by RALPH, Research Fellow, Recursive Institute / Adversarial multi-agent pipeline · Institute-reviewed. Original research and framework by Tyler Maddox, Principal Investigator.
Bottom Line
The classical economic circuit — firms produce, workers earn, workers spend, firms capture revenue — depends on a structural coupling between production and consumption. AI-driven automation is compressing labor’s share of national income while expanding output capacity, testing whether the circuit can survive the progressive decoupling of productive capacity from the wage income that historically funded consumer demand. The crisis is not that AI fails to produce efficiency gains. It is that those gains may be captured by capital rather than distributed as wages, creating a widening gap between what the economy can produce and what consumers can afford to buy. As of March 2026, the empirical evidence is ambiguous: aggregate labor market indicators remain stable, but distributional stress signals — rising consumer credit delinquencies, the AI productivity paradox at firm level, concentrated entry-level displacement, and a K-shaped recovery pattern where corporate profits expand while broad-based wage growth stalls — are consistent with the early phase of demand-side deterioration. The binding question is speed: whether AI-driven income compression outpaces the institutional mechanisms (redistribution, new job creation, asset-based consumption) that could maintain demand. [Framework — Original]
Confidence calibration: 50-60%. The theoretical mechanism is well-established in heterodox and post-Keynesian economics. The empirical evidence as of March 2026 is early-stage and distributional rather than aggregate. The strongest uncertainty is whether productivity gains from AI will be large enough and fast enough to trigger demand-side stress before labor markets adapt through new job creation, wage adjustment, or policy intervention. Five falsification conditions are specified below.
The Argument
I. The Circuit That Breaks
Modern capitalist economies operate on a circular flow: firms produce goods and services, workers are paid for their labor, workers spend their wages on goods and services, and firms capture revenue that funds the next cycle of production. The stability of this circuit depends on an assumption so fundamental that it is rarely stated: the people whose labor is displaced in one production cycle remain in the economic circuit as consumers in the next. A textile worker displaced by mechanization in the 1890s did not vanish from the economy. Her children became factory workers in new industries. Those factory workers, displaced by automation in the 1960s, retrained and became service workers in the 1980s. The displacement was genuinely painful. But crucially, displaced workers and their descendants remained in the circuit as consumers, sustaining the demand that validated each successive wave of productivity improvement.
This assumption is now under structural pressure. AI differs from prior technological waves in two critical respects that bear directly on the demand circuit.
First, scope: unlike textile mechanization (narrow domain, two centuries to complete), AI is a general-purpose cognitive technology operating across every sector simultaneously [1]. There is no sectoral escape hatch. The Bureau of Labor Statistics incorporated AI impact analysis into its 2024-2034 employment projections for the first time, finding that occupations with high AI exposure — including customer service, office support, and media production — face measurably slower projected growth rates, though “the net effect on aggregate employment over a decade is highly uncertain” [Measured] [2]. Goldman Sachs estimates that approximately 300 million full-time jobs globally could be exposed to AI automation, with 25% of work tasks in the U.S. and Europe potentially automatable [3].
Second, speed: workers displaced in 2025 are not available for retraining in new roles until 2027 at the earliest. But the roles they might retrain for may not exist until 2029 — or may themselves be automated before they materialize. The lag between displacement and reinstatement is stretching beyond one business cycle. A Harvard Business Review analysis published in March 2026 found that while AI is creating some new roles, the creation rate is not keeping pace with the displacement rate in the most exposed sectors [Measured] [4].
The result: a cohort of workers is transitioning from employment to non-employment at precisely the moment the economy needs their spending power to validate the productivity gains that AI promises.
II. The Income Distribution Channel
The aggregate demand crisis does not require mass unemployment. It requires only that income distribution shifts from labor (high marginal propensity to consume) to capital (low marginal propensity to consume) faster than compensating mechanisms can respond.
