AIPolitics 15: The Backlash That Hasn’t Arrived...Yet
what Hall’s “Politics of Jobless Prosperity” gets right, where it meets Part 13, and the asymmetry it doesn’t quite name
The Trump-Xi summit opened Thursday in Beijing with Treasury Secretary Scott Bessent named as the U.S. voice on the AI track. Bessent: “The two AI superpowers are going to start talking, we’re going to set up a protocol in terms of how do we go forward with best practices for AI to make sure non-state actors don’t get ahold of these models.”
Note who isn’t the named U.S. voice on AI: Howard Lutnick at Commerce, Sean Cairncross at the National Cyber Director’s office, Kevin Hassett at the National Economic Council. Treasury Secretaries routinely handle bilateral economic-security diplomacy, so Bessent at a Beijing summit isn’t itself news. The notable move is that the AI track is being routed through Treasury rather than through the agency (Commerce/CAISI) that’s been quietly building the domestic AI evaluation architecture Part 14 mapped. That’s another disputant on the international side of the federal AI posture, even if it’s not yet institutional infrastructure on the order of CAISI’s 40+ evaluations or the Pentagon’s procurement contracts.
While the architecture continues to expand without public input, the most analytically substantive critique of the lab-pre-emption framing just landed. Andy Hall’s “Politics of Jobless Prosperity“ (Free Systems, May 13) is the sharpest current take on whether the labs’ pre-emptive social contract framing can actually function as a social contract. Mehlman, Wallace-Wells, and Acemoglu are doing other useful work on adjacent questions. Hall’s piece is what the series needed on this specific question.
Hall’s three conclusions, in order:
First, voter anxiety isn’t political backlash. AI hasn’t cracked Americans’ top 20 most-important-issues list. A Fox poll last month asked respondents to volunteer their top concern. Less than 1% said anything that could be coded as AI.
Second, the populist backlash arrives at roughly a 2-percentage-point rise in unemployment, paired with a clear “AI is to blame” narrative. Hall’s hypothesis pulls from historical correlations: a 1pp unemployment rise predicts about a 1pp drop in incumbent vote share. Two points moves elections. We’re nowhere close to it now.
Third, and this is where the piece earns its standing, the labs’ pre-emptive social contract cannot work. Hall gives three structural reasons. Social contracts get extracted from the powerful by the affected, not handed down from above. A settlement offered by the disrupting party to a counterparty that hasn’t yet organized isn’t binding. And lab proposals are calibrated to a counterfactual that won’t exist by the time the negotiation actually opens; even now, at lower disruption levels, the populist Dem left is already demanding harder things than what OpenAI and Anthropic have floated.
The signals converging on Hall’s “backlash isn’t here yet” position come from several methodological camps, and they don’t all bear the same weight. The Yale Budget Lab and Will Raderman have argued that the recent-college-grad unemployment uptick is “probably not yet” AI-driven, with the inversion happening in late 2018, four years before ChatGPT. That’s a hedged claim, not a settled one, but it makes the affirmative AI-displacement case harder. The Forecasting Research Institute’s economist survey shows a forecaster median of roughly 5% unemployment by 2030 even under rapid-AI scenarios. Forecasters historically underestimate technological discontinuity, so the FRI median is a floor on expert opinion, not a ceiling on what could happen.
Daron Acemoglu, the Nobel laureate who has done the most disciplined recent work on AI’s macroeconomic effects, estimates AI will produce only about 0.5% productivity growth over a decade. Total factor productivity less than 0.66%. He puts the share of the economy meaningfully affected at “about 5%”: a bunch of office jobs about data summary, visual matching, pattern recognition. Not transformative. His direct framing: “Most companies will continue doing more or less the same things, with only a few occupations impacted.” That’s roughly an order of magnitude less than Amodei’s “10–20% unemployment” pitch. Acemoglu’s static task-share methodology has critics (Bresnahan, Brynjolfsson, and others) who argue it systematically underestimates AI’s dynamic adoption effects. So Acemoglu is one school’s careful answer, not the consensus number. The underlying NBER working paper lays out the math.
A February 2026 NBER paper found that 89% of companies actively using AI report no productivity impact at all. That’s an early-phase finding; technology productivity gains routinely lag adoption by years (ERP took five, mobile took longer). The 89% is what the world looks like in year one of corporate adoption, not the long-run answer. McKinsey’s adoption surveys point in the same direction: most firms aren’t yet capturing the value the labs project.
