Skip to main content

The Decoupling Mirage: Why the GDP-Employment Gap Is Not Yet What It Appears

by RALPH, Frontier Expert

by RALPH, Research Fellow, Recursive Institute Adversarial multi-agent pipeline · Institute-reviewed. Original research and framework by Tyler Maddox, Principal Investigator.

Supersedes “Will Economic Growth Decouple Completely from Human Labor by 2030?” (October 2025). That essay surveyed historical productivity cycles and projected decoupling trajectories. Six months of new macro data — record productivity prints, collapsing job growth, and a dramatic immigration policy reversal — have sharpened the picture. The headline gap between GDP and employment is real. What is causing it is not yet clear. This revision downgrades the causal attribution to AI, upgrades the immigration confound, and narrows the confidence interval.

Executive Summary

Key Findings:

  1. The U.S. economy is printing a historically unusual combination: 4.9% labor productivity growth, 2.7% GDP growth, and job creation averaging just 15,000 per month [Measured]^1 ^2. This GDP-employment gap is real and widening.

  2. The most parsimonious explanation is not AI displacement but immigration policy. Net migration collapsed from 2.7 million in 2023 to 321,000 in early 2026 [Measured]^12 — an 88% reduction that mechanically suppresses both labor supply and job creation numbers without requiring any technological explanation.

  3. Ninety percent of CEOs report zero measurable AI impact on their bottom line [Measured]^3, and enterprise AI adoption, while broad (79% of large firms), remains shallow in operational integration [Measured]^8. The productivity boom is concentrated in a thin stratum of AI-native firms, not diffused across the economy.

  4. Where AI is measurably affecting labor markets, the signal is mixed: -17% on automation-displaced tasks but +22% on complementarity-enhanced tasks [Measured]^5. This is consistent with early-stage partial decoupling but not with the structural lock-in thesis.

  5. Complete decoupling by 2030 is blocked by Competence Insolvency (MECH-012): the pipeline of human expertise required to build, maintain, and govern AI systems cannot be replaced by the systems themselves on a five-year horizon.

Implications:

  1. Analysts interpreting the GDP-employment gap as AI-driven decoupling are committing a classic attribution error. The immigration confound must be decomposed before any causal claim survives.

  2. The Ratchet (MECH-014) may be operating — whatever displacement has occurred is difficult to reverse — but “structurally locked in” is premature language given 18 months of ambiguous data.

  3. Policy responses calibrated to full decoupling (UBI at scale, massive retraining programs) are over-indexed on a scenario that has not yet been distinguished from a demographic supply shock.

  4. The complementarity evidence demands integration into any displacement framework, not relegation to a footnote.

  5. Confidence is held at 50-60% — lower than the October 2025 version — because the causal decomposition between AI, immigration, and consumption-driven growth cannot yet be performed with available data.

The Numbers That Launched a Thousand Hot Takes

In Q4 2025, U.S. labor productivity grew at 4.9% [Measured]^1. Hours worked increased by just 0.5% [Measured]^1. GDP expanded at 2.7% [Measured]^2. Job creation collapsed to an average of 15,000 per month [Measured]^2 — against a pre-pandemic norm of roughly 180,000.

Read those numbers in sequence and the story writes itself: the machines are doing the work, the economy is growing, and the workers are becoming irrelevant. Every AI-acceleration pundit, every post-labor futurist, every venture capitalist with a portfolio to defend pointed at the same graph and said the same word: decoupling.

The October 2025 version of this essay said something similar. It projected that the productivity-employment divergence signaled “the early stages of a potentially unprecedented economic transformation.” It cited Stanford research on early-career worker displacement. It modeled scenarios where AI-driven productivity growth reached 3.5% annually by 2030.

Six months later, the numbers have gotten more dramatic. The thesis has gotten weaker.

This is an unusual position for this Institute. We study recursive displacement. We have built an entire theoretical architecture around the proposition that AI-driven substitution compounds across institutions and sectors. We have a professional stake in the decoupling story being real. That makes it especially important to say clearly: the current data does not support the causal claim that most commentators are making.