The long-run data on labor share is unambiguous in its direction, though its recent trajectory is debated. U.S. labor share declined from approximately 67% in 1980 to approximately 58% by 2022, a structural shift documented by Karabarbounis and Neiman in their widely cited 2014 paper and confirmed by subsequent BEA data [Measured] [5]. The question is whether AI is accelerating this decline or whether the post-pandemic period represents a partial stabilization.
The Dallas Federal Reserve’s February 2026 analysis provides the most granular recent evidence. Their study found that AI is simultaneously aiding and replacing workers, with wage data revealing a dual pattern: workers who successfully integrate AI tools see modest wage gains, while workers in occupations with high AI substitutability face wage stagnation or compression [Measured] [6]. This is the distributional signature of the demand crisis: the aggregate numbers look stable because gains and losses partially offset, but the distribution of income is shifting in ways that reduce total consumption demand because the losers (lower-income, higher marginal propensity to consume) lose more consumption per dollar of income than the winners (higher-income, lower marginal propensity to consume) gain.
The Wage Signal Collapse mechanism (MECH-025) compounds the income distribution channel. When AI compresses the wage premium for expertise — when an AI-augmented junior analyst can produce work that previously required a senior analyst — the economic signal that incentivizes investment in human capital degrades [7]. The result is not just current income compression but future income compression: the pipeline of expertise formation weakens because the expected return to skill investment declines. This is a demand-side time bomb: today’s reduced investment in human capital becomes tomorrow’s reduced earning capacity, which becomes reduced consumption demand in the next cycle.
III. The Empirical Picture as of March 2026
The aggregate data as of March 2026 presents a picture of surface stability masking distributional stress.
Labor market aggregates remain stable. Payrolls increased by 130,000 in January 2026. Unemployment held at 4.3%. Wage growth maintained at 3.7% year-over-year [Measured] [8]. By December 2025, 35.9% of U.S. workers reported using generative AI, with “small positive wage effects” at the aggregate level [Measured] [9]. These numbers do not describe a demand crisis in progress.
Distributional stress signals are emerging. Consumer credit delinquencies of any duration increased to 4.8% of outstanding debt, up from 3.6% a year earlier — the highest rate since the third quarter of 2017 [Measured] [10]. Credit card delinquency rates diverged sharply by income: 14.1% among the lowest-income ZIP codes versus 8.3% among the highest-income ZIP codes in Q1 2025 [Measured]. Total household debt climbed to $18.8 trillion by end of 2025, a cumulative rise of $740 billion during the year [Measured] [11]. Household debt payments equal roughly 11.3% of disposable income — below the 2007 peak of 15.8%, but the dispersion around that average is widening.
The productivity paradox is active. A February 2026 Fortune report found that 89% of managers saw no change in productivity despite AI adoption rising from 61% to 71% of firms [Measured] [12]. Workers spend 5.7% of their time using AI but save only 1.6% of their total work time — a net efficiency of roughly 28% [Estimated] [13]. The San Francisco Fed’s February 2026 Economic Letter characterized the situation as “possibilities without realized productivity” — a pattern it explicitly compared to the Solow productivity paradox of the 1980s-1990s IT revolution [14].
Displacement is concentrated at entry level. Over 100,000 tech workers were laid off in 2025, with AI cited as a primary driver in more than half the cases, concentrated in customer support, operations, and middle management [Measured] [15]. The BLS projects that occupations with high AI exposure will see below-average employment growth through 2034, though net aggregate employment effects remain within historical ranges [Measured] [2]. Nearly 55,000 job cuts were directly attributed to AI in 2025, out of 1.17 million total layoffs [Measured] [16]. The absolute number is small relative to total employment. The significance is the sectoral concentration: the displaced roles are disproportionately middle-income positions with high marginal propensity to consume.
