Jevons paradox explains why efficiency can expand demand. It does not explain who captures that demand.
Reshuffle concept · deep dive
Jevons Misunderstanding.
People overuse Jevons paradox to argue that AI will save workers - cheaper work will create demand for more work.
Demand does increase. It does not flow to workers. As the data below shows, workers are often left disadvantaged.
About this page
This is an argument with data to play with. Not an evidence base.
The work draws from years of thinking and writing about platform structure, market mechanics, and tech-driven reshuffles. The data is an editorial synthesis from public sources: citable where the numbers can be sourced, and indicative where the picture is bigger than any one source.
This isn't empirical evidence and shouldn't be treated as such. It's a structured attempt to reveal patterns and structure in prior analysis and public data.
Play with the controls. Click through the references. Drag the sliders. And look for where the pattern doesn't yet fit a case you know.
The popular claim
Where does this argument come from?
Stanley Jevons noticed that more efficient steam engines didn't reduce coal consumption. They expanded it. Coal output grew several-fold. Miner employment and wages rose with it. The modern AI argument borrows this precedent directly.
The popular claim
"AI will create more knowledge work than it absorbs. Augmented workers stay productive; the market expands; the labor market wins."
Where this claim shows up 5 sources
-
McKinsey Global Institute June 2023 The economic potential of generative AI: The next productivity frontier"Generative AI could add the equivalent of $2.6 trillion to $4.4 trillion annually across 63 use cases. Most of the value will come from augmenting current job activities, not replacing them."
View source → -
World Economic Forum January 2025 Future of Jobs Report 2025"AI and information processing technologies are expected to create 11 million jobs while displacing 9 million - a net positive shift in the global labour market."
View source → -
Goldman Sachs Research March 2023 The Potentially Large Effects of Artificial Intelligence on Economic Growth"Generative AI could raise global GDP by 7% (almost $7 trillion) and lift productivity growth by 1.5 percentage points over a 10-year period. Most exposed workers will see their work complemented, not substituted."
View source → -
LinkedIn Workforce Reports October 2024 Future of Work Report: AI at Work"Jobs mentioning generative AI skills have grown 6× year-on-year. AI-skilled professionals are commanding wage premiums and seeing accelerated hiring across categories."
View source → -
OpenAI 2023–2024 Public statements + GPT-4 economic impact whitepaper (Eloundou et al.)"GPTs are general-purpose technologies. They will augment workers - including knowledge workers - and expand the surface of what individuals can produce."
View source →
Where the claim comes from
1865 Coal · Steam engine
- Coal output grew several-fold
- Miner employment rose with output
- Miner wages rising in real terms
Stanley Jevons noticed in 1865 that more efficient steam engines did not reduce coal consumption - they expanded it. Miners sat above the engine. The engine made coal cheaper to extract; miners still did the extraction. Demand grew; labour demand grew with it.
Sources
UK coal output and miner employment 1860-1900 drawn from B. R. Mitchell, British Historical Statistics (Cambridge University Press, 1988). Real-wage series from Feinstein (1995) and Allen (2007). The pattern holds across the major coal-producing economies of the period.
The evidence
Is it actually playing out this way?
In AI-mediated markets, the surplus flows to whoever sits above the algorithm: the actor who owns the customer, workflow, accountability, standard, system, or outcome.
Workers below the algorithm may remain in the loop, but they do not benefit from the augmentation or from the increasing demand.
Across six AI-exposed roles, industry revenue rose.
Per-worker value held flat or fell.
Translators
Worker feeds AIPlatform capture
Market ↑Per worker ↓
Industry revenue grew.
Per-word translator pay dropped roughly in half.
Customer support
Worker feeds AIPlatform capture
Market ↑Per worker ↓
AI customer-support vendors hit hundreds of millions in revenue.
Worker pay shifted to gig contracts.
Data annotation
Worker feeds AIPlatform captureAuthority rent
Market ↑Per worker ↓
AI training-data labour pool exploded.
Per-task pay collapsed to floor.
Performance marketing
Platform capturePlatform arbitrage
Market ↑Per worker ↓
Programmatic ad spend kept growing.
Marketer fee-share kept shrinking.
Recruiter sourcing
Data platformPlatform capture
Market ↑Per worker ↓
Application volume scaled to billions.
Sourcing-recruiter pay compressed.
Junior software
Data platformAI-capital rent
Market ↑Per worker ↓
AI-coding-tool revenue exceeded GitHub's pre-acquisition revenue.
Junior-coding pay compressed.
Counterfactual
What Jevons paradox misses about AI
In 1865, coal output, miner employment, and miner wages all rose together. The market expanded and the workers expanded with it. In translation, and in every AI-exposed role behind it, output and employment move in opposite directions. The Jevons mechanism is the same. The structural conditions that let workers benefit are not.
Coal, 1860 – 1900
Jevons working
Steam engine becomes radically more efficient at burning coal.
- Coal output grew several-fold
- Miner employment rose with output
- Miner wages rising in real terms
Miners sat above the engine. The engine made coal cheaper to extract; miners still did the extraction. Demand grew; labor demand grew with it.