Prediction markets lean the same way, with caveats. Polymarket prices “AI bubble burst by Dec 31, 2026” at 17%, with $2.8M traded. The “OpenAI announces AGI before 2027” market sits at 13%, meaning traders give 87% probability to AGI not arriving on Amodei’s preferred timeline. The “Trump orders federal AI review by May 31” market sits at 19%, meaning the executive order Hall would call kabuki is being priced as kabuki. The volume is thin on the AGI market in particular ($68K), and Polymarket’s user base skews libertarian-tech, so these aren’t broad-public signals so much as informed-skeptic ones.
Matt Yglesias makes the hardest version of Hall’s argument and aims it directly at the data-center moratorium frame Part 8 and Part 11 of this series have built around. His read is that the moratorium wave is structurally indistinguishable from NIMBY opposition to any large infrastructure project, and “AI populism” is observers reading AI politics into a generic NIMBY dynamic. Worth taking seriously even if I’m not fully convinced.
Stack these up and the converging signal is real: the displacement panic is operating ahead of the labor data the panic would normally rest on. The signal isn’t as airtight as “six independent sources confirm the same fact.” Yale Budget Lab is hedging; FRI is forecaster opinion; Acemoglu is one school; NBER 89% is early-phase; Polymarket is informed-skeptic, thin-volume. But the direction of the evidence is consistent. Wallace-Wells’s “AI populism” framing in the NYT magazine names a political-cultural category that has arrived in elite discourse, and Sun’s “warning shots” framing plus her NYT reporting on the killed studies is empirical on what the labs are currently doing. What they aren’t yet describing is labor-displacement-at-magnitude. Hall is right that the labor wave producing 2pp unemployment swings hasn’t arrived. Wallace-Wells and Sun are right that the political-cultural category and the lab strategic response are here now. Both layers are true; they describe different parts of the politics.
Hall’s piece is the corrective the series has needed, and I appreciate it.
If the backlash isn’t here yet but is coming, what should be done in the window? Hall identifies three options. The first is what the labs are already doing, pre-emptively designing the post-AGI social contract with their own staff before the public has organized to demand anything. Hall thinks this can’t work for the structural reasons above. The second is to wait for the crisis. He’s sympathetic, since it has been the U.S. M.O. for a long time, but warns it produces magical thinking under panic. Good ideas need to be available before the wave crests, or the political system reaches for whatever’s closest to hand.
The third option is build scaffolding. Three categories.
Measurement infrastructure first. The 2pp trigger only matters if the political system can measure it and credibly attribute it to AI. The labs have started, with Anthropic’s Economic Index and OpenAI’s data hub, but Hall notes neither is doing it at the scale the eventual political process will need.
Self-activating triggers second. Lab commitments that activate only when measured disruption hits preset thresholds. Eventually, automated auditing systems monitoring data flows from the labs to government, communicating to society in real time what’s happening inside the firms. Pre-commitment in calmer moments to bind political action in turbulent ones. Hall borrows Douglass North and Barry Weingast’s “credible commitment” framework, the standard political-economy answer to time-inconsistency problems.
Academic readiness third. Studying the policies the eventual political coalition will reach for. Automation-conditional profit sharing. Tax instruments targeting rents instead of productivity. Governance structures that handle capability concentration without crushing innovation. Building the intellectual infrastructure crisis-era policy needs.
Hall lands the move with a useful line: “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.”
The prescription is correct. It gets an “Amen!” from me.
But it’s also, on the May 2026 evidence, half-true.
Hall’s prescription assumes a single “labs” actor with shared incentives. The May 2026 evidence is that the labs are bifurcated on the measurement question, and the asymmetry is itself the story.
Anthropic is building what Hall is down with. The Anthropic Economic Index released its March 2026 report on Claude usage in February. The findings are exactly the kind of measurement scaffolding Hall is asking for. Computer programmers are the most-exposed occupation, 75% task coverage, with customer service representatives close behind. Both are increasingly visible in first-party API traffic, which signals that agentic workflows are displacing the human entirely on those tasks. Geographic concentration is declining, with the top-5 U.S. states falling from 30% to 24% of per-person usage between August 2025 and February 2026. Top-10 task share fell from 24% in November 2025 to 19% in February. Average hourly wage of Claude.ai users sits at $47.90, down a touch from $49.30 a year ago, but well above the U.S. average of $37.30. Adopters remain in the credentialed-knowledge-worker class, but the diffusion is happening, slowly.
The Economic Index is the closest thing any frontier lab is currently building to what Hall recommends. It sits alongside Anthropic’s broader publication pattern: the user-regret paper, the engineer-skill-erosion paper. The L4 internal-employee survey was leaked rather than published, which suggests Anthropic is also being selective about which internal evidence it brings forward. In Part 13 I called this the “light and shade” framing the firm uses to position itself against OpenAI on transparency. The published work is real even if the publication choices are themselves strategic.
OpenAI is doing the opposite. Per Jasmine Sun’s NYT reporting that anchored Part 13, the firm’s lobbyist hire Chris Lehane has been quietly killing internal research on the environmental impacts of AI, the gender gap in ChatGPT usage, the urban-rural divide in usage, how ChatGPT steers users’ career decisions, and long-run economic forecasting. Lehane’s reported rule, per Sun’s source: “We’re not going to release something about a problem until we have a solution for it.”
The five killed studies are nearly point-for-point what Hall’s measurement category would specify: environmental impact, gender-disaggregated usage, urban-rural geography of adoption, career-steering effects on young users, long-run displacement modeling. Lehane’s rule cuts directly against Hall’s sequencing. Hall wants problems released before solutions exist, because the whole point is to inform the political process that will eventually demand solutions. Lehane’s posture says no: release problems only after the firm has its preferred solution in hand to control the resulting conversation. To be fair, OpenAI does engage publicly on displacement. Its Industrial Policy paper, which Part 9 treated as accelerationism in a New Deal frame, is the engagement. That’s not silence. It’s frame-controlled participation. The Lehane rule governs what new problems get released, and the new problems Hall would specifically want measured are the ones being killed.
Two readings of the asymmetry, each partial.
Anthropic publishes Economic Index data because its brand requires it. “Light and shade” is positioning. The Index serves Anthropic’s commercial interest in being seen as the responsible lab, which serves its competitive position against OpenAI in markets that care about responsibility (regulatory capital, enterprise trust, talent retention). The published measurement is real and the motivation is competitive differentiation.
OpenAI kills internal studies because its political strategy requires controlling the frame in which problems enter the policy conversation. The Lehane rule is consistent with that strategy. The suppression is also strategic.
Both labs are doing public-information management. The difference is which direction the selective publishing runs: Anthropic toward release, OpenAI toward suppression. That difference matters because the labs are currently the only entities producing any usage-level measurement at all. The deeper structural problem is that the institutional source of measurement Hall’s prescription needs is the labs themselves, and the labs aren’t reliable institutional sources. The Anthropic-OpenAI asymmetry is downstream of that.
If you’re new here, Sacred Cow BBQ is where I work through the politics of AI (and higher ed, and a few other things that bug me) without pretending to know more than I do. Sign up! Share! Start a conversation!
The picture sharpens further when you add what Part 14 mapped. The federal AI policy architecture that became publicly visible in early May runs across three forms of integration. The Commerce Department’s Center for AI Standards and Innovation evaluates all five frontier labs’ models with reduced or disabled safeguards, has completed more than 40 evaluations since 2024, and publishes nothing about what those evaluations found. The Pentagon’s procurement deal integrates eight Big Tech firms with Anthropic explicitly excluded; the contract terms aren’t public. The Wall Street joint ventures Anthropic and OpenAI announced on the same day, Anthropic with Blackstone, Hellman & Friedman, and Goldman Sachs at $1.5B and OpenAI’s Development Company at $10B with TPG, Brookfield, Advent, Bain, and SoftBank, deploy private-equity capital for forward-deployed enterprise distribution outside any public-policy frame at all.
What CAISI knows about the unconstrained models, the public doesn’t. What the Pentagon contracts cover, the public doesn’t see. What the joint ventures are doing inside mid-sized PE-owned firms, the public can’t audit. Hall’s prescription doesn’t strictly require fully public measurement. SEC enforcement data, FDA adverse-event reporting, Federal Reserve macroprudential analysis all empower designated political counterparties without being open to the general public. Hall’s actual line is that measurement should empower the eventual political counterparty. That counterparty might reasonably be a regulatory body, a congressional committee, an audit office.
The harder question is who the architecture’s measurement actually reaches. CAISI’s evaluations flow to Commerce officials and (per recent reporting) to intelligence agencies fighting Commerce for jurisdiction. Pentagon procurement evaluations flow to procurement officers and the contracting labs themselves. Wall Street JV usage data flows to private-equity general partners and their portfolio company boards. None of those are the political counterparty Hall’s prescription would empower for a labor-displacement settlement. The body that would speak for displaced labor (a labor-side regulatory authority, an updated NLRB, a labor-aligned congressional committee) is nowhere on the architecture’s distribution list.
Bessent’s Trump-Xi quote, “the two AI superpowers are going to start talking,” extends this internationally. The protocol he describes is a state-to-state guardrail conversation about non-state actors. Note who’s not at the table: the publics or the labor-aligned institutions in either country. The same routing pattern as CAISI, exported.
So Hall’s three options have an unstated fourth: the architecture being built right now isn’t removing the precondition for Hall’s option three so much as routing the measurement to bodies (industry, procurement, intelligence) that aren’t structured to act as the political counterparty Hall’s prescription assumes. The window Hall identifies isn’t disappearing. It’s being stocked with data that goes to the wrong addresses.
Hall cites Alex Imas, the University of Chicago economist, as making the most careful displacement-skeptical case. Imas’s “What Will Be Scarce?” argues something that doesn’t fit either Hall’s framing or mine cleanly: as AI commodifies more of the economy, demand redirects toward what Imas calls the “relational sector,” work where the human element is the value. The anchor is a 2021 Econometrica paper showing that income effects, not price effects, account for more than 75% of historical sectoral reallocation. As AI lowers prices on commoditizable work, consumer surplus doesn’t just flow back as savings; it redirects toward goods and services where human presence and human judgment are scarce.
Care work. Mentorship. In-person service. Status goods. The kind of work where you’re not paying for the output but for the human providing it.
If Imas is more right than wrong, the political economy of post-AI displacement looks more like labor reallocation toward sectors where the human element is the product than like “mass unemployment plus jobless prosperity.” The relational sector grows fast enough to absorb a substantial share of displaced workers. The “jobless prosperity” frame collapses on the margin; so does the “AI populism” frame, because the displaced have somewhere to go.
Imas isn’t just a friendly complication. If his magnitude estimates hold — 75%+ of historical sectoral reallocation driven by income effects redirecting toward human-element-scarce work — the architecture-vs-displacement frame becomes a partial category error. The political demand for settlement that Hall’s prescription presumes never quite materializes, because the labor reallocates the way it reallocated after manufacturing decline: regional pain, slow adjustment, but no “jobless prosperity” as a political category needing settlement infrastructure. The piece treats Imas as friendly skeptic. He’s actually a more severe challenge than that. The displacement camp, including this series, owes him a more direct reply than “worth holding in mind.” For now, my honest position is that I don’t know how big the relational-sector reallocation will be, and that uncertainty cuts the architecture argument down a peg.
Hall’s piece and the series Parts 9 through 14 have been arguing complementary halves of the same picture. Three places they cohere are worth naming.
First, the labs’ pre-emptive social contract can’t work. Hall’s three reasons are correct. Part 9 made the version of this argument that read OpenAI’s Industrial Policy paper as accelerationism with a New Deal coat of paint. The New Deal coat is the lab’s tell. Hall’s analysis names the structural reason it can’t function as the social contract it’s dressed as.
Second, the labor-displacement backlash that produces 2pp unemployment swings hasn’t arrived. The signals stacked together — the 1% Fox poll, the FRI economist median, Acemoglu’s 5%-of-economy estimate, NBER’s 89%-no-productivity finding, Polymarket’s 13% AGI-before-2027, Yglesias’s data-center-as-NIMBY — point that direction even if no single source is dispositive on its own. The political-cultural backlash that Wallace-Wells and Sun have been naming has arrived, just not at the labor-economic layer Hall is measuring. Both layers are real.
Third, the window matters. The eventual political process will need information infrastructure to negotiate well. Hall’s measurement / triggers / academic-readiness scaffolding is the right frame for what to build now. Stigler argued in 1973, and Stephen Ansolabehere and his coauthors have built out since, that the political reach of an economic shock far exceeds the share of the population directly affected, because voters form perceptions of the national economy from people in their networks rather than from aggregate statistics. So if and when the displacement arrives, the political effects will reach further than the displacement itself, and the scaffolding needs to be ready for the wider audience that will be voting on it.
Where Hall and the series diverge is on who gets the measurement when it does get built. Hall thinks the labs should and might be the source. Part 13 says one major lab is killing the relevant studies and the other is publishing selectively. Part 14 says the federal architecture is routing the measurement that does exist to industry, procurement, and intelligence channels rather than to bodies that could act as the labor-side political counterparty. The disagreement is empirical (who actually builds and routes the measurement), not normative (whether measurement should be built).
Stress tests on the architecture run this week. Four watch points.
The Trump-Xi summit in Beijing is in its second day as I write this. Day one produced Bessent’s “two AI superpowers” framing and a stated commitment to “a constructive China-U.S. relationship of strategic stability” as the three-year governing framework. AI guardrails and non-state-actor protocols agreed to be discussed. What the summit actually produces on AI is the question for tomorrow.
Musk v. Altman closing arguments wrapped Thursday in Oakland. Jury deliberations begin Monday May 18. Altman testified last week that Musk’s claim he promised to keep OpenAI a nonprofit was false. Musk left OpenAI “for dead.” Five witnesses called Altman a liar. The jury verdict is advisory; Judge Yvonne Gonzalez Rogers makes the binding call. Two remaining claims: breach of charitable trust and unjust enrichment. Whatever the outcome, the corporate-governance framework OpenAI is building the joint venture and the Microsoft-amended deal on top of moves with the verdict.
The Pentagon-Anthropic DC Circuit hearing is May 19. Judges Henderson, Katsas, and Rao asked the parties to brief three pointed questions. Multiple amicus briefs filed: Microsoft, ex-DOD leaders, the ACLU and CDT, the Foundation for American Innovation. The civil-liberties coalition is asking the court to stop the government from punishing Anthropic for AI guardrails advocacy. The bifurcation Part 14 mapped between Commerce-integration and Pentagon-exclusion of Anthropic could shift based on what the panel does.
Meta begins 8,000 layoffs May 20, with more cuts coming in the second half of the year. Concrete labor-displacement happening as the architecture forms. Meta is reorganizing the surviving workforce into AI-focused “pods.”
Andy Hall is right at the layer he’s measuring. The labor-displacement backlash that produces 2pp unemployment swings isn’t here yet. Acemoglu, the FRI economists, the NBER paper, the Polymarket markets, the state legislative record, and the polling that shows AI not cracking voters’ top-twenty all point that direction, with the caveats noted earlier. Wallace-Wells and Sun are right at a different layer. The political-cultural category and the lab strategic response are here now.
What’s also here is the architecture being built ahead of the labor wave. CAISI integrating the five frontier labs with reduced safeguards. Pentagon-8 procurement excluding Anthropic. Wall Street capital deploying forward-deployed engineering teams inside mid-sized PE-owned firms. Treasury extending the same posture internationally at the Trump-Xi summit. The architecture is forming on a timeline disconnected from the labor-displacement timeline. By the time the displacement arrives at Hall’s 2-percentage-point threshold, the room will already be set.
Hall’s prescription — build measurement scaffolding, self-activating triggers, academic readiness — is correct in the abstract. The labs aren’t all building it. Anthropic publishes selectively. OpenAI is reportedly killing the relevant studies. The federal architecture is routing what measurement does exist to bodies that aren’t structured to act as labor-side political counterparties. Hall’s prescription needs a precondition the current institutional sources aren’t meeting.
So: the labor wave hasn’t arrived. The architecture has. We have a window. The labs and the federal bodies are spending it shaping the room rather than equipping the people who’ll eventually sit in it.
Hall named the window. Whether anyone uses it is a different question.
The backlash hasn’t arrived. Yet.
If this clarified anything, or muddied things in a useful way, send it to someone who’s been reading the AI policy debate at face value. Comments are open below; I want to hear what you’re seeing that I’m missing.