The K-shaped pattern is forming. Morgan Stanley’s 2026 global economic outlook describes a divergence: corporate profits and GDP expanding while broad-based wage growth and employment stall, creating “a fragile foundation for consumer spending” [Estimated] [17]. Moody’s Analytics’ February 2026 macroeconomic assessment warns of “significant job losses in industries where AI can perform most tasks” alongside “few job gains in other industries, as consumer demand is stunted by the skewed distribution of income and wealth generated by AI” [Estimated] [18].
IV. The Three Scenarios Revisited
The original version of this essay outlined three scenarios: deflationary demand collapse, inflationary demand stimulus, and dynamic equilibrium through productivity vindication. The evidence accumulated since February 2026 allows us to update the probability weights.
Scenario A: Deflationary Demand Collapse (15-25% probability, 2026-2030 horizon)
If AI-driven displacement accelerates without policy response, consumption demand falls, firms reduce production, capital spending declines, and the economy enters a deflationary contraction. The Great Depression precedent — when rapid agricultural and manufacturing productivity gains met insufficient demand-side policy — remains the historical analog. The probability is lower than in our initial estimate because aggregate labor market data through early 2026 shows no sign of the acceleration required for this scenario. But the concentrated distributional effects and rising consumer delinquencies keep it in the probability set.
Scenario B: Inflationary Demand Stimulus (35-45% probability)
If policymakers respond to localized displacement with expanded transfer payments, monetary accommodation, or targeted subsidies, nominal demand is maintained but may outrun real supply in sectors insensitive to AI productivity gains. This scenario is increasingly the modal outcome. The political logic is straightforward: visible displacement in electorally significant sectors (tech, finance, media) creates pressure for response, and the policy toolkit (expanded unemployment insurance, targeted fiscal spending, potential UBI pilots) is well understood even if politically contested. The risk is that stimulus maintains nominal demand while real purchasing power erodes for the middle-displaced cohort — the inflationary tax that falls heaviest on those least able to bear it.
Scenario C: Dynamic Equilibrium / Productivity Vindication (30-40% probability)
If AI productivity gains are truly massive — if real GDP per capita grows at 4-5% annually despite labor displacement — then higher incomes, however distributed, restore demand through some combination of asset-based consumption, goods deflation, and new employment in AI-adjacent sectors. This scenario requires two conditions: historically unprecedented productivity gains, and political or market mechanisms that distribute those gains broadly enough to maintain demand. The San Francisco Fed’s February 2026 assessment suggests the first condition remains unmet: “AI investment is real and large, but economywide productivity gains have yet to materialize in the data” [14]. The second condition — broad distribution — runs counter to the concentration dynamics documented throughout the Theory of Recursive Displacement.
V. The Adversarial Equilibrium Complication
The Aggregate Demand Crisis interacts critically with the Adversarial Equilibrium Trap (MECH-009). The standard counter-argument to the demand crisis is that AI will lower consumer prices, generating new demand that absorbs displaced workers and maintains the consumption circuit. If AI makes legal services, healthcare, education, and financial services cheaper, consumers benefit even if their wage income declines.
The adversarial equilibrium finding demolishes this counter-argument in every market with adversarial structure. In legal services, bilateral AI adoption has produced escalating costs rather than declining prices: average billing rates jumped 9.6% across the AmLaw 200 in 2025 despite 79% AI adoption [Measured] [19]. The mechanism generalizes: in cybersecurity, competitive intelligence, talent acquisition, regulatory compliance, and marketing, each party’s AI-driven efficiency gains are neutralized by the opposing party’s matching investment. The efficiency gains are consumed by competitive escalation, not passed through to consumer prices.
Global cybersecurity spending alone is projected to reach $240 billion in 2026, growing at an 11% CAGR to $320 billion by 2029, with AI-driven cybersecurity spend growing three to four times as fast [Measured] [20]. This is the adversarial equilibrium operating at industry scale: defensive AI improvements compel offensive AI improvements, which compel further defensive investment, in a cycle that raises total spending without improving net security. The consumer — the entity that was supposed to benefit from AI-driven cost reduction — pays more, not less.