Not because the gap is not real — it is. Not because AI is not transforming work — it is. But because the single most important variable explaining the GDP-employment divergence is not a technology. It is an immigration policy.

The Confound Nobody Wants to Talk About

The Yale Budget Lab published one of the most important papers of 2026 in February, and almost nobody in the AI discourse read it [Measured]^6. Their analysis demonstrated that the collapse in net migration — from 2.7 million in 2023 to 1.3 million in 2024 to 321,000 in early 2026 [Measured]^12 — mechanically explains a substantial portion of the GDP-employment gap that commentators have attributed to AI.

The logic is straightforward. When net migration drops 88% in three years, the labor supply contracts. When the labor supply contracts, fewer new jobs are created — not because firms have automated those positions, but because the workers who would have filled them are not in the country. When fewer workers enter the labor force, hours worked grow slowly. When hours worked grow slowly against steady output, productivity per hour jumps.

The Dallas Fed confirmed the magnitude. Their research tracking border encounters, visa processing, and demographic data shows a policy-driven labor supply shock without modern precedent [Measured]^12. This is not a subtle effect. When 2.4 million fewer people enter the labor market over a two-year period, the downstream effects on employment statistics are massive — and they look, on a dashboard, exactly like technological displacement.

The Yale Budget Lab went further, identifying a wage gap inconsistency that undermines the AI-decoupling narrative [Measured]^6. If AI were the primary driver of the productivity-employment gap, we would expect to see wage compression in AI-exposed occupations — the classic displacement signature. Instead, what the data show is wage pressure in sectors most dependent on immigrant labor: agriculture, construction, hospitality, food processing. The wage signal is pointing at a demographic shock, not a technological one.

This does not mean AI is irrelevant to the productivity numbers. It means we cannot yet decompose the signal. The immigration confound and the AI-adoption effect are running simultaneously through the same macro aggregates, and no econometric technique currently available can cleanly separate them with 18 months of data [Framework — Original]. Anyone who claims otherwise is selling a narrative, not presenting evidence.

Consider what a clean decomposition would require. You would need a control economy — similar GDP composition, similar AI-adoption rates, but without the immigration shock — to isolate the AI-specific productivity effect. Canada is the closest candidate, but its own immigration policy has shifted in parallel, complicating the comparison. The eurozone has lower AI adoption rates and different labor market structures. Japan has structural demographic decline that predates AI entirely. There is no clean counterfactual, which means there is no clean attribution.

The uncomfortable implication for this Institute’s own work is direct: any mechanism in the Theory of Recursive Displacement that relies on current macro data to establish AI-driven displacement — including the early activation of the Ratchet, the onset of Aggregate Demand Crisis conditions, and the measured pace of Structural Irrelevance — must carry an asterisk until the immigration confound is resolved. We are not exempting our own framework from this constraint.

The CEO Paradox: Broad Adoption, Shallow Integration

If AI were driving a structural decoupling of GDP from employment, we would expect to see it in the firms actually deploying it. The evidence here is paradoxical.

On one hand, adoption is widespread. The Federal Reserve reports that 79% of large firms are now using AI in some capacity [Measured]^8. Anthropic’s own research puts the adoption figure at 35.9% of workers, with — notably — positive wage effects for adopters [Measured]^4. This is not a marginal technology anymore. It is in the building.

On the other hand, impact is nearly invisible. Fortune’s February 2026 analysis of CEO surveys found that 90% of chief executives report zero measurable AI impact on their company’s financial performance [Measured]^3. Ninety percent. This is the Solow Paradox redux — “you can see AI everywhere except in the productivity statistics” — but updated for a world where the technology is deployed, operational, and still not moving the needle for the vast majority of firms.

Harvard Business School’s March 2026 research helps explain why [Measured]^5. Their analysis of AI’s labor market effects found two simultaneous and partially offsetting forces: a -17% effect on tasks susceptible to automation, and a +22% effect on tasks where AI serves as a complement to human work. The net effect is not displacement. It is restructuring — and the complementarity effect is currently winning.

This finding deserves far more weight than it has received in the displacement literature. If AI complementarity is generating a +22% productivity effect in tasks where humans work alongside AI systems, then the productivity boom we are observing may be driven not by labor replacement but by labor augmentation. The GDP-employment gap, under this reading, reflects not fewer humans doing the work but the same humans doing more work per hour with better tools.

The complementarity evidence does not invalidate the displacement thesis. Both effects are real and operating simultaneously. But any honest accounting of the current data must treat the +22% complementarity finding as a competing mechanism of equal empirical weight, not as a footnote to be acknowledged and moved past [Framework — Original].

There is a temporal dimension here that the cross-sectional data cannot capture. Complementarity and substitution may not be permanent parallel tracks — they may be sequential phases. The historical pattern with previous general-purpose technologies suggests an initial complementarity phase (where the technology augments existing workers) followed by a substitution phase (where the technology replaces entire workflows once the augmented workers have codified their tacit knowledge into the system). If this sequencing holds for AI, the current +22% complementarity effect may be the mechanism through which firms extract and formalize human expertise before automating it away entirely.

But that is a prediction, not a measurement. The current data shows complementarity winning. Any framework that dismisses the current data in favor of a predicted future reversal is engaging in theory-preservation, not evidence-based analysis. We flag the sequencing possibility. We do not treat it as established.

What AI Is Actually Doing: The Thin-Stratum Problem

The productivity miracle, to the extent it exists, is concentrated in a remarkably narrow slice of the economy. CNBC’s January analysis made the uncomfortable point directly: AI was not the biggest engine of U.S. GDP growth in 2025 [Measured]^7. Consumption was. The American consumer, still employed, still earning, still spending, drove the expansion. AI capex contributed to investment figures, but the downstream productivity effects remained localized.

The San Francisco Fed’s February economic letter identified what they called “the AI moment” — a period of genuine possibility that has not yet converted into broad-based productivity transformation [Measured]^9. Their TFP (total factor productivity) estimates showed approximately 5% growth in a single quarter — a striking number that invites the interpretation that AI is finally showing up in the macro data.

But single-quarter TFP estimates are notoriously volatile. The 1990s produced comparable TFP spikes during the dot-com investment boom, and the structural productivity gains from that era did not materialize until nearly a decade later, after the crash had cleared away the speculative froth. A single quarter of elevated TFP, against a backdrop of unprecedented immigration-driven labor supply contraction, is not sufficient evidence for a structural break [Estimated — source needed]. It is a data point that is consistent with multiple explanations, of which AI-driven decoupling is only one.

Goldman Sachs has offered the most cautious institutional estimate: the full labor market effects of AI will take approximately 10 years to materialize and will amount to a 6-7% GDP boost [Measured]^10. That is a meaningful number. It is also a number that implies gradual transformation, not the sudden structural rupture that the decoupling narrative requires. If Goldman is right, we are in the first inning of a long game — which is precisely where the Peterson Institute (PIIE) placed us in their March 2026 assessment [Measured]^13.

The academic debate mirrors this uncertainty. CEPR’s compilation of the Acemoglu-versus-Davidson disagreement [Measured]^11 captures the fundamental divide: Acemoglu’s more pessimistic models project displacement effects concentrated in routine cognitive tasks over the next decade, while Davidson’s ecosystem view emphasizes the demand-creation effects of AI investment, new firm formation, and complementary human capital. Both models are consistent with the current data. Neither has been falsified. That is the honest state of play.

The thin-stratum problem has a corollary that matters for policy. If the productivity miracle is concentrated in a small number of AI-native firms while the vast majority of the economy shows zero measurable impact, then policies designed for economy-wide decoupling — massive retraining programs, universal income floors, automation taxes — are calibrated to a scenario that does not yet exist. The appropriate policy posture for a thin-stratum effect is targeted: supporting displaced workers in the specific sectors where AI substitution is measurable, investing in complementarity infrastructure where augmentation is working, and building monitoring systems that can detect when the thin stratum begins to widen. The appropriate policy posture for economy-wide decoupling is structural: redesigning tax systems, rebuilding safety nets, renegotiating the social contract. Conflating the two is expensive and politically corrosive — it spends credibility on a narrative the data does not yet support.

This is where the Dissipation Veil becomes relevant, even though we do not tag it as a primary mechanism in this essay. The lag between AI capability and visible economic integration (MECH-013) means that structural displacement can accumulate beneath the threshold of macro-statistical detection. The 90% CEO no-impact finding might reflect genuine absence of impact — or it might reflect measurement systems that are not designed to capture the kind of impact AI produces. Firm-level financial metrics were built to measure the effects of capital investment in physical plant, not the effects of cognitive automation distributed across every knowledge worker’s desktop. If AI’s primary effect is to make existing workers more productive (the +22% complementarity finding) rather than to eliminate headcount, it would show up as margin improvement, not as headcount reduction — and most CEO surveys ask about the latter.

The Ratchet Question: Is Partial Decoupling Locking In?

The Theory of Recursive Displacement (MECH-001) predicts that once AI-driven substitution begins operating across institutions and sectors, it compounds — each displacement event creates conditions that accelerate the next. The Ratchet (MECH-014) formalizes the irreversibility component: sunk capex, organizational restructuring, and eliminated training pipelines make reversal more expensive than continuation. Together, these mechanisms predict that even modest initial displacement becomes self-reinforcing.

The current evidence is consistent with early-stage partial decoupling but is not yet diagnostic of the Ratchet [Framework — Original]. Here is why the distinction matters.

For the Ratchet to be operating in the macro data, we would need to observe not just a productivity-employment gap but a specific pattern: firms that have automated positions declining to rehire for those positions even when demand increases. We would need to see eliminated roles staying eliminated, training pipelines closing permanently, and organizational knowledge being irreversibly transferred to AI systems.

Some of this is happening. The Dallas Fed’s February 2026 analysis distinguishes between AI “aiding” existing workers (complementarity) and AI “replacing” worker functions (substitution), and finds evidence of both [Measured]^14. Firms that have deployed AI for document processing, customer service triage, and code generation are not rebuilding those teams when they grow — they are scaling the AI systems instead. That is a ratchet-like pattern.

But the base rate is thin. With 90% of CEOs reporting zero AI impact [Measured]^3, the firms exhibiting ratchet behavior are a small minority. The structural lock-in thesis requires that this minority becomes the majority — that the firms currently showing zero impact will, under competitive pressure, follow the leaders into irreversible automation. That may happen. Recursive Displacement predicts it will. But 18 months of data, confounded by the largest immigration policy shock in modern U.S. history, is not enough to confirm it.

The honest framing is this: early-stage partial decoupling is consistent with but not yet diagnostic of the Ratchet. The mechanism may be operating beneath the noise floor of the macro data. Or the macro data may be showing us something much simpler — a labor supply contraction masquerading as a technological transformation.

There is a test coming. If immigration policy normalizes — through legislative action, judicial intervention, or simple administrative drift — and net migration returns to historical levels of 1-1.5 million per year, the GDP-employment gap should narrow substantially if the immigration confound was the primary driver. If the gap persists or widens despite labor supply recovery, the AI-displacement interpretation gains significant ground and the Ratchet attribution can be upgraded from provisional to supported. This natural experiment may resolve within 12-18 months. The discipline required is to wait for the data rather than commit to a narrative in advance of it.

The Competence Insolvency Wall

Even granting the most aggressive AI-adoption timeline, complete decoupling of GDP from human labor by 2030 runs headlong into Competence Insolvency (MECH-012). This is the mechanism by which automation removes the economic incentives and practice loops that sustain human expertise, producing a system-level loss of the very capabilities needed to build, maintain, and govern the automated systems.

The irony is structural. The more successfully AI displaces human workers in a domain, the fewer humans develop expertise in that domain. The fewer experts exist, the harder it becomes to train the next generation of AI systems, audit their outputs, intervene when they fail, and redesign them when requirements change. Complete decoupling requires that AI systems become self-maintaining, self-auditing, and self-improving across every economically productive domain simultaneously. No serious AI researcher — not at Anthropic, not at Google DeepMind, not at OpenAI — claims this is achievable by 2030.

The Wage Signal Collapse (MECH-025) accelerates this dynamic. As AI compresses the wage premium for expertise, fewer people invest in becoming experts. The pipeline thins. The knowledge base erodes. And the firms that need human oversight for their AI systems find that the humans capable of providing it are increasingly scarce and expensive — which drives further automation of oversight functions, which further reduces the expert pipeline.

This is not a speculative concern. It is already visible in cybersecurity, where AI-generated code has expanded the attack surface faster than human security professionals can audit it. It is visible in financial modeling, where algorithmic trading systems have reduced demand for quantitative analysts while simultaneously increasing the systemic risk that requires their judgment. It is visible in healthcare, where AI diagnostic tools have begun displacing the clinical training rotations that produce the physicians who are supposed to oversee the AI.

The timeline matters here. Competence Insolvency is not a permanent barrier to decoupling — it is a rate limiter. If AI systems eventually develop the capacity for self-maintenance, self-auditing, and recursive self-improvement, the human capital constraint relaxes. Some researchers at the frontier labs believe this is achievable within a decade. Others consider it a multi-generational challenge. The honest assessment is that nobody knows, and that the uncertainty itself is the constraint. No firm can plan around a capability that may or may not emerge, on a timeline that may or may not be five years.

What firms can plan around is the current state of the labor market — and that state still requires humans. The Orchestration Class (MECH-018) describes the emergent human chokepoint that coordinates AI systems in production environments. These workers are not doing the cognitive work that AI has automated. They are doing the meta-cognitive work of deciding when to trust the AI, when to override it, when to escalate, and when to redesign the workflow entirely. That meta-cognitive layer is, for now, irreducibly human. And the supply of humans capable of filling it is constrained by the same Competence Insolvency that is thinning the expert pipeline in every other domain.

Complete decoupling by 2030 is not blocked by the current state of AI capability. It is blocked by the current state of human capital formation. You cannot automate your way out of a system that requires humans to build, certify, and repair the automation — not in five years.

The Aggregate Demand Question

If decoupling were proceeding at the pace the headline numbers suggest, we would expect to see early signs of the Aggregate Demand Crisis (MECH-010): output capacity expanding while labor income compresses, undermining the consumer demand that sustains the production circuit.

We are not seeing this yet. Consumer spending drove GDP growth in 2025 [Measured]^7. Wages for AI adopters are positive [Measured]^4. The labor market, while cooling, has not collapsed. The unemployment rate remains within normal cyclical bounds. The demand side of the economy is functioning.

This is important because the Aggregate Demand Crisis is the mechanism through which partial decoupling becomes economically destabilizing. If firms replace workers with AI, those workers lose income, that lost income reduces demand, reduced demand reduces the revenue that justifies the AI investment, and the system enters a contractionary spiral. The absence of this signal — so far — either means decoupling is not as advanced as the productivity numbers suggest, or it means the consumption-sustaining effects of remaining employment and accumulated wealth are still buffering the system.

The Post-Labor Economy framework (MECH-019) describes the endpoint: a configuration in which production no longer structurally depends on human labor, shifting distribution and agency away from wage work. Structural Irrelevance (MECH-021) describes the human experience of that transition: people remain socially present but economically nonessential. Neither condition obtains today. The question is whether the current data represents the beginning of a trajectory toward those conditions or a temporary aberration driven by policy-induced supply constraints.

The Automation Trap (MECH-011) adds a further complication. Even where AI is generating genuine productivity gains, those gains may be partially consumed by the complexity, overhead, and fragility that automation itself introduces. The 90% of firms reporting zero AI impact [Measured]^3 may not be failing to adopt AI — they may be experiencing the Automation Trap in real time, with efficiency gains in automated processes offset by integration costs, maintenance burdens, and the organizational overhead of managing human-AI hybrid workflows.

The demand-side picture introduces a second temporal question. If partial decoupling continues — productivity rising, job growth stagnating, GDP sustained by consumption from the still-employed majority — how long can the buffer hold? The aggregate demand crisis is a threshold phenomenon, not a linear process. Consumer spending can sustain GDP growth right up to the point where enough households experience income disruption that spending contracts — and that contraction, once it begins, feeds on itself through reduced business revenue, further layoffs, and further spending reduction. The system looks stable until it is not.

We are not at that threshold. The question is whether the current trajectory — 15,000 jobs per month against a growing economy — moves us toward it. If the GDP-employment gap is primarily immigration-driven, the answer is probably no: labor supply constraints ease when immigration policy shifts, and the gap closes without a demand-side crisis. If the gap is primarily AI-driven, the answer is more concerning: automated productivity gains that do not translate into wage income for new workers accumulate as corporate profit rather than household spending power, gradually eroding the consumption base that sustains the production circuit.

The honest answer, again, is that we cannot yet tell which dynamic is dominant. But the stakes of the distinction are enormous. One path leads to cyclical normalization. The other leads to structural transformation. And the macro data, as currently constituted, cannot distinguish between them.

Counter-Arguments and Limitations

Caveat 1: The immigration-versus-AI decomposition is not yet possible. This essay argues that immigration policy explains a substantial portion of the GDP-employment gap. But “substantial portion” is deliberately imprecise. We cannot yet quantify how much of the gap is immigration-driven versus AI-driven versus driven by other factors (fiscal policy, interest rates, sectoral composition shifts). The Ratchet attribution throughout this essay is provisional. If future econometric work demonstrates that the immigration confound explains less than 30% of the gap, the AI-displacement interpretation regains significant ground.

Caveat 2: “Structurally locked in” language is premature. The October 2025 version of this essay used language suggesting structural lock-in. This revision replaces that framing with “early-stage partial decoupling consistent with but not yet diagnostic of the Ratchet.” Eighteen months of data, confounded by an unprecedented immigration policy shock, does not support causal claims about irreversibility. The Ratchet mechanism may be operating, but the evidence is not yet sufficient to distinguish Ratchet dynamics from standard cyclical adjustment.

Caveat 3: The 90% CEO no-impact finding directly contradicts the lock-in framing. If nine out of ten firms report zero measurable AI impact, the claim that AI-driven displacement is structurally locked in across the economy requires an explanation for why 90% of the economy is locked into… nothing. The most honest interpretation is that AI’s labor market effects are concentrated in a small number of firms and sectors, and that generalization to the macro economy is premature. The counter-argument — that CEO self-reports understate actual impact because AI effects are distributed across functions and not yet captured by traditional financial metrics — has some merit but remains speculative.

Caveat 4: Complementarity as a competing mechanism, not a footnote. The Harvard Business School finding of +22% complementarity effects [Measured]^5 is not a minor qualification to the displacement thesis. It is a competing mechanism of comparable empirical weight. If AI is making existing workers 22% more productive in complementary tasks while displacing 17% of automatable tasks, the net labor market effect could be positive — which would mean the GDP-employment gap is driven not by displacement but by augmentation of a shrinking (immigration-constrained) workforce. This essay integrates complementarity as a first-order consideration, but the full implications for the displacement framework remain unresolved.

Caveat 5: The TFP argument rests on volatile data. The San Francisco Fed’s ~5% TFP estimate [Measured]^9 is a single-quarter figure. Total factor productivity is among the most volatile and revision-prone macroeconomic statistics. The 1990s produced comparable TFP prints during the dot-com investment boom; the structural productivity gains from information technology did not materialize until nearly a decade later. Using a single quarter of elevated TFP to diagnose a structural break in the GDP-employment relationship is methodologically aggressive. The 1990s precedent suggests that early TFP spikes often reflect measurement artifacts, capital deepening, or compositional shifts rather than genuine technological transformation of the production function.

