work might even treat those very different use cases as the same. For obvious reasons, such abstract metrics will do little to shape nuanced policy.  

Anthropic has discovered a much better approach. Their Clio tool—short for Claude Insights and Observations—analyzes Claude usage in real time. By relying on privacy-protective tools, Anthropic can gain a deep sense of which tasks are being augmented or automated by AI. They can also break this information down by geography, user demographics, and other essential fields. The net result is a detailed window into how AI is shaping work. Policymakers could use this information to support retraining opportunities like those previously mentioned, getting a sense of which tasks and jobs seem more AI-proof than others. Additionally, this information may be critical to shaping AI literacy programming by revealing which communities and groups seem to be behind the AI curve.  

However, Anthropic is only one lab, which diminishes the value of its insights. Congress can and should weigh conditioning federal procurement of AI tools on whether the AI company has implemented similar tools and made the resulting data generally available. This procurement hook could also specify standardized definitions and metrics, as well as safeguards for user privacy. Assuming broad compliance across the frontier labs, this information could be shared in real-time and made broadly available to local and state officials as well as nonprofits studying workforce development.  

Third, and finally, Congress should closely monitor and consider supporting or mirroring OpenAI’s certification courses. The company recently launched two initiatives: a generic, AI basics course that aims to provide workers with training on how to use today’s AI tools, and a course specific to teachers. Students will earn a certification that they can then share with future employers. There are several open questions about this initiative, including how the certification is regarded by employers (i.e., does it bolster the job prospects and pay of certification holders), whether job-specific training should be carried over to other professions, especially those that have a clear nexus with public well-being, and whether certification holders continue to advance in their AI prowess on their own accord over time.  

The government has several potential roles to play in spreading the benefits from such courses. For one thing, it should help individuals looking to upskill or retrain find the best certification program for their needs. Imagine, for example, a dashboard of different AI certification programs based on price, tailored to any profession, and efficacy in terms of job outcomes. Additionally, the government may need to assist with devising common certification metrics so that these programs do not become a valueless signal to employers. If everyone is “AI-certified” by different labs using different metrics, then such training is unlikely to be rewarded in the labor market. Congress need not get into the weeds of such a nuanced endeavor, but may want to designate an expert body to oversee that coordinating function in the event that fragmentation is not resolved by market forces.  

Where to Go From Here  

The through-line connecting these three proposals is not a grand legislative vision but a posture — one that treats AI labs and other private actors as the first movers in discovering policy interventions that Congress can then scale, with a particular eye toward younger Americans. That posture has historical warrant. The EEA succeeded, to the extent it did, because it moved quickly and let states and localities do the designing. The contemporary analog is not a single omnibus AI workforce bill but a tiered approach in which private programs run experiments, the federal government identifies what works, and Congress codifies the standards, metrics, and support that enable successful pilots to scale. This is unglamorous work. But it may be the straightest path to move the needle for the 22-to-27-year-olds currently facing a 7.8 percent unemployment rate. 

There is also a narrower point worth making to the AI labs themselves. The programs highlighted here — LevelUp, Clio, and the OpenAI certifications — are promising precisely because they were built by the firms closest to the technology driving the disruption. That proximity is an asset, but it is also an obligation. Labs that have argued, persuasively, that their tools will reshape the labor market cannot then treat the workforce response as somebody else’s problem.  

The next iteration of these programs should be designed from the outset with public partnership in mind: standardized metrics, shared data, interoperable certifications, and an honest accounting of placement outcomes. Congress will eventually act on AI and labor. The question is whether it acts on a foundation the labs helped build or on one assembled in their absence. The former is almost certainly better policy. For the labs, it is almost certainly better politics. 

Kevin Frazier leads the AI Innovation and Law Program at the University of Texas School of Law and is a Senior Fellow at the Abundance Institute.