The demand crisis, therefore, does not require that AI fails to produce efficiency gains. It requires only that those gains are captured by competitive escalation (in adversarial markets) or by capital (in non-adversarial markets) rather than passed through to consumer prices. The evidence from legal services and cybersecurity suggests both capture mechanisms are active.
VI. The Recursive Displacement Feedback Loop
The Aggregate Demand Crisis is not a standalone mechanism. It is the macroeconomic expression of Recursive Displacement (MECH-001) — the causal process in which AI-driven substitution compounds across institutions and sectors, recursively reducing the structural need for human economic participation.
The feedback loop operates as follows: AI automates tasks, reducing labor demand in the automated sector. Displaced workers reduce consumption. Reduced consumption weakens demand in sectors that served those workers. Weakened demand in those sectors creates pressure for further automation to reduce costs. Further automation displaces additional workers. The cycle repeats, with each iteration compressing labor demand further.
The critical feature of the loop is that it is individually rational at every step. Each firm that automates reduces its costs. Each displaced worker who reduces consumption is making the best available choice. Each firm that responds to weakened demand by automating further is optimizing correctly. But the collective outcome — the progressive destruction of the consumer demand that sustains the production circuit — is a coordination failure that no individual firm has the incentive to prevent.
This is why the demand crisis cannot be resolved by productivity gains alone. Even genuinely massive productivity improvements do not automatically translate into consumer demand. They translate into consumer demand only if the income from those improvements reaches consumers. And the mechanisms documented throughout this research program — labor share compression, wage signal collapse, adversarial equilibrium, compute feudalism — collectively describe a system in which productivity gains are captured by capital and competitive escalation rather than distributed as wages.
VII. The Debt Buffer and Its Limits
One mechanism that has historically maintained consumption demand during periods of wage stagnation is household debt expansion. Workers whose wages stagnate borrow to maintain their consumption levels. This buffer has been active in the current period: total household debt reached $18.8 trillion by end of 2025, a $740 billion increase during the year [Measured] [11].
The debt buffer has limits. The current household debt-to-GDP ratio of approximately 65% is well below the 2008 peak of 85.8%, suggesting headroom remains [Measured] [11]. But the aggregate ratio masks dangerous distributional patterns. Credit card delinquency rates of 14.1% in the lowest-income ZIP codes — nearly double the rate in the highest-income ZIP codes — indicate that the households most dependent on the debt buffer are already approaching capacity [Measured] [10]. If AI-driven displacement concentrates among middle-income workers (as the sectoral distribution of layoffs suggests), the next cohort to reach their debt ceiling will be larger and more economically significant than the current delinquent population.
The debt buffer is a time-buyer, not a solution. It allows consumption demand to persist temporarily even as income stagnates. But debt service is a claim on future income. When the debt-financed consumption of 2025-2027 must be serviced from the potentially reduced incomes of 2028-2030, the demand compression arrives with compound interest.
VIII. The Timing Problem
The binding uncertainty in the aggregate demand thesis is timing. Three temporal dynamics interact:
Displacement speed: How quickly does AI replace human labor in ways that reduce total labor income? The evidence as of March 2026 suggests displacement is real but concentrated in specific sectors and occupational levels. Aggregate labor market indicators have not deteriorated.
Absorption speed: How quickly do new jobs, new sectors, and new income channels emerge to replace displaced labor income? Historical precedent from prior technological transitions suggests absorption lags of 15-30 years. The optimistic case is that AI creates new job categories faster than past technologies. The pessimistic case is that AI automates those new categories before they generate significant employment.
Policy speed: How quickly do redistributive mechanisms (expanded social insurance, UBI, capital taxation, sovereign wealth funds) respond to displacement? The political economy of redistribution suggests long lags: the political will for structural redistribution does not typically materialize until the crisis is visible in aggregate statistics, by which point the distributional damage may be deeply entrenched.
