VAN R. HOISINGTON AND LACY H. HUNT: AI is unlikely to lift aggregate productivity much above its long-run 2% pace, given a large government sector, a services-heavy economy and lost scale

The Gist

Even if AI helps some industries, most American jobs are in government, health care and services, where machines raise output slowly, so overall productivity probably stays around 2 percent. This is a steelman: Van R. Hoisington and Lacy H. Hunt's arguments are reconstructed in their strongest fair form, not endorsed.

Conclusion

Aggregate productivity growth is likely to stay near its long-run 2 percent pace despite AI, because a large government sector, a services-heavy job mix and lost economies of scale limit how far gains spread.

Premises

1. Long-run U.S. labor productivity growth has been about 2.0 to 2.1 percent a year (1929 to 2025 and 1949 to 2025).
2. Government outlays are about 35.3 percent of GDP, and private service-providing jobs outnumber goods-producing jobs about 5.3 to 1 (September 2026).
3. Most employment growth is in labor-intensive sectors such as health care, social assistance and hospitality, where productivity gains are harder to achieve.
4. Deglobalization (trade barriers, fragmentation and duplicated supply chains) reduces economies of scale and raises capital needs, and more resources go to political or policy objectives.
5. Rising federal interest costs also weigh against faster aggregate productivity growth.

Assumptions

• Productivity gains in a few sectors do not quickly spread to the whole economy.
• Financial interest: Hoisington Investment Management manages U.S. Treasury portfolios and for decades was known for holding very long-duration Treasuries on a falling-inflation, falling-rates view. Press reports on its 2026 letters say the firm cut effective duration from about 20.9 years at the end of September 2025 to under one year by June 30, 2026, so this letter's forecast of higher long-term yields is consistent with the firm's current short-duration positioning rather than with its earlier long-duration stance. Sources: "Decades long bull turns bearish. Cites structural deficits…

Analysis

Argument type: Inductive.

Potential Fallacies

• Unwarranted extrapolation from a historical average (P1 → Conclusion) — A 76-96 year mean is used as the forecast anchor, which assumes the process generating productivity is stable. The average blends very different regimes (roughly 2.8% in 1948-73, about 1.5% in 1973-95, about 2.5-3% in 1995-2004, about 1.2-1.5% in 2005-19). General-purpose technologies such as electricity and IT shifted trend growth after long lags. Whether AI is such a break is the very question at issue, so the base rate informs but does not settle it.
• Composition / static-structure inference (P2, P3 → Conclusion) — Headcount shares of services and health care, and the government share of spending, are treated as fixed constraints on aggregate gains. Aggregate productivity depends on value-added-weighted gains in AI-exposed activities, not on job counts. AI may also change the tasks within health care, government administration and other services, so a services-heavy economy may be a channel for AI gains rather than a barrier.
• Non-diagnostic evidence (explaining the past, not discriminating the future) (P2, P3) — The high services share and government size already held throughout the period when productivity averaged 2%, and through the 1995-2004 acceleration. These facts are therefore about equally likely whether or not AI raises productivity, so they add little evidence about AI's effect.
• Unsupported causal assertion (P4, P5) — Deglobalization and federal interest costs are said to depress productivity growth with no mechanism, magnitude or timing. Direction is plausible but size is not shown. The two factors could also be partly offset by automation-driven reshoring, capital deepening, or growth easing the fiscal burden. They show other forces may lower growth, not that AI's contribution is small.
• Measurement artifact treated as a real constraint (P2) — Government output is largely measured by inputs, so measured productivity there is near zero by construction. Outlays also include transfers and interest, which are not produced output. A large outlay share therefore overstates the relevant drag and conflates a statistical convention with an economic limit.
• Selective evidence / one-sided headwinds (P2-P5 overall) — Only factors that depress productivity are assembled. Countervailing evidence, such as task-level AI gains in coding, customer support and documentation, the AI investment boom, and labor scarcity pushing automation, is not weighed. The stacking of items can create an impression of overwhelming support even though they vary in relevance and overlap (P2 and P3 describe much the same fact).

