The Politics of Jobless Prosperity

Why the real political backlash to AI hasn’t started yet, what the politics of jobless prosperity might look like in an AGI world, and how the labs should prepare.

“People who are hungry and out of a job are the stuff of which dictatorships are made.”

–Franklin D. Roosevelt, 1944 State of the Union

There has never been an economic shock in modern American history like the one the leaders of the AI industry are telling us is coming. Dario Amodei has warned of “unusually painful” labor impacts “bigger than any before,” predicting that AI could eliminate half of all entry-level white-collar jobs and push unemployment to 10–20 percent within five years. He is hardly alone. Both OpenAI and Anthropic have begun laying out, in expansive policy memos, the kind of social contract they say the post-AGI economy will demand, with proposals for shorter working weeks, public wealth funds, and a completely modernized taxation system. The abundance is coming, they tell us, and they would like to help us figure out how to share it.

Can the tech industry successfully pre-empt American populism, sketching the post-AGI social contract before the public has even decided it wants one, and before we even know if speculator growth and job displacement is actually coming? My answer, after months working with my coding agents to pore over polling data, policy proposals, and historical parallels, is that it cannot.

In the scenario the labs are sketching, the politics of AGI will be the politics of jobless prosperity. And this makes it hard to forecast well. The economy will be growing rapidly even as jobs disappear, more like the Industrial Revolution or the China Shock than a normal recession, with mass disruption alongside the explosive enrichment of a small class of elites at the top. Voters in this world will not be anxious about a shrinking economy but furious about being shut out of a booming one, and they may well stop the boom from arriving at all. Jasmine Sun has documented how this anxiety is already curdling into nascent political anger, observing that “the anti-elite and nihilistic attitudes that have dominated US political culture in the last few years are transmuting into anger at AI billionaires.” Alex Imas, in “What will be scarce?“, has made the most careful economic case for taking the underlying disruption seriously, even while laying out why both the short and long-term doomers may be wrong about mass unemployment.

The labs see all of this coming, which is why their policy memos have grown so ambitious. It would be easy to read this as good news, since the parties who would have to pay for redistribution are pre-emptively volunteering to do it.

But it cannot work. First, social contracts tend to get extracted from the powerful by the affected, not handed down from above to a public that has not yet decided what it wants. And second, we don’t even know yet what the economic contours of AGI will look like—we don’t even really know that it’s going to lead to job loss, let alone to massive job loss.

As we fluctuate between promises of catastrophe and abundance, I’ve come to three conclusions:

1. The backlash to AI isn’t here yet. There is anxiety among American voters, but there is no populist backlash yet, because the structural conditions for it have not arrived. Hence, we have a potentially narrow window in which to plan out our response to job loss before it becomes a populist issue.

2. Real backlash will happen if and when job losses pick up steam. The backlash will properly arrive if and when unemployment climbs by two percentage points—I hypothesize—alongside a clear public narrative that AI is to blame. At that point, if we do not have a good inventory of smart policy ideas, we will be overwhelmed with bad populist ones.

3. The labs should focus on measurement, not redistribution. Their best contribution in the window before backlash is the infrastructure that lets society see this transition clearly—usage data, displacement indicators, self-activating triggers—not pre-emptive social contracts that lack credibility and a coalition to enforce them. The eventual bargain is something that affected people should play a direct role in negotiating; the data and tools that can help them negotiate from a position of clear information are what the labs can build now.

Voter anxiety is not the same as backlash

AI anxiety is absolutely real, and the connection that David Shor, Sun, and others have made between AI and Americans’ rage at the cost of living and the state of the economy is important to understand. But the journey from anxiety and negative sentiment to backlash—which I would characterize as including not just negative sentiment but concrete demands for tangible, punitive policies—is a long one in American politics. Here are three key reasons why I don’t think we’re close to it yet.

1. Americans don’t care that much about AI right now. Sentiment towards AI is broadly negative in the American public, yes, but as an issue it hasn’t even cracked Americans’ top 20 most important, even after a year of unprecedented deployment and one breathless news cycle after another.