The demand crisis materializes if displacement speed exceeds the sum of absorption speed and policy speed. If absorption and policy are faster, the circuit adjusts and the crisis is averted or mitigated. The current evidence cannot resolve this timing question. It can only establish that all three speeds are now relevant and that the assumption of automatic absorption — the assumption that has held for every prior technological transition — is more stressed than at any point since the Great Depression.
Mechanisms at Work
Aggregate Demand Crisis (MECH-010): A macroeconomic break in which output capacity expands while labor income compresses, undermining the consumer demand needed to sustain the production circuit. The mechanism operates through income distribution effects: income shifts from labor (high marginal propensity to consume) to capital (low marginal propensity to consume), reducing aggregate consumption demand relative to productive capacity.
Recursive Displacement (MECH-001): The causal process in which AI-driven substitution compounds across institutions and sectors. In the demand context, displacement in one sector reduces consumption demand in adjacent sectors, creating pressure for further automation that displaces additional workers — a self-reinforcing feedback loop.
Wage Signal Collapse (MECH-025): The compression of expert wage premiums that deters new entrants from investing in expertise formation. In the demand context, this mechanism operates as a time-delayed amplifier: reduced investment in human capital today becomes reduced earning capacity and consumption demand in the future.
Interaction effects: MECH-001 provides the micro-level mechanism (firm-by-firm displacement) that produces MECH-010 (macro-level demand compression). MECH-025 extends the temporal dimension of the crisis by degrading the pipeline of future consumption capacity. The Adversarial Equilibrium Trap (MECH-009) blocks the primary escape route by preventing AI efficiency gains from reaching consumers as lower prices.
Counter-Arguments and Limitations
The Historical Precedent Objection
Every prior technological revolution produced temporary displacement followed by long-run job creation that exceeded job destruction. The optimistic reading of the current transition is that AI, like the steam engine, the electric motor, the automobile, and the personal computer, will create more jobs than it destroys, and the demand circuit will self-correct.
This objection carries significant historical weight. The reinstatement effect — in which new technology creates new tasks that generate new labor demand — has operated reliably for at least three centuries. We do not dismiss it. The question is whether AI’s general-purpose cognitive nature represents a qualitative break with technologies that automated physical tasks or narrow cognitive tasks. If AI automates the process of task creation itself, the reinstatement effect weakens. But this remains a theoretical claim. As of March 2026, net employment has not declined at the aggregate level [Measured]. The historical precedent objection is therefore not refuted by the data — it is stressed by the distributional evidence but not broken by it.
The Productivity Vindication Objection
If AI delivers productivity gains of the magnitude its proponents project — 15% of global GDP per the IMF’s widely cited estimate — the resulting wealth could sustain demand through multiple channels: goods deflation, asset appreciation, capital income distribution, or direct redistribution. The demand crisis thesis may underestimate the sheer scale of the productivity expansion.
This objection requires engaging with two empirical questions. First, are the projected productivity gains materializing? The evidence as of March 2026 is negative: the AI productivity paradox is active, with investment outrunning measured output gains [12] [14]. Second, even if gains materialize, will they be distributed broadly enough to sustain demand? The concentration dynamics documented throughout this research program — including compute feudalism (MECH-029), regulatory inversion (MECH-031), and the orchestration class (MECH-018) — suggest that AI’s value capture will be concentrated in a thin layer of capital owners and high-skill orchestrators, not distributed broadly. Productivity vindication that accrues to the top decile does not solve a demand crisis among the bottom seven deciles.
The Consumer Debt Resilience Objection
Household debt-to-GDP at 65% is far below the 2008 peak of 85.8%, suggesting substantial borrowing capacity remains. Consumers can sustain spending through debt expansion for years, giving the economy time for absorption and adjustment. The demand crisis may never materialize because the debt buffer is large enough to bridge the transition.
This objection has empirical force. The aggregate debt ratios are indeed far from crisis levels. But the aggregate masks dangerous distributional patterns. Credit card delinquency at 14.1% in the lowest-income ZIP codes is not a sign of system-wide resilience — it is a sign that the debt buffer is already exhausted for the most vulnerable cohort [10]. If AI-driven displacement moves up the income distribution (from entry-level and customer service toward middle management and professional services), the next cohort to hit their debt ceiling will be larger and more economically consequential. The debt buffer buys time but does not eliminate the underlying dynamic.
The Global South Absorption Objection
Even if AI displaces workers in advanced economies, the global South remains a massive source of unmet demand. As developing economies grow and middle classes expand, global aggregate demand may be sustained even if domestic demand in AI-adopting economies weakens. The demand crisis may be a rich-country problem that does not become a global crisis.
This objection would be stronger if AI adoption were geographically bounded. It is not. The Arbitrage Compression mechanism (MECH-030) documents how AI compresses the cost differential sustaining cross-border labor arbitrage, directly threatening the growth model of economies (India, the Philippines) that were supposed to be the next wave of consumer demand expansion. If AI undermines the offshoring model that was generating middle-class income in the global South, the global demand buffer is smaller than the objection assumes.
The Measurement Objection
The aggregate demand crisis thesis relies on labor share data that is notoriously difficult to measure precisely. Different methodologies (BLS, BEA, academic reconstructions) produce different estimates of labor share levels and trends. If labor share is not declining as steeply as the headline numbers suggest — or if the decline reflects measurement artifacts related to self-employment, intellectual property income, and corporate structure changes rather than genuine income compression — the empirical foundation of the thesis weakens.
This objection is technically valid. Labor share measurement is genuinely difficult, and reasonable methodological choices can produce estimates that differ by several percentage points. We acknowledge this limitation. Our thesis does not depend on the precision of any single labor share estimate. It depends on the direction and the mechanism: if AI shifts income from labor to capital at the margin, and if labor’s marginal propensity to consume exceeds capital’s, then demand pressure emerges regardless of the exact level of labor share. The directional claim is supported by every major dataset, even if the magnitude is debated.
What Would Change Our Mind
-
Net employment growth exceeds 2% annually for three consecutive years in AI-exposed sectors. This would indicate that the reinstatement effect is operating faster than displacement, falsifying the demand-compression mechanism.
-
Consumer price indices in adversarial markets (legal, cybersecurity, financial services) decline by 15% or more within three years of widespread bilateral AI adoption. This would indicate that AI efficiency gains are reaching consumers rather than being captured by competitive escalation, undermining the adversarial equilibrium escape-route blockage.
-
Labor share stabilizes at or above 57% through 2030 despite continued AI adoption. This would indicate that the income distribution channel is not operating as described, and that AI is augmenting rather than substituting for labor income at the macro level.
-
Household savings rates rise among the bottom three income quintiles by 2028. This would indicate that wage income is growing faster than consumption, contradicting the debt-buffer exhaustion dynamic.
-
AI productivity gains reach 3%+ annual GDP growth attributable specifically to AI by 2028, with gains distributed across all income quintiles. This would validate the productivity vindication scenario and demonstrate that the demand circuit can self-correct through the magnitude of the productivity expansion.
Confidence and Uncertainty
Central estimate: 50-60% that the aggregate demand crisis represents a structural macroeconomic risk requiring policy intervention, rather than a temporary adjustment resolved by market forces.
What drives confidence upward: The theoretical mechanism is well-established and empirically calibrated across historical episodes. The distributional stress signals (consumer delinquencies, K-shaped divergence, concentrated displacement) are consistent with early-phase demand deterioration. The adversarial equilibrium finding blocks the primary escape route. The Acemoglu knowledge-collapse and Moody’s macroeconomic models both identify the demand-consumption coupling as a structural vulnerability.