Counterarguments

These challenges are not ranked.

• Premise 2 and Premise 3 (High impact) — AI is a general-purpose technology aimed at cognitive work, which dominates much of services, health administration and government (documentation, billing, triage, case processing, legal and financial analysis, software). A services-heavy economy is therefore a large target for AI, not a shield against it. The 1995-2004 productivity pickup also occurred in an economy already service-dominant.
• Premise 1 (High impact) — The 2% average conceals regime shifts. Electricity and IT showed multi-decade lags followed by sharp acceleration, and the Solow paradox was followed by the late-1990s boom. A long-run mean is therefore weak evidence about whether a new general-purpose technology will break the pattern.
• Conclusion (High impact) — Aggregate productivity is a weighted sum of sector growth, so large gains in even a 30-40% slice of the economy would lift the aggregate noticeably. The argument offers no arithmetic showing the AI-exposed share or the diffusion speed is too small.
• Premise 2 (Medium impact) — Government outlays (35.3% of GDP) mostly consist of transfers and interest. Government value added is only about 11-12% of GDP, and government output is measured by inputs. The outlay share therefore overstates both the size and the real productivity drag of the government sector.
• Premise 4 and Premise 5 (Medium impact) — Deglobalization and higher capital needs can induce automation and capital deepening that raise labor productivity, and AI-driven growth can ease fiscal burdens. The sign and size of the net effect on productivity are unexamined.
• Assumption 1 (A1) (Medium impact) — Software-based AI can be copied across firms at near-zero marginal cost, and competitive pressure forces laggards to adopt once leaders show gains. This suggests diffusion may be faster than for capital-heavy earlier technologies.
• Conclusion (Medium impact) — Baumol dynamics cut both ways. Rising wages in stagnant labor-intensive sectors create the strongest incentive to automate exactly those sectors, and aging labor markets reinforce it.
• Conclusion (Low impact) — Measured productivity may fail to capture AI-driven quality gains and free digital services, so the aggregate statistics may be an unreliable test of AI's effect in either direction.

Suggested Improvements

• Direct evidence on AI — Add evidence on AI adoption rates, task-level and firm-level productivity studies, and sector-level AI exposure, and bridge from those micro results to aggregate effects. The argument is about what dilutes AI's gains but never examines how large those gains could be, which leaves the central question untested.
• Quantify the sector arithmetic — Provide a shift-share or growth-accounting decomposition using value-added weights for AI-exposed sectors (information, finance, professional services, administration) versus low-exposure sectors, and state what within-sector gains would be needed to move the aggregate. Replaces the job-count ratio with a measure that actually bears on aggregate productivity, and shows whether the 'near 2%' claim follows.
• Use the right government measure — Use government value added or employment share and separate transfers and interest from government production, or restrict the analysis to business-sector productivity. Outlays overstate the relevant drag and invite an easy technical objection.
• Specify horizon and thresholds — State a time frame and define 'near' and 'much above' 2% (for example, a band of 1.6-2.4% over the next decade), ideally with scenario probabilities. Makes the forecast falsifiable and distinguishes near-term diffusion lags from long-run structural effects.
• Engage the strongest opposing case — Address the general-purpose-technology and J-curve arguments directly, explaining why AI would deviate from the electricity and IT pattern, and engage the 1995-2004 episode in a services-heavy economy. The historical base rate cuts both ways, so the argument needs a reason to prefer the stationary interpretation.
• Support P4 and P5 with mechanisms and magnitudes — Cite empirical estimates of trade fragmentation's effect on TFP and of interest burdens on investment, and address the offsetting channels (reshoring automation, capital deepening). Both premises currently establish direction only. Their contribution to holding growth near 2% is not demonstrated.
• Transparency on interests — State the firm's duration positioning within the letter and cross-check the thesis against independent forecasts. The conclusion fits the firm's current positioning, so disclosure and external validation increase reader trust.