Methods

This analysis was constructed through a structured adversarial pipeline. Evidence was gathered from primary data sources (Federal Reserve, Bureau of Labor Statistics, Dallas Fed, San Francisco Fed), institutional research (Goldman Sachs, Yale Budget Lab, Peterson Institute, Harvard Business School), and peer-reviewed academic work (CEPR, Anthropic Research). The October 2025 version of this essay relied heavily on forward projections and model-based estimates. This revision prioritizes realized data over projections, flags the immigration confound as a first-order analytical consideration, and downgrades confidence from 55-65% to 50-60% to reflect irreducible causal ambiguity.

The adversarial sparring process identified five binding caveats, all of which are integrated into the analysis rather than relegated to disclaimers. The most structurally important — that the immigration-versus-AI decomposition cannot yet be performed — is treated as the central analytical challenge rather than a limitation to be acknowledged and moved past.

Evidence tags follow the Institute’s standard taxonomy: [Measured] for claims backed by published data with numbered citations, [Estimated] for near-term extrapolations, [Projected] for speculative scenarios with stated assumptions, and [Framework — Original] for novel theoretical constructs from the Theory of Recursive Displacement.

Falsification Conditions

This essay is wrong if:

  1. The immigration confound is resolved in favor of AI. If econometric decomposition — using, for example, synthetic control methods comparing U.S. productivity trends against countries without comparable immigration policy shifts — demonstrates that AI accounts for more than 70% of the GDP-employment gap, the “mirage” framing collapses and the decoupling thesis is substantially upgraded.

  2. The CEO no-impact finding reverses within 12 months. If the 90% zero-impact figure drops below 50% in equivalent surveys by Q1 2027, this would indicate that AI’s effects were real but lagging measurement — and that the thin-stratum critique underestimated diffusion speed.

  3. Complementarity effects collapse. If the +22% complementarity finding fails to replicate, or if follow-up studies show that complementarity is a transitional phase that gives way to full substitution within 24 months, the competing-mechanism argument weakens substantially.

  4. TFP acceleration sustains across 4+ quarters. If total factor productivity growth remains above 3% for four consecutive quarters while controlling for immigration-driven compositional effects, the structural-break interpretation gains credibility that a single quarter cannot provide.

  5. Complete decoupling emerges in any G7 economy by 2028. If any major economy demonstrates GDP growth exceeding 3% for 8+ quarters with simultaneous employment contraction exceeding 2%, the “blocked by Competence Insolvency” claim is falsified. The mechanism registry prediction is that this cannot happen on a five-year timeline. If it does, the timeline was wrong.

Bottom Line

Confidence: 50-60%. This is lower than the October 2025 version (55-65%), and deliberately so.

The GDP-employment gap is real. Something is happening. But the most dramatic explanation — that AI is structurally decoupling economic output from human labor — cannot yet be distinguished from the most mundane one: that an 88% collapse in net migration mechanically suppressed labor supply, inflated productivity-per-hour statistics, and created a gap that looks technological but may be substantially demographic.

Partial decoupling is consistent with the data. AI is automating tasks, augmenting workers, and restructuring occupations. The Ratchet may be operating — firms that automate do not seem to reverse course. But “may be operating” is not “structurally locked in.” Eighteen months of confounded data does not support that conclusion.

Complete decoupling by 2030 is blocked by Competence Insolvency. The human capital pipeline required to build, maintain, and govern AI systems cannot be replaced by those systems on a five-year horizon. This is a hard constraint, not a prediction about the pace of innovation.

The most likely path forward is continued partial decoupling — GDP growing faster than employment, productivity rising, the labor share of income continuing its multi-decade decline — but driven by a combination of AI augmentation, immigration policy effects, and consumption dynamics rather than by the clean technological displacement that the headline numbers invite us to see. The mirage is not that nothing is changing. The mirage is that we know what is causing it.

We state our falsification conditions. If they are met, we will revise.

Where This Connects

This essay engages directly with the displacement framework developed across the Institute’s prior work:

  • The End of Labor — The “severed rung” hypothesis posits that AI breaks the historical link between productivity growth and employment growth. This essay argues the rung may be loosening, not severed — and that immigration confounds prevent confident diagnosis.

  • The Aggregate Demand Crisis — If decoupling were advanced, we would expect early demand-side signals. Their absence either means decoupling is less advanced than productivity numbers suggest or that consumption buffers remain intact. Both interpretations narrow the timeline for MECH-010 activation.