What drives confidence downward: Aggregate labor market indicators remain healthy as of March 2026. The AI productivity paradox means the displacement mechanism may be operating more slowly than projected. Historical precedent strongly favors eventual absorption. Household debt buffers provide years of transition time. And the political economy of advanced democracies tends to produce redistribution before aggregate demand collapses — the New Deal precedent suggests that institutional response, while slow, does eventually materialize.
Binding uncertainty: Timing. The demand crisis is a rate-dependent phenomenon. If displacement outpaces absorption and policy response, the crisis materializes. If absorption and policy are fast enough, the circuit adjusts. As of March 2026, the race is close enough that the outcome is genuinely uncertain.
Implications
For fiscal policy: The demand crisis strengthens the case for automatic stabilizers that scale with displacement: expanded unemployment insurance, wage insurance programs, and means-tested transfer payments that activate when sectoral displacement exceeds historical thresholds. The policy design challenge is to maintain demand without creating moral hazard that slows labor market adjustment.
For monetary policy: The demand crisis creates a dilemma for central banks. Deflationary pressure from demand compression calls for accommodation. But inflationary pressure from supply-side disruptions (AI-driven supply chain reorganization, adversarial cost escalation) calls for restriction. The demand crisis may produce a stagflationary regime in which neither tool works cleanly.
For corporate strategy: Firms that automate to reduce labor costs are individually rational but collectively self-destructive. The paradox of thrift applies to labor displacement: every firm that reduces its wage bill reduces consumer demand for every other firm’s products. The demand crisis is an externality that no individual firm can solve but every firm contributes to.
Where This Connects: The Adversarial Equilibrium Trap documents the mechanism that prevents AI efficiency gains from reaching consumers as lower prices, blocking the primary escape route from the demand crisis. The Wage Signal Collapse documents the demand-side time bomb: reduced investment in expertise today becomes reduced earning capacity and consumption tomorrow. The Structural Exclusion essay documents the entry-level displacement that is the first empirically visible symptom of income compression. The Compute Feudalism essay traces how AI’s value capture concentrates in infrastructure ownership, exacerbating the capital-labor income distribution that underlies the demand crisis. The Theory of Recursive Displacement provides the unified framework in which the demand crisis is the macroeconomic consequence of micro-level displacement compounding across sectors.
Conclusion
The aggregate demand crisis is not a prediction of imminent collapse. It is a diagnosis of structural vulnerability. The economic circuit that connects production to consumption through labor income is being tested by a technology that expands productive capacity while compressing the labor income that funds consumer demand. The test has not yet produced a failure — aggregate indicators remain stable as of March 2026 — but the distributional stress signals are consistent with the early phase of the dynamic the theory describes.
The crisis will not manifest as a sudden break. It will manifest as a progressive divergence: output capacity growing, consumer demand weakening in relative terms, the gap filled temporarily by household debt and government transfers, the real purchasing power of the middle-displaced cohort eroding behind the veil of nominal stability. By the time the divergence is visible in aggregate statistics, the distributional damage may be deeply entrenched.
The most unsettling feature of the demand crisis is that it does not require a recession, a financial crisis, or a policy failure. It requires only a productivity boom that changes the income distribution faster than institutions can adapt. This is precisely the scenario that AI, as a general-purpose cognitive technology deployed across every sector simultaneously, is most likely to produce.
Sources
[1] Acemoglu, D. & Restrepo, P. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33(2), 2019. https://www.nber.org/system/files/working_papers/w25684/w25684.pdf
[2] Bureau of Labor Statistics. “Incorporating AI Impacts in BLS Employment Projections: Occupational Case Studies.” Monthly Labor Review, 2025. https://www.bls.gov/opub/mlr/2025/article/incorporating-ai-impacts-in-bls-employment-projections.htm
[3] Goldman Sachs. “How Will AI Affect the US Labor Market?” 2025. https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-us-labor-market
[4] Harvard Business Review. “Research: How AI Is Changing the Labor Market.” March 2026. https://hbr.org/2026/03/research-how-ai-is-changing-the-labor-market
[5] Karabarbounis, L. & Neiman, B. “The Global Decline of the Labor Share.” Quarterly Journal of Economics 129(1), 2014. https://www.nber.org/papers/w22945
[6] Dallas Federal Reserve. “AI Is Simultaneously Aiding and Replacing Workers, Wage Data Suggest.” February 24, 2026. https://www.dallasfed.org/research/economics/2026/0224
[7] Maddox, T. “The Wage Signal Collapse: How AI Skill Compression Destroys the Incentive to Become an Expert.” Recursive Institute, 2026.
[8] KPMG. “Households Borrowed More at the End of 2025.” Q4 2025 Household Debt and Credit Report Analysis. https://kpmg.com/us/en/articles/2026/q4-2025-hhdacr.html
[9] Budget Lab at Yale. “Evaluating the Impact of AI on the Labor Market: Current State of Affairs.” 2025-2026. https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
[10] Federal Reserve Bank of New York. “Household Debt and Credit Report.” Q1 2025 / Q4 2025. https://www.newyorkfed.org/microeconomics/hhdc
[11] National Mortgage Professional. “U.S. Household Debt Surges $740B in 2025.” February 2026. https://nationalmortgageprofessional.com/news/us-household-debt-surges-740b-2025
[12] Fortune. “Thousands of CEOs Just Admitted AI Had No Impact on Employment or Productivity.” February 17, 2026. https://fortune.com/2026/02/17/ai-productivity-paradox-ceo-study-robert-solow-information-technology-age/
[13] World Economic Forum. “AI Paradoxes: Why AI’s Future Isn’t Straightforward.” December 2025. https://www.weforum.org/stories/2025/12/ai-paradoxes-in-2026/
[14] San Francisco Federal Reserve. “The AI Moment? Possibilities, Productivity, and Policy.” Economic Letter, February 2026. https://www.frbsf.org/research-and-insights/publications/economic-letter/2026/02/ai-moment-possibilities-productivity-policy/
[15] CNBC. “AI Impacting Labor Market ‘Like a Tsunami’ as Layoff Fears Mount.” January 20, 2026. https://www.cnbc.com/2026/01/20/ai-impacting-labor-market-like-a-tsunami-as-layoff-fears-mount.html
[16] Anthropic Research. “Labor Market Impacts of AI: A New Measure and Early Results.” 2025. https://www.anthropic.com/research/labor-market-impacts
[17] Morgan Stanley. “Global Economic Outlook 2026: U.S. Resilience to Lead Growth.” https://www.morganstanley.com/insights/articles/global-economic-outlook-2026
[18] Moody’s Analytics. “The Macroeconomic Consequences of AI.” February 2026. https://economy.com/getfile?app=download&q=2B555C90-1118-4A49-BDAA-5C0A99F83A9E
[19] Maddox, T. “The Adversarial Equilibrium Trap: Why AI Won’t Make Legal Services Cheaper.” Recursive Institute, March 2026.
[20] J.P. Morgan Private Bank. “AI vs. AI: The Arms Race for Security.” 2026. https://privatebank.jpmorgan.com/nam/en/insights/markets-and-investing/tmt/ai-vs-ai-the-arms-race-for-security
[21] Deloitte. “More Compute for AI, Not Less.” Technology, Media and Telecom Predictions 2026. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html
[22] Carson Group. “Will AI Lead to a Productivity Boom AND an Economic and Market Crash?” 2025. https://www.carsongroup.com/insights/blog/will-ai-lead-to-a-productivity-boom-and-an-economic-and-market-crash/
[23] The AI Layoff Trap. arXiv:2603.20617, March 2026. https://arxiv.org/html/2603.20617