Shape: Three arrows up
Sources
UK coal output and miner employment 1860–1900 from B. R. Mitchell, British Historical Statistics (Cambridge University Press); real-wage series from Feinstein (1995) and Allen (2007).
Translation, 2020 – 2026
Jevons insufficient
AI achieves acceptable-quality machine translation at near-zero marginal cost.
- Words translated grew several-fold
- Translator employment BLS occupation: roughly flat
- Per-word rate post-editing pay $5–18/hour
Translators sit below the algorithm - their post-editing work feeds the machine-translation pipeline rather than going to a client. The market expanded; the workers did not catch the expansion.
Shape: Scissors
Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Platform captureWhere value re-anchorsCustomer-channel captureWhen a customer-channel platform aggregates fragmented workers, the platform captures the customer-of-record, the price discovery, and the reputation signal. Each worker is repriced as commodity throughput against the AI floor.Worker-stays mythLiterature blind spotWorker-stays-assumptionStandard labour-market analyses assume the job stops as a coherent bundle when AI absorbs a task. Exposure-by-occupation methods can't see the bundle dissolving - the role can stay alive as a name while becoming an input slot in someone else's system.
Sources
Industry words machine-translated: Slator Language Industry Market Report and CSA Research industry-volume series. Translator employment: BLS Occupational Series 27-3091 (Interpreters and Translators). Per-word rates: ProZ freelance compensation surveys and industry post-editing rate ranges.
Jevons misunderstanding
Does expanded demand mean expanded wages?
Jevons paradox predicts that AI's cost collapse expands the market for the work. The popular reading assumes wages follow the expansion. The labor data says otherwise.
The pattern kicks in when criticality stays high but compensation drops to throughput. Most of the skills AI made more needed live in this gap.
Hover a dot for the role name; click to jump to its card on the right.
Flag Criticality gapDurability modifierTuareg trap (criticality-vs-compensation gap)Named after the Tuareg navigators of the Sahara - indispensable to caravan crossings but historically paid at commodity rates. A scarce capability is hired as commodity labour when the buyer refuses to acknowledge its criticality. The work looks defensible (the task can't be done without the worker) but compensation prices it as throughput. Augmentation mythLiterature blind spotAugmentation-as-positive-coding'Augmentation' reads as positive for the worker by default in most economic writing. Often the augmentation is just platform or AI-capital arbitrage with extra steps - the platform or model provider extracts the productivity gain, not the worker.
-
Translation post-editor (low-resource languages) Market: Machine translation has made it economical to translate into languages that previously had no commercial market - Irish, Brazilian Portuguese SME review, and similar long-tail languages. Translation volumes have grown several-fold as a result.
Wage: Per-word post-editing rates have fallen from roughly $0.10 to $0.03-0.05 over the past several years. Native-fluent reviewers remain irreplaceable in the language, but the work is priced as throughput rather than expertise.
-
RLHF labeler (Kenya / Philippines) Market: Every major model release depends on human-curated reward signal and preference data. Demand for labeling work from frontier AI labs has grown roughly an order of magnitude year over year.
Wage: Pay sits at the global labor-market floor, typically $1-2 per hour in Kenya and the Philippines. The labels train the models everyone else uses, but the economic rent accrues to the lab rather than the labeler.
-
Customer support agent Market: AI chatbots now handle roughly two-thirds of routine customer queries. The residual third - complex cases that require judgment, escalations, and trust-building - routes to human agents, and volume per agent has grown as overall ticket counts rise.
Wage: Headline rates look high (Klarna paid $41/hr at the peak of its AI-replaces-CS strategy), but the tier is mostly 1099 contractors without benefits. Klarna publicly reversed its automation strategy in May 2025 when the math of scaling humans for the residual third didn't work.
-
Per-stamp architect-of-record reviewer Market: AI-generated architectural and engineering drawings have multiplied the review pipeline. Each project now requires more architect-of-record stamps to verify and certify the output.
Wage: Per-stamp rates remain flat at $80-120, prices that were set decades ago and have not been repriced for the new volume. The professional liability stays with the architect, but the rate is paid as throughput.
-
Content moderator Market: User-generated content volumes have grown across every major platform. Automated classifiers catch the obvious violations, leaving the cases that require human judgment - often the worst content - to moderators.
Wage: Compensation sits in the BPO tier, typically $1-3 per hour for outsourced contractors. The platform captures the moderation economic rent; the worker absorbs both the throughput pressure and the psychological cost.
-
Junior paralegal Market: Firms accept more matters because AI legal tools - Harvey and its competitors - lower the cost per matter. Document review, due diligence, and contract analysis pipelines have grown substantially.
Wage: The junior billable-hour tier is compressing as fewer hours bill out per associate. BLS lists the median paralegal salary at $59,200/yr, but the entry tier is shrinking - the partner captures the markup that used to flow through junior hours.
-
Junior software developer Market: AI coding assistants have expanded what companies can produce per engineer. Codebases grow faster, product teams take on work they previously declined, and the volume of routine code shipped has multiplied.
Wage: Brynjolfsson, Chandar, and Chen's 2025 Stanford/ADP study finds that employment for software developers aged 22-25 fell nearly 20% from its late-2022 peak. AI is compressing the entry-level/routine coding layer while increasing the premium on senior engