AI Doom Overstated: Citadel Says Historical Precedent Suggests Slow Adoption, Economic Resilience

Following the AI doomsday reports from Anthropic CEO Dario Amodei (a turbulent "rite of passage" for humanity leading to rapid job disruption and a major shock to the economy), and Citrini Research (a hypothetical scenario where rapid AI agent advances create a negative feedback loop, leading to >10% US unemployment, a ~38% S&P 500 drop, and economic contraction by mid-2028), Citadel Securities macro maven, Frank Flight, takes a different tack, pushing back on this dystopian future view.

Specifically, Flight argues that recursive AI progress does not equate to recursive economic adoption; diffusion follows historical S-curves, constrained by costs, infrastructure, regulation, energy limits, and organizational friction.

Furthermore, he points out that compute-intensive automation faces rising marginal costs, creating natural barriers to substitution, concluding that AI likely offsets structural headwinds, sustaining ~2% trend growth without rendering labor obsolete: "The macroeconomy remains governed by substitution elasticities, institutional response, and the persistent elasticity of human wants."

The 2026 Global Intelligence Crisis

The year is 2026. The unemployment rate just printed 4.28%, AI capex is 2% of GDP (650bn), AI adjacent commodities are up 65% since Jan-23 and approximately 2,800 data centers are planned for construction in the US*. In spite of the current displacement narrative – job postings for software engineers are rising rapidly, up 11% YoY.

Despite the macroeconomic community struggling to forecast 2-month-forward payroll growth with any reliable accuracy, the forward path of labor destruction can apparently be inferred with significant certainty from a hypothetical scenario posted on Substack: The 2028 Global Intelligence Crisis.

We wrote last week that we see the near-term dynamics around the AI capex story as inflationary, but given markets are focused on the forward narrative, we outline a more constructive take on the end state below. Before that, however, it’s worth reflecting that the imminent disintermediation narrative rests on the speed of diffusion.

Job Postings For Software Engineers Are Rapidly Rising
Indeed Job Postings: Software Engineers + Overall Postings, Daily and 21dma

Source: Citadel Securities, Indeed. Figures are for illustrative purposes only. Past performance figures do not guarantee future results.

What Does the Data Actually Say on AI Diffusion Speed?

The St Louis Fed has data on AI adoption from the Real Time Population Survey. The first order presentation of AI adoption is generally a binary question: Do you use AI? The more important question insofar as it relates to the AI displacement narrative is: how intensely is AI being used for work? We can tease out the answer from a subset of the St Louis Fed data that buckets by frequency of AI use. We would posit that if AI represents imminent displacement risk, the real time population data would show an inflection upwards in the daily use of AI for work. The data seems unexpectedly stable and presents little evidence of any imminent displacement risk (solid lines at the bottom of the chart).

AI Adoption Trends Do Not Look Non-Linear
Share of Working Age Adults Using Generative AI, Real Time Population Survey

Source: Citadel Securities, St Louis Fed, Real Time Population Survey. Figures are for illustrative purposes only. Past performance figures do not guarantee future results.

Recursive Technology ≠ Recursive Adoption

The current debate around artificial intelligence conflates the recursive potential of the technology with expectations of recursive economic deployment. In other words, because AI systems can improve themselves or accelerate their own capabilities, commentators are extrapolating a future in which automation and productivity compound indefinitely at exponential rates. Technological diffusion has historically followed an S-curve. Early adoption is slow and expensive. Growth accelerates as costs fall, and complementary infrastructure develops. Eventually, saturation sets in, and the marginal adopter is less productive or less profitable which causes growth to decelerate.

Despite this – markets often extrapolate the acceleration phase linearly but history implies pace of adoption plateaus as organizational integration is costly, regulation emerges and diminishing marginal returns exist in economic deployment. The risk of displacement declines with a slower pace of adoption.

Adoption Rate of Generative AI at Work and Home versus the Rate for Other Technologies
Share of working age adults using AI at work for three technologies: generative AI, PCs and the internet

Source: Citadel Securities, BLS, St Louis Fed, Real Time Population Survey International Telecommunication Union. Figures are for illustrative purposes only. Past performance figures do not guarantee future results.

Furthermore, it is well acknowledged that training and inference requires significant semiconductor capacity, data centers, and energy. Displacing white collar work would require orders of magnitude more compute intensity than the current level utilization. If automation expands rapidly, demand for compute definitionally rises, pushing up its marginal cost. If the marginal cost of compute rises above the marginal cost of human labor for certain tasks, substitution will not occur, creating a natural economic boundary. This dynamic contrasts sharply with narratives assuming frictionless replication of intelligence. Even if algorithms improve recursively, economic deployment remains bounded by physical capital, energy availability, regulatory approvals, and organizational change. Recursive capability does not imply recursive adoption.

Productivity Shocks Are Supply Shocks

At its core, AI-driven automation is a productivity shock. Productivity shocks are positive supply shocks: they lower marginal costs, expand potential output, and increase real income. They are in isolation disinflationary and growth-enhancing in the medium term. Historically, every major technological advance: steam power, electrification, the internal combustion engine, computing, has followed this pattern. The counterargument suggests that AI differs because it displaces labor income directly, thereby suppressing aggregate demand. If firms produce more at lower cost, prices fall or margins expand (or both). Lower prices increase real purchasing power, which generally increases consumption. Higher margins increase retained earnings and investment capacity. If output rises and real GDP increases then by national income accounting identity something must be rising on the demand side: Consumption, investment, government spending, or net exports must be increasing (more here). A scenario in which productivity surges but aggregate demand collapses while measured output rises violates accounting identities. For AI to generate a sustained macro contraction one must assume that labor income falls and no compensating rise occurs in investment, fiscal transfers, or external demand. The surge in new business formation is an interesting point of reference here.

New Business Formation is Rapidly Expanding
New Business Applications, US Census Bureau

Source: Citadel Securities, US Census Bureau. Figures are for illustrative purposes only. Past performance figures do not guarantee future results

Substitution Elasticity Constraint

The critical variable in AI displacement is the elasticity of substitution between AI capital and labor. If that elasticity is extremely high – i.e. firms can substitute nearly all human labor with automated systems at relatively stable cost – then labor’s share of income could collapse. In such a world, capital income rises dramatically while wage income contracts. But even here, aggregate demand does not automatically implode. Capital income has a lower marginal propensity to consume than wage income, but it does not have zero spending velocity. Profits can be reinvested, distributed, taxed, or spent. For demand to fall structurally, redistribution mechanisms would need to fail persistently, and investment opportunities would need to dry up simultaneously. Democratic nations facing such displacement risk would generally be expected to err towards in regulatory and fiscal policy shifts that offset the worst-case outcomes, further limiting substitution elasticity. Moreover – there is little evidence of AI disruption in labor market data as of today. In fact, the forward-looking components of our labor market tracking have improved and AI data center construction appears to be driving a pick-up in construction hiring.

US Labor Market Tracking Continues To Point to Improvement
6m Z-scores of selected labor market indicators, Cital Securities Labor Market Tracker, NFP Private Payrolls

Source: Citadel Securities, Bloomberg, BLS, ADP, Revelio, Indeed. Figures are for illustrative purposes only. Past performance figures do not guarantee future results

The economy contains a vast array of tasks: physical, relational, regulatory, supervisory – that are costly to automate. Even cognitive automation faces coordination frictions, liability constraints, and trust barriers. It seems more likely that AI will be a complement rather than a substitute for labor is many areas. Historically, technological revolutions have altered task composition rather than eliminated labor as an input. To produce a negative demand shock large enough to overwhelm output expansion, one must assume near-total automation of economically relevant labor combined with extremely weak redistributive responses. To frame this debate correctly one can simply ask, was the advent of Microsoft office a complement or substitute for office workers? Ex-ante the concern skewed towards substitution, ex post it appears a clear complement.

Data Centre Construction is Boosting Construction Hiring
United States, Employment, Payroll, Construction, SA

Source: Bloomberg, Citadel Securities, Feb 2026. Figures are for illustrative purposes only. Past performance figures do not guarantee future results.

The 15 Hour Work Week

In 1930, John Maynard Keynes wrote “Economic Possibilities for our Grandchildren,” predicting that productivity growth would be so powerful that by the early twenty-first century the workweek would fall to fifteen hours. He was directionally correct about productivity growth, but profoundly wrong about labor market implications. Rather than working dramatically less, societies consumed dramatically more. Why? Because rising productivity lowered costs and expanded the consumption frontier. Preferences shifted toward higher quality goods, new services, and previously unimaginable forms of expenditure. Leisure increased modestly, but material aspiration expanded far more. History suggests productivity gains do not automatically translate into labor withdrawal or demand collapse as they alter the composition of demand, expand real incomes and generate new industries. Keynes underestimated the elasticity of human wants.

Conclusion

For AI to produce a sustained negative demand shock, the economy must see a material acceleration in adoption, experience near-total labor substitution, no fiscal response, negligible investment absorption, and unconstrained scaling of compute. It is also worth recalling that over the past century, successive waves of technological change have not produced runaway exponential growth, nor have they rendered labor obsolete. Instead, they have been just sufficient to keep long-term trend growth in advanced economies near 2%. Today’s secular forces of ageing populations, climate change and deglobalization exert downward pressure on potential growth and productivity, perhaps AI is just enough to offset these headwinds. The macroeconomy remains governed by substitution elasticities, institutional response, and the persistent elasticity of human wants.