When Code Gets Cheaper, More Gets Built: AI, Jevons Paradox, and the Shape of Intellectual Work
AI is changing how intellectual work is practiced, and software is a clear case study in that transition. The shape of the change is visible, and fits a pattern identified with the steam engine.
The Mechanism
In 1865, William Stanley Jevons observed something counterintuitive about the steam engine. As engines became more fuel-efficient, total coal consumption rose rather than fell. Cheaper energy made previously unviable industries viable, and demand expanded to fill the space that efficiency created. Coal use went up.
This pattern — efficiency gains increasing rather than reducing total resource consumption — is now called Jevons Paradox. It’s since been documented across fuel efficiency, agricultural water use, LED lighting, and computing. When the cost of a core input falls dramatically, demand for the output it enables expands faster than the efficiency gain. Absolute consumption rises.
Efficiency ↑ ⟹ Cost of Production ↓ ⟹ Demand for Output ↑↑ ⟹ Resource Use ↑
This is what’s happening in software, and in intellectual work more broadly.
AI dramatically reduces the cost of generating code, writing first drafts, synthesizing research, and producing analysis. The displacement thesis assumes demand for that output is fixed — that cheaper production simply reduces the headcount required to meet it. Jevons Paradox predicts the opposite: lower production cost unlocks demand that previously could not be served. The backlog of software the world wants but could not afford to build, the analyses that were too expensive to run, the documentation that was always deferred — these expand to consume the surplus capacity.
There is one property of AI output that has no direct analog in prior technology transitions. Steam and electrical failures are visible — the boiler explodes, the light goes out. AI produces errors that are fluent, confident, and structurally plausible. They don’t look like errors. This changes what verification requires, and it is the reason verification has become the binding constraint rather than simply a quality preference.
What History Shows
The steam and electrical transitions show what the adjustment period looks like, and what distinguishes the organizations that capture the efficiency gain.
The steam engine increased output volume. That was the first-order effect. The durable gains required precision tolerances, standardized parts, and the emergence of mechanical engineering as a formal discipline to manage the complexity of high-pressure systems. Higher output forced stronger engineering.
The electrical transition makes this clearer. Early factories replaced one large steam motor with one large electric motor — roughly unchanged output, different power source. Productivity gains only materialized when engineers recognized that electricity allowed something structurally different: a small motor on every individual machine. This required re-architecting the factory floor entirely. New configurations became possible that steam could never reach — new equipment, new layouts, new ways of organizing work that hadn’t existed before. Economists studying this period documented a lag of roughly a decade between electrical adoption and measurable productivity gains, precisely because the re-architecture took time. (Paul David documented the identical pattern in computing — his 1990 analysis of the dynamo and computer is the canonical source for this lag. PDF)
The implication for AI is direct. AI injected into existing workflows without structural change is the motor swap. More output of the same kind. The gains documented in practice come from something different.
Two Paths in Practice
Organizations using AI are bifurcating along a clear line, and the outcomes are measurable.
Generation-First uses AI to produce output faster, volume rises, and the architectural thinking that should precede generation is not strengthened to match. The failure mode here is structural, not incidental: uncertainty in AI outputs compounds when steps are chained — each output feeding the next — rather than adding linearly. A pipeline that looks reliable at each individual stage can produce genuinely untrustworthy results at the end. Code churn — code rewritten or deleted within weeks of creation — is rising in organizations following this path (Amazon’s engineering all-hands in March 2026 addressed a spate of outages the company attributed to AI-assisted changes deployed without established safeguards [archive]). The velocity is real. Technical debt is being produced at the same speed.
Spec-First reverses the sequence. Human judgment defines the requirements and the test suite first. AI generates candidate implementations. Automated verification eliminates the candidates that fail. A human reviews and approves what remains. The focus shifts from production to selection. This pattern appears across software and is beginning to appear in other knowledge-intensive fields — legal drafting, financial analysis, research synthesis — wherever the volume of AI-generated output has outrun the capacity to verify it informally.
Both paths are operating now, in the same industries, sometimes in the same organizations. The difference in outcome is substantial.
The Roles Emerging
Software job titles are lagging the underlying shift, but the functions separating out can be described plainly, and they are not specific to software.
Orchestration: directing AI systems that generate output — deciding which tool handles which component, how components connect, and what the overall system is trying to accomplish. Less about production, more about architecture and sequencing.
Context engineering: managing what the AI is given — the specific rules, constraints, domain knowledge, and organizational logic that determine whether AI output is useful or generic. This function did not previously exist as a distinct role.
Verification: confirming that what was generated is correct, coherent, and trustworthy. As generation speed increases, this becomes the constraint. It is closer to structural inspection than writing, and its importance scales with the volume of AI output being produced.
These three functions were previously bundled. They are separating, in software and across intellectual work wherever AI adoption is dense enough to surface the distinction.
Practitioners are also discovering capabilities that have no precedent in prior workflows — the equivalent of new floor plans rather than rearranged motors. One finding from software practice: articulating work to an AI while explaining what you built and why surfaces errors through the act of explanation itself, before any feedback is given. The forcing function is the explanation, not the response. This is a new kind of cognitive step that the old workflow didn’t contain, and it transfers across any intellectual work that involves building and verifying something complex.
One structural concern worth noting: the entry-level roles that have traditionally served as on-ramps to senior expertise are in directional decline, because the syntax-level and first-draft work that juniors used to do is now handled by generation tools. How professions develop the next generation of senior practitioners is an open question across fields, not only in software.
The Historical Pattern
The Slop period — high-volume generation without architectural coherence — is structurally identical to the decade of electrified but unreorganized factories. The productivity data didn’t move until organizations redesigned around what the new medium actually made possible. Boiler explosions in the steam era preceded the development of mechanical engineering standards. Failure under high output created the pressure toward stronger practice.
The organizations making the shift from Generation-First to Spec-First earlier are following the historical pattern. The value accumulates in judgment — the capacity to specify clearly what should be built, and to confirm rigorously that what was built is correct. In software this is now explicit. In other intellectual fields, the same shift is forming.