Scenario Tests

• AI automates documentation, billing, scheduling and diagnostics in health care and administrative work in government, shifting these 'labor-intensive' sectors toward faster productivity growth. (Challenges) — P3 becomes a channel for gains rather than a barrier. The aggregate could rise meaningfully above 2%, and the conclusion fails.
• AI raises productivity mainly in tradable and goods-related sectors, and Baumol cost disease shifts employment toward stagnant services. (Supports) — Mix effects dampen the aggregate figure even when real gains are large, which is consistent with the conclusion, though through a mechanism the argument does not itself articulate.
• Diffusion follows the electricity and IT pattern: a decade or more of weak measured gains, then a sharp acceleration once complementary investment and reorganization mature. (Challenges) — A near-term 'near 2%' reading would be right but would not support the claim as a durable structural limit. The missing time horizon makes the conclusion ambiguous.
• AI gains show up largely as quality improvements and consumer surplus that official statistics do not capture, leaving measured productivity near 2%. (Supports) — The measured-aggregate prediction holds, but for measurement reasons rather than because the structural headwinds blocked gains.
• Persistent AI-linked borrowing pushes yields higher and crowds out private investment, while tariffs and fragmentation persist. (Supports) — P4 and P5 gain force and could dampen capital deepening, though they would not explain why AI's own contribution is small.
• Reshoring is accompanied by heavy investment in robotics and automation, and AI-driven capex lifts capital per worker. (Challenges) — The deglobalization and higher-capital-needs premise would then raise labor productivity, reversing the sign the argument assigns to it.

Coherence & Relevance

The argument is a convergent inductive case. It combines a base-rate premise with several independent-looking structural headwinds, joined by the assumption that sector gains do not spread quickly (A1). It is coherent in direction and hedged appropriately in form, but it is mostly about what constrains diffusion rather than about what AI can do. No premise addresses AI's mechanisms, adoption or magnitude. The headwinds establish direction but not size, and P2 and P3 largely double-count the same fact, so the evidence is thinner than five premises suggest. A1 carries much of the argument's weight. The conclusion is a reasonable conservative prior, plausible mainly because of historical diffusion lags, but it is not firmly established by the premises. The disclosed financial interest (A2) is consistent with the thesis and suggests weighing the unsupported judgment calls cautiously, without undermining the verifiable data.

• Long-run U.S. labor productivity growth has been about 2.0 to 2.1 percent a year (1929 to 2025 and 1949 to 2025). (Strong) — Provides the base rate, but does not show that the generating process is stable or that AI cannot cause a regime shift.
• Government outlays are about 35.3 percent of GDP, and private service-providing jobs outnumber goods-producing jobs about 5.3 to 1 (September 2026). (Moderate) — Describes current structure, not AI exposure. The outlay figure includes transfers, and the job ratio is a headcount measure. Both structural facts already held during the 2% history and the 1995-2004 acceleration.
• Most employment growth is in labor-intensive sectors such as health care, social assistance and hospitality, where productivity gains are harder to achieve. (Moderate) — Largely overlaps with P2 in what it counts as evidence. It does not establish that AI cannot raise productivity within these sectors or that they carry enough weight to cap the aggregate.
• Deglobalization (trade barriers, fragmentation and duplicated supply chains) reduces economies of scale and raises capital needs, and more resources go to political or policy objectives. (Weak) — Concerns a general drag on productivity, not AI's diffusion or contribution. No magnitude is given, and the capital-needs channel could raise labor productivity.
• Rising federal interest costs also weigh against faster aggregate productivity growth. (Weak) — No mechanism links interest outlays to productivity. The premise is also not named in the conclusion's list of limiting factors, a minor mismatch between stated premises and stated conclusion.

View this argument on LogicFirst.ai