Sunil Venkataram
Back to garden
•
🌰 seedling
Jevons Paradox vs Cognitive Displacement - The Unresolved Tension
type/noteailaboreconomics

Jevons Paradox vs Cognitive Displacement — The Unresolved Tension

The two positions

Position A — Jevons Paradox applies

Classical Jevons paradox: when a resource becomes more efficient to use, total consumption of the resource increases rather than decreases, because the efficiency unlocks applications that were previously too expensive to pursue. Jevons studied coal in 1865: steam engines became more efficient, coal got “cheaper per unit of work done,” and total coal consumption went up, not down.

Applied to cognitive labor: AI makes human cognition effectively cheaper (either by substituting for human workers or by making human workers more productive). The reduction in cost unlocks applications that were previously uneconomic — personalized legal help for anyone with $20, medical second opinions for anyone with a question, tutoring for anyone learning anything, bespoke software for anyone with an idea. Each of these uses was rationed by cognitive labor scarcity before, and the scarcity is going away. Total demand for cognitive work goes up, not down, because the new use cases more than offset the reduced cost per existing use case.

The historical case: ATMs were supposed to eliminate bank tellers. Instead, bank branches became cheaper to staff, so banks opened more of them, and total teller employment increased for decades after ATMs became universal. Legal software was supposed to eliminate lawyers; legal spending grew. Computerization was supposed to eliminate administrative jobs; those jobs grew. Every previous wave of “labor-saving technology” produced this pattern.

See TAM of Intelligence is Infinite and Three Waves of AI Opportunity - Unhobbling, Physical Interface, Robotics for developed versions of this argument.

Position B — Cognitive Displacement

From Sahaj Garg’s 2026 essay The Displacement of Cognitive Labor and What Comes After, among others. The argument:

  • The productivity jump is categorical, not marginal. When a single well-crafted prompt produces what was previously four weeks of engineering work, the ratio is not 10% or 50% — it’s roughly 200x. Categorical jumps break historical analogies because reabsorption depends on the ratio being small enough that demand expansion can outrun displacement.
  • The timeline is compressed. Previous transitions (agricultural → industrial → service → knowledge) happened over generations, giving displaced workers time to retrain and new sectors time to mature. Cognitive labor displacement is happening in 3-5 years.
  • The usual escape hatch is gone. In every previous transition, workers displaced from one category moved into another. Farm workers moved to factories. Factory workers moved to offices. Office workers could move to… what, exactly? Garg’s core claim is that all three major labor categories (cognitive, emotional, physical) are being automated on overlapping timelines (cognitive: 3-5 years, emotional: 3-5 years, physical: 5-10 years via accelerated R&D). See Cognitive Automation Accelerates the Robotics Timeline. There is no next category.
  • Jevons expansion is slower than displacement contraction. Even if new cognitive use cases emerge, the pace at which humans can learn them and get paid for them is capped by retraining speed and trust formation. The AI workers filling the new use cases are already available; the human workers are not.

The prediction: majority of cognitive jobs substantially automated within 3-5 years. The $80K-$400K class gets hit hardest. See The Knowledge Work Cliff - Displacement of the Upper-Middle Class.

Where they actually disagree

The two positions agree on several things that might seem like disagreements:

  • Both agree AI is genuinely transformative. Neither dismisses the technology as hype.
  • Both agree total economic output will grow. The disagreement is about who captures the growth and what happens to specific workers in the interim.
  • Both agree physical labor is not a permanent refuge. Jevons proponents usually concede that robotics is coming; they just disagree on timeline.
  • Both agree the transition will be turbulent. Neither imagines a frictionless glide path.

What they actually disagree about:

1. The ratio of new-demand growth to displacement speed. Jevons proponents believe new cognitive use cases scale faster than old ones get automated. Displacement proponents believe the new use cases are being captured by AI too, not by humans moving into them.

2. Whether the reabsorption pattern holds when the thing being automated is cognition itself. Jevons proponents treat AI as a productivity tool like previous labor-saving technologies. Displacement proponents argue cognition is not like previous categories — because cognition was the input that let humans move from one automated category to the next, and when cognition itself is automated, that adjustment mechanism breaks.

3. How much of human cognitive work is genuinely substituted vs. merely augmented. Jevons proponents look at tools like Claude Code and see a productivity multiplier. Displacement proponents look at the same tool and see substitution that will complete as the tool improves. Both could be right for different slices of the labor market.

4. Political and social buffer. Jevons proponents assume the reabsorption will happen somewhere in the economic system. Displacement proponents argue the displaced class (educated, politically engaged, networked) is precisely the class that historically produces political instability when displaced, and that instability will shape the transition in ways the pure economic argument misses.

The empirical test

The way to actually resolve the tension: watch whether human employment in cognitive-work-adjacent categories grows or shrinks over the next 24-36 months, and at what velocity.

Specific things to watch:

  • Total software developer employment. The canonical test case. If Jevons is right, SWE employment should grow rapidly even as AI coding assistants take on more of the work, because cheaper code production unlocks new demand. If displacement is right, SWE employment should be flat or declining despite massive growth in total code produced.
  • Law firm employment at the mid-tier level. Contract review, document drafting, and compliance work are the most directly substitutable. If Jevons holds, total legal employment grows because legal services reach new customer segments. If displacement holds, the mid-tier lawyer class shrinks.
  • The ratio of AI-tool revenue to displaced-labor value. If AI companies are capturing most of the productivity gain (high revenue per user, few jobs created around the tools), displacement is winning. If the ecosystem around AI tools creates substantially more employment than the tools themselves destroy, Jevons is winning.
  • Entry-level knowledge work hiring. The leading indicator. Juniors are hired partly on the expectation that they will become the senior workers of the future. If companies stop hiring juniors because AI handles the work juniors used to do, the displacement thesis is being validated by the most cost-sensitive part of the labor market.

Early 2026 data is genuinely mixed — some categories show strong Jevons-style expansion, others show clear displacement. The next 12-24 months of hiring data will be the most important empirical signal for resolving the question.

Why it matters which is right

The two positions imply completely different policy and investment postures:

If Jevons is right:

  • Investment in AI tools companies is a productivity-multiplier bet
  • Policy should focus on accelerating adoption and retraining for the new use cases
  • The transition is bumpy but ultimately positive-sum for labor
  • Displaced workers are temporarily dislocated, not permanently redundant
  • The worry is inequality of access to the new tools, not aggregate job loss

If displacement is right:

  • Investment in AI tools companies is a capture-the-surplus bet — they are taking value from workers, not multiplying it
  • Policy must focus on income maintenance, identity support, and political stability through a transition with no natural reabsorption
  • The transition is zero-sum or negative-sum for labor at the aggregate level, not just at the margins
  • Displaced workers are structurally redundant, not temporarily dislocated
  • The worry is mass unemployment of the educated class and its political consequences

These are not small differences. A generation of policy and investment decisions will be made on which of these turns out to be closer to the truth.

The honest position

The intellectually honest answer as of April 2026 is that the question is unresolved and the people claiming certainty in either direction are overconfident. Both positions are held by serious people with serious arguments. Both are internally coherent. The empirical data does not yet decisively support either.

The useful posture:

  • Hold both positions simultaneously as live hypotheses, and track the leading indicators above to update in real time
  • Accept that the resolution will be specific by category — software development may follow Jevons while law follows displacement, or vice versa, and the aggregate answer may be a weighted average of very different category-specific answers
  • Prepare policy and personal strategy for a range of outcomes, not a point forecast
  • Watch for the empirical tell around 12-24 months from now — by late 2027 the direction will be much clearer than it is today
Backlinks (5)
Connected Notes