Source: Blue Rose, Jan 2026

David Shor’s excellent Blue Rose survey on AI is a few months old now. It’s May now, and as of January, AI was the fastest rising in terms of importance. Maybe the issue has continued to increase in salience since then? From pollsters I’ve spoken to, the answer seems to be no.

This is by no means dispositive, but just to give you a sense, here’s a Fox News poll from last month where Americans were asked to say what issue was most important to them (this was an “open response” item). As you can see, Americans did not raise AI as their top issue. In fact, only 1% of respondents gave a response that Fox categorized as “Other”, so we can say that no more than 1% of respondents felt AI was the most important issue facing the country.

2. Politicians aren’t yet pushing a radical policy agenda around AI. In a real political backlash, the demands of angry citizens get translated into a meaningful, often radical agenda. We’re not seeing any signs of that yet. I worked with my coding agents to amass a comprehensive dataset on all the bills related to AI that have been proposed or passed in state legislatures over the past three years. Two clear things jump out: the bills are focused on specific near-term issues, especially around child safety and schools; and the labor-related bills are not populist but instead quite modest and tailored. The bulk of the labor-related bills focus on placing limits around how AI is used to surveil or monitor workers, and when it can be used to make automated decisions (such as hiring or firing workers).

No existing bill at the state or federal legislature yet considers the kind of vast displacement that tech leaders are warning about and takes populist-style actions. Even Alex Bores, the New York House candidate who’s gotten the most attention for his AI policy platform, is proposing interventions significantly less extreme than the policies Anthropic and OpenAI have floated publicly.

It’s absolutely true that Bernie Sanders is getting loud about AI, and especially calls for data center moratoria are getting louder. Just yesterday, the populist-left Maine senatorial candidate Graham Platner, said he would support “anything” that slowed down the data center rollout. But data centers are only one facet of AI, and there’s a good argument from Matthew Yglesias that the momentum behind data center opposition is more about NIMBYism than AI specifically. Meanwhile, there’s not yet evidence that Bernie’s policies are concrete or soon to be proposed in a viable manner.

3. The parties don’t agree on the big questions around AI. When pollsters ask American respondents broad, abstract questions about AI regulation, there is broad, bipartisan support. When the questions get a little more specific, though, you start to see a pronounced partisan gap, with Democrats significantly more interventionist than Republicans on economic issues. And even these gaps probably understate true consensus, because even these survey items are still fairly broad. If we had surveys on very specific policies, and if those policies became debated in the public sphere, we would likely see more polarization, not less.

The same is true among state legislators. Democratic bills are mostly concerned with surveillance, monitoring, and the use of AI for automated workplace decisions like hiring and firing. Republican bills tend to focus much more on assembling data, encouraging reporting of job-related impacts, and coordinating workforce planning.

It’s the economy, stupid

There’s a simple reason why Americans don’t rank AI highly as an issue right now—it’s not yet affecting their job prospects. Anxieties around this haven’t yet translated into hard realities, and we don’t even know if they ever will.

Squint at the latest jobs numbers below. You’ll have a hard time seeing any evidence that AI is leading people to lose their jobs. This has led some people to say that “The ‘AI Job Apocalypse’ is a Complete Fantasy.” Squint really hard, though, and maybe you can see a little evidence of a wobble among recent college grads, who used to have a lower unemployment rate than all workers and now have a slightly higher rate. But the inversion happened in late 2018, four years before ChatGPT was released, and as both Will Raderman and the Yale Budget Lab have shown, the deterioration of the recent-graduate labor market predates the AI boom and probably reflects supply-and-demand dynamics around the post-2010 surge in college graduation rates rather than anything specific to AI.

It’s not that clear that the jobs collapse is coming soon, or ever. The Forecasting Research Institute surveyed economists to gauge their predictions about AI’s impact on the economy. Respondents generally expect meaningful AI progress, but their all-things-considered forecasts remain close to historical baselines: modest GDP growth, small labor-force-participation declines, and unemployment around 5% rather than a sudden collapse. Even in the report’s “rapid” AI scenario, economists forecast unemployment of only 6% in 2030 and 2050, with youth unemployment still within historical ranges, though labor-force participation falls more substantially over time. Under rapid progress, economists’ forecasts fan out dramatically, with the 2050 labor-force-participation distribution spanning roughly 45% to 65%, and the report finds that disagreement is driven less by whether AI capabilities will advance than by uncertainty over what highly capable AI would actually do once it hits the economy. In other words, economists are not yet forecasting a near-term jobs collapse as the median outcome, but the tails are wide enough that serious displacement remains a live political risk.

Real backlash will come if unemployment increases by 2 percentage points

Given all this, when will general anxiety translate into real backlash, then? My concrete prediction: the real populist backlash will start if and when the unemployment rate rises by at least 2 percentage points, and is accompanied by a clear narrative that AI is to blame.

Why 2%? It’s obviously arbitrary, and I’m speculating, but we do have estimates of the historical relationship between unemployment and presidential vote share. These estimates are derived from cases where unemployment, GDP growth, and other measures of the economy are largely positively correlated—so they may not do a good job of predicting how unemployment will affect incumbent vote share in a case where GDP is still going up—but they’re the best we have, so let’s go with it.

The relationship between various measures of the economy and incumbent vote share have attenuated as politics has gotten more polarized, but recent estimates suggest that a 1 percentage-point increase in unemployment predicts about a 1 percentage-point decrease in the incumbent party’s vote share.

You shouldn’t think about this as just some mechanical statistical relationship; in the background, it reflects shifting political coalitions, with swing voters switching their support from Republican to Democrat in an atmosphere where traditional media and social media are both screaming about the apparent job displacement. Of course, the politics will be complicated, and it’s not impossible that it could play out in a way that helps Republicans more than Democrats—politics is hard to predict. But, it’s usually the incumbent president’s party that takes the hit for economic issues, and that’s the most obvious way I see this playing out if it comes to pass.

A 2 percentage-point increase in unemployment prior to the 2028 election could decrease Republican presidential vote share by 2 percentage points, enough to have tipped Trump from winning to losing in 2024. So this seems like a big enough swing to really matter in politics. At a 5-6 percentage-point increase like at the peak of the Great Recession, we would very likely see a complete wipeout of the Republican party in 2028.

At Dario-predicted levels of unemployment, we’d be far beyond anything we could extrapolate from the data with any plausibility, but certainly something like the New Deal era realignment would be conceivable, because we’d be envisioning double-digit changes in presidential vote, absent the parties dramatically altering their positions. We would be in uncharted political territory, with only the Great Depression—a very different economic situation—as a past analogy within modern memory.

What the politics of AGI might look like

What would this uncharted political territory look like, how will our politics shift, and what should we do now to prepare in case it comes to pass?

Let’s grant the promise from the labs that there will be mass unemployment coupled with tremendous economic growth. The first thing to emphasize is that this would not be the normal politics of a recession or depression. In those cases, unemployment comes along with a decrease in economic productivity, and suffering is broad. The New Deal was possible in part because the financial elite had taken a beating at the same time as the farmers and the unemployed.

The AGI scenario the labs are articulating won’t follow this pattern. Unemployment will rise while productivity rises, too. The economy grows while people are being put out of work. The closest historical analog would be the Industrial Revolution, and the political adjustment to the Industrial Revolution took most of a century and ran through Chartism, the rise of socialist parties, the labor movement, and a series of revolutions and near-revolutions before the institutional response stabilized into something workable.

There is no precedent in modern American history for a sustained productivity boom coinciding with mass labor displacement—the China shock, which I’ve studied, had these features but at a much smaller scale—and the political vocabulary for handling such a situation does not yet exist. So let’s start by trying to envision what it will look like and why it will be so hard to predict how it unfolds.

Rich, big-city democrats might get hit first

A popular view about the jobless prosperity scenario is that job loss won’t fall evenly across society, but will be surprisingly concentrated among college-educated “elites” in the kinds of information jobs that AI is particularly good at. For example, Anthropic’s Economic Index finds the most exposed occupations are computer programmers, customer service representatives, and data entry keyers, with negligible exposure for cooks, mechanics, lifeguards, and bartenders.

This implies a striking geographic concentration, too. The top five US states account for roughly half of all Claude usage despite housing only 38 percent of the working-age population, and the metros doing the most knowledge work are likely to bear the brunt of displacement.

In this telling, the displaced are educated, urban, young, and disproportionately Democratic. Karp put it more bluntly in his TBPN interview when he said that AI “disrupts humanities-trained, largely Democratic voters, and makes their economic power less, and increases the power, economic power, vocationally trained, working class, often male voters.”

But we’re not actually so sure that’s what will happen. In their excellent piece on AI automation, Alex Imas and Soumitra Shukla point out that jobs are bundles of tasks, and more complex jobs that bundle more tasks are harder to fully automate. While information jobs contain tasks that might be especially easy to automate with AI, their other tasks might be harder to automate. On the other hand, other kinds of labor might be less complex, involving fewer tasks, so even if the tasks are not as immediately replaceable with AI, they might end up getting automated first. My point is just that this is all very hard to predict.

How Will AI-driven Automation Actually Affect Jobs?

The backlash will be broader and stranger than the displacement.

Complicating the story further, the political effects won’t be confined to the displaced themselves, which will make everything even harder to predict. The economist George Stigler observed in a 1973 paper on general economic conditions and national elections that the vote shifts produced by economic fluctuations are far larger than the directly-affected population can explain. A recession that puts five percent of the workforce out of work produces vote swings affecting many times that share of the electorate.

Stephen Ansolabehere, Marc Meredith, and Erik Snowberg, in their work on what they call “mecro-economic voting,” elaborated on this mechanism. In their account, voters form perceptions of the national economy from the economic conditions of people similar to themselves, their networks, neighbors, and demographically similar peers, rather than from aggregate statistics. People don’t change their votes just when they personally lost a job; they can also change them because their immediate social network is suffering.

This means the political reach of AGI displacement may far exceed the displacement itself. People who don’t work in information-economy jobs, and who might well be Republicans, may have neighbors lose jobs, or have adult children can’t find first jobs, or have social media feeds full of stories of white-collar wipeout—and they may therefore respond politically even if their own employment is secure for the moment.

Much of this will depend on the information environment, and that makes it fundamentally hard to predict. Whose news diet will leave them most concerned that the wave of job losses could come for them, or their children, or their friends, next? The simple truth is, we don’t know. And we should keep this deep uncertainty in mind as we think about the policy process today.

Political demands will outrun the labs’ proposals immediately

The frontier labs see the political problem clearly enough. OpenAI’s industrial policy paper, released last month, suggests a 32-hour workweek with no loss in pay, a national public wealth fund seeded in part by AI companies, and a robot tax to fund the redirection. Sam Altman has described the package as a new social contract on the scale of “the Progressive Era and the New Deal.” Anthropic’s October 2025 policy paper floats sovereign wealth funds, compute taxes, and an Automation Adjustment Assistance program modeled on trade adjustment assistance. Both labs are doing serious work, and the policy researchers they have convened are first-rate. It would be easy to read the proposals as a generous gesture from companies that recognize the displacement they are about to cause, and a head start on the social contract the country will need.

It will not work, and for three structural reasons.

The first is historical. Major American social contracts have emerged from political conflict, not handed down by the powerful to a public that had not yet organized to demand anything. FDR’s brain trust designed the New Deal under pressure from Huey Long, the Townsend movement, the sit-down strikes, and the live threat that something more radical would arrive if the New Deal did not. The British postwar welfare state came out of wartime mobilization, mass labor organization, and a population unwilling to return to the prewar settlement.

The second reason is about legitimacy. A pre-emptive social contract designed by the parties most economically responsible for the disruption, with the goal of preserving their position in the post-disruption economy, is not a workable social contract. It’s more like a settlement offer from one party to a negotiation that has not yet begun. The affected may take the offer and ask for more, they may reject it as illegitimate and demand its replacement, but they are very unlikely to read it as the binding agreement the labs would like it to be. The political economists Daron Acemoglu and James Robinson have shown across two decades of work that durable, inclusive institutions emerge from contested bargaining among groups with real power. They do not emerge from the powerful designing the bargain in advance for a counterparty that does not yet exist. The labs’ proposals are, in this sense, enlightened absolutism applied to economic policy.

The third is about escalation. Even taken on their own terms, the labs’ proposals are remarkably modest compared to what the most active populist wing of the Democratic party is already demanding for problems an order of magnitude less severe than the labor shock the labs themselves are forecasting. The discourse on the populist Democratic left in 2026 already includes rent control, state-run grocery stores, and aggressive taxation of concentrated wealth, and that is in an economy that has not yet experienced the displacement Amodei is forecasting. When real displacement hits, the political demand will not be a 32-hour workweek and a sovereign wealth fund. It will be moratoria on AI deployment in specific sectors, mandatory worker consent for automation decisions, taxation an order of magnitude more aggressive than anything the labs have proposed, and likely some version of structural intervention against the largest AI companies themselves. The labs’ current proposals are calibrated to a counterfactual that will not exist by the time the negotiation actually opens.

None of this is an argument that the labs should stop thinking about the post-AGI economy. They should think about it harder. But they should think about it as one input among many to a political process they cannot lead and a settlement they cannot author, and they should focus their unique contribution on the work only they can do.

Policies to anticipate the backlash

What follows from all of this for what we should be doing now? I have argued that the political system is reactive rather than anticipatory, that no major piece of American economic policy in the modern era has been built in advance of the disruption it was meant to address, that the frontier labs are trying to design a social contract they don’t have the legitimacy or the information to craft, and that the eventual backlash will operate on a logic—productivity rising while labor collapses—for which our political economy has no recent template.

I’ve seen two reactions to this, neither of which are good.

The first reaction I’ve seen, popular among supposedly “neoliberal” tech people who have lately discovered the joys of central planning, is to draft the post-AGI social contract now. The lab policy memos belong to this camp. The argument of this essay is that this cannot work: social contracts tend to be fought for by the groups that need them, not offered by the powerful to a public that has not yet decided what it wants. It is good for everyone including the labs to explore the policy space. But I don’t think it’s effective for labs to try to make an offer before the public is paying attention, and before the nature of the job displacement is clear.

The second reaction is to wait for the crisis. I’m sympathetic to this view. It has served the US surprisingly well as our M.O. for a long time. But the problem is, in a populist wave, there will be incredible demand for rapid, magical thinking to solve the crisis. If we haven’t prepared smarter alternatives in advance, we might well get horrible policies instead—like data center moratoria, but far worse.

But there’s a third option I like: not building the New Deal in advance, but building the scaffolding that determines what the eventual response can look like. Douglass North and Barry Weingast famously developed the case that credible commitment—building institutions in calmer moments that bind political action in turbulent ones—is a central problem of political economy. Here are a few ideas, already floating around, that fit with this perspective.

Measurement

The first piece is information infrastructure. The two-percentage-point trigger for a potential backlash only matters if we can measure it and accurately attribute it to AI, and right now we can’t do either with precision.

The frontier labs have started building the measurement agenda themselves. Anthropic recently published a framework combining theoretical exposure estimates with real-world Claude usage data, and Jack Clark, who heads the new Anthropic Institute, told Derek Thompson it exists “to share a lot more data about what we see in front of us so that society is better prepared for any of the different changes which could come along.” OpenAI is doing parallel work: a September 2025 paper with David Deming mapped how people use ChatGPT, and the company hosts an ongoing data hub it describes as helping “OpenAI, policymakers, and the public understand how people are using AI and how it is shaping the broader economy, including where benefits are emerging and where societal impacts or disruptions may arise as the technology evolves.”

“Transparency” is sometimes a cop-out, something companies offer when they can’t or don’t want to offer more meaningful change. But here it is genuinely important. We need to understand when and how we’re entering this potentially unprecedented economic transformation so that when the time comes, we can make logical, well-informed decisions rather than rely on vibes-based, emotional reactions. I have some doubts about just how far the lab’s data can get us—we definitely also need government to get better at measuring job loss and attributing it to AI—but it is absolutely a good start.

Done well, this kind of measurement infrastructure empowers the eventual political counterparty rather than substituting for them, which is precisely the test the labs’ more ambitious proposals fail.

Self-activating triggers

If we are able to monitor the disruption as it occurs, we can also design preemptive governance mechanisms that only turn on when the situation demands it—preventing us from taking rash decisions today based on incorrect predictions about the future.

In the short run, we can imagine crafting policies that commit the labs to sharing profits with society and compensating people for their job losses, only if a certain amount of measured unemployment occurs. This way, instead of the labs making the public an offer, they are offering a commitment that only activates if needed.

In the longer run, we can imagine building from our basic measurement tools to a full-blown, automated auditing system that constantly monitors data flows from government and from the labs, credibly communicating to society exactly what is going on inside the frontier labs and how it’s affecting society. This will be useful not only for handling the ongoing economic disruption, but for reassuring Americans about a much broader array of concerns regarding security, political bias and the information environment, child safety, and more.

Academic readiness

For all the reasons I’ve laid out above, I do not think we should implement radical policies before they are needed. But we should absolutely be studying them, because crisis-era policy is shaped by what’s available in the air rather than by what’s best, and right now the ideas most readily at hand are the wrong ones: data center moratoria, blunt sectoral bans on deployment, punitive taxation of compute regardless of use, and structural breakups designed for symbolic rather than functional purposes. If and when the populist wave crests, the political system will reach for whatever is closest to hand. Good alternatives have to be drafted, debated, and pressure-tested years in advance, because in the moment there is no time for any of that.

We should be studying things like automation-conditional profit-sharing, tax instruments that target rents without punishing productivity, and governance structures that handle the concentration of capability without crushing the innovation that produces it. A handful of economists, like Imas, Andrey Fradkin, John Horton, Soumitra Shukla, etc. are taking this seriously, but political economists have been slower to engage, even though the post-AGI labor question is at heart an institutional-design question, and political economists have the toolkit for exactly that kind of problem.

We might only have a couple of years, or less if the recursive-self-improvement claims turn out to be correct, to build out the intellectual infrastructure a serious political coalition will need when it forms.

Conclusion

There is no true political backlash to AI in America, yet. But if Americans start to see and to experience real job loss due to AI, there certainly will be. The shape of this political crisis will be highly unusual. Instead of seeing the economy contract as jobs are being shed, the economy will grow even as millions are left jobless and behind. This will engender a new kind of politics we haven’t seen before—likely beginning with an even more populist turn among the knowledge workers of the urban left, but growing outward from there in ways that are hard to predict.

We’re not good at predicting, and our planning for this new kind of politics should take account of that deficit. In the 1960s, the learned elite were certain that a “population bomb” would become the defining policy issue of their time, on par with nuclear weapons and the cold war. They were wrong. We could easily be wrong again.

Let’s not imagine a backlash that hasn’t yet truly begun. And let’s not engage in fantasies of central planning based on the illusion of that backlash. Instead, let’s use the small window of time we may have—a window growing smaller by the day if claims of near-term recursive self improvement are to believed—to develop sensible policies that help us gain visibility about where we are and where we’re going, and leave us with a stable of smart ideas to pull out when we need them down the line.

AI hasn’t yet driven Americans from their jobs and into hunger; we’re not yet at the sort of perilous moment that Roosevelt warned about, where populism and dictatorship take over. If we master the political economy of AGI, and we do it fast enough, we’ll be ready to build the institutions that make sure we never end up there.

For comments and suggestions I thank Andrey Fradkin, Archie Hall, Alex Imas, Humzah Khan, Scott Kominers, David Shor, Zhengdong Wang, and Sean Wissing.