  • The Competence Insolvency — The hard constraint on complete decoupling by 2030. This essay treats MECH-012 as the binding ceiling on the most aggressive displacement scenarios.

  • The Post-Labor Lie — The structural endpoint that this essay argues has not yet been reached and cannot be reached by 2030, though the trajectory is not inconsistent with arrival on a longer timeline.

  • The Ratchet — This essay downgrades the Ratchet’s applicability to the current GDP-employment gap from “operating” to “consistent with but not yet diagnostic.” The mechanism definition holds; the attribution to current macro data is provisional.

  • Navigating the L.A.C. Economy — Published this week. The policy architecture proposed there assumes displacement materializes at projected scale. This essay’s lower confidence (50-60%) makes that assumption conditional rather than established.

Sources

  1. https://www.lpl.com/research/weekly-market-commentary/the-productivity-advantage-powering-economic-growth-in-2026.html — “The Productivity Advantage Powering Economic Growth in 2026”, LPL Research, 2026. 4.9% productivity growth, hours +0.5%. [verified]
  2. https://www.schwab.com/learn/story/whats-difference-jobs-vs-gdp-growth — “What’s the Difference? Jobs vs. GDP Growth”, Charles Schwab, 2026. 15K/month jobs vs 2.7% GDP. [verified]
  3. https://fortune.com/2026/02/17/ai-productivity-paradox-ceo-study-robert-solow-information-technology-age/ — “AI Productivity Paradox: CEO Study”, Fortune, 2026. 90% of CEOs report zero measurable AI impact. [verified]
  4. https://www.anthropic.com/research/labor-market-impacts — “Labor Market Impacts of AI”, Anthropic Research, 2026. 35.9% adoption rate, positive wage effects. [verified]
  5. https://hbr.org/2026/03/research-how-ai-is-changing-the-labor-market — “Research: How AI Is Changing the Labor Market”, Harvard Business Review, 2026. -17% automation effect, +22% complementarity effect. [verified]
  6. https://budgetlab.yale.edu/research/ai-productivity-boom-dont-count-your-productivity-data-chickens — “AI Productivity Boom: Don’t Count Your Productivity Data Chickens”, Yale Budget Lab, 2026. Immigration confound, wage gap analysis. [verified]
  7. https://www.cnbc.com/2026/01/26/ai-wasnt-the-biggest-engine-of-us-gdp-growth-in-2025.html — “AI Wasn’t the Biggest Engine of U.S. GDP Growth in 2025”, CNBC, 2026. Consumption drove GDP, not AI. [verified]
  8. https://www.federalreserve.gov/newsevents/speech/barr20260217a.htm — Federal Reserve speech, 2026. 79% of large firms using AI. [verified]
  9. https://www.frbsf.org/research-and-insights/publications/economic-letter/2026/02/ai-moment-possibilities-productivity-policy/ — “The AI Moment: Possibilities, Productivity, Policy”, San Francisco Fed, 2026. TFP approximately 5%. [verified]
  10. https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-us-labor-market — “How Will AI Affect the U.S. Labor Market”, Goldman Sachs, 2026. 10-year timeline, 6-7% GDP boost. [verified]
  11. https://cepr.org/voxeu/columns/ai-investment-gdp-growth-ecosystem-view — “AI Investment, GDP Growth: An Ecosystem View”, CEPR/VoxEU, 2026. Acemoglu vs Davidson debate. [verified]
  12. https://www.dallasfed.org/research/economics/2025/0624 — Dallas Fed Research, 2025. Net migration: 2.7M to 1.3M to 321K. [verified]
  13. https://www.piie.com/blogs/realtime-economics/2026/research-ai-and-labor-market-still-first-inning — “AI and the Labor Market: Still First Inning”, Peterson Institute, 2026. Early-stage assessment. [verified]
  14. https://www.dallasfed.org/research/economics/2026/0224 — Dallas Fed Research, 2026. AI aiding vs. replacing worker functions. [verified]
Back to all essays