AI creates more jobs than it destroys - the Jevons Paradox

For the last two years, the consensus AI story has been simple and scary: large language models will gut white‑collar work, and software engineers will be first in line. The latest research from Citadel Securities suggests the opposite is happening - at least for software engineers.

In a recent global macro note, Citadel highlights a striking fact: job postings for software engineers are up 11% year over year, even as overall job postings across the economy are flat or falling. The data comes from Indeed and is being flagged not by a Silicon Valley booster, but by Ken Griffin’s roughly $65 billion market‑making empire.

If AI is already destroying knowledge‑work jobs, the labor market has a strange way of showing it.

What the Citadel data actually says

Citadel’s “2026 Global Intelligence Crisis” report sets the stage with a macro snapshot: U.S. unemployment sitting at 4.28%, AI capital expenditure running at about 2% of GDP (roughly 650 billion dollars), AI‑linked commodities up 65% since early 2023, and around 2,800 data centers planned in the U.S. alone. This is what a major investment boom looks like, not a labor market collapse.

Within that backdrop, their Indeed chart shows two lines: one for overall job postings and one for software engineer postings.

·       Overall postings are flat to down.

·       Software engineer postings are rising sharply, up 11% year on year.

Citadel’s point is blunt: in spite of all the “imminent disintermediation” rhetoric, the real‑time labor data does not display an inflection in AI‑driven job destruction today. If anything, the tight labor market and rising demand for technical roles sit awkwardly alongside the more sensational AI unemployment narratives.

Why Jevons Paradox matters for AI strategy

To understand what’s going on, you have to go back to a 19th‑century economist, William Stanley Jevons. Jevons observed that making coal use more efficient in steam engines didn’t reduce coal consumption; it made coal so economically attractive that total consumption soared.

This Jevons Paradox shows up whenever a key input becomes dramatically cheaper:

·       The unit cost drops.

·       New use cases become viable.

·       Total demand for the underlying capability explodes.

AI is doing this to software. Coding has become faster and cheaper thanks to tools like code assistants and model‑driven scaffolding. That doesn’t lead companies to build less software. It leads them to ask, “What else can we automate, optimize, or create that was previously too expensive?”

Citadel frames AI as a productivity shock: a positive supply shock that lowers marginal costs, expands potential output, and, in the medium term, raises real incomes. Historically, technologies that deliver this kind of shock - electrification, internal combustion, computing - shift the structure of demand and create new categories of work rather than simply destroying it.

For executives, the implication is clear: if you treat AI primarily as a cost‑cutting tool, you’ll miss the larger opportunity. The real strategic question is how to redeploy the productivity dividend into new products, markets, and capabilities.

The hidden split in “tech jobs”

The Citadel framing also helps untangle an important labor market nuance that many dashboards blur: not all technical roles are created equal.

Across U.S. data, you see a sharp divergence between two broad categories:

·       “Computer programmers” - narrowly defined roles focused on translating given specifications into code.

·       “Software developers / software engineers” - broader roles that combine design, architecture, integration, and delivery.

Since 2023, employment in “computer programmer” roles has fallen by roughly 27%, taking it to the lowest level since around 1980. At the same time, “software developer” employment is only marginally down in the near term (about 0.3%) and is projected by the Bureau of Labor Statistics to grow about 18% across this decade. (These figures are drawn from BLS categories and cited by multiple analyses; the Citadel piece emphasizes the broader point that the displacement risk is over‑stated in aggregate labor data.)

This split is exactly what you would expect in a Jevons world:

·       The narrow “code monkey” job - manual translation of logic into syntax - is the most automatable task.

·       The higher‑leverage role - defining problems, architecting systems, integrating AI, managing complexity - becomes more valuable as the unit cost of code falls.

Citadel explicitly argues that AI looks more like a complement than a substitute to skilled labor in many domains, echoing what happened with earlier office technologies. When spreadsheets and word processors arrived, the fear was that they would eliminate office jobs; instead, they redefined them and expanded what office workers could do.

For leaders, this isn’t a semantic distinction. It’s a roadmap for workforce strategy: the risk is concentrated in low‑autonomy, narrowly scoped work, not in end‑to‑end product builders.

AI adoption is powerful, but it’s not instant

Another important contribution from Citadel’s work is a reminder that “recursive technology” does not mean “recursive adoption.” Because AI systems can improve themselves and accelerate their own development, it’s tempting to assume the economic deployment will follow a smooth exponential curve.

The historical record suggests otherwise. Citadel points to the classic S‑curve of technological diffusion:

·       Early adoption is slow and expensive.

·       Then you hit a steep middle phase as costs fall and infrastructure matures.

·       Eventually, adoption saturates and growth slows as the marginal adopter becomes less productive or profitable.

Using data from the St. Louis Fed’s Real Time Population Survey, Citadel notes that the share of working‑age adults using generative AI for work, especially daily use, has been unexpectedly stable so far. There is no evidence of a runaway curve in work‑related AI intensity yet, which undercuts the idea of an imminent labor shock.

This has three strategic implications:

·       AI will be a multi‑year integration challenge, not a one‑quarter flip.

·       Organizational change, regulation, and compute constraints will act as brakes.

·       Firms that learn to integrate AI systematically - process by process - will pull away from those that wait for a mythical “steady state” to arrive.

For boards and executives, the key risk is not missing a one‑time automation window; it’s under‑investing in the messy, incremental work of adoption while competitors quietly build AI‑native processes and products.

What this means for business leaders

Put together, the Citadel analysis and the surrounding labor data support a different AI narrative from the popular doom loop.

Instead of “AI kills jobs, especially in software,” the more accurate macro story today looks like:

·       AI is driving one of the largest capex cycles in recent history - data centers, chips, and infrastructure - while unemployment remains low.

·       Software engineer demand is rising even as overall postings stagnate, suggesting that software is becoming more central to competitive advantage, not less.

·       The roles under real pressure are narrowly scoped, low‑autonomy coding jobs, not high‑leverage engineering and product roles.

·       Adoption is powerful but bounded by S‑curve dynamics, organizational friction, and physical constraints like compute and energy.

For general business leaders, the strategic questions shift:

·       From “How many people can we replace with AI?” to “What new products and capabilities become viable now that software is cheaper to build?”

·       From “Will AI destroy our workforce?” to “How do we retrain and reorganize our workforce so that AI amplifies our best people and eliminates low‑value tasks?”

·       From “Can we time the AI peak?” to “How fast can we build the internal muscles to deploy AI across functions over several years?”

If Citadel is right, the real AI risk for most companies in the late 2020s is not mass unemployment; it is falling behind in a world where your competitors have turned cheaper intelligence into more software, more experiments, and more ways to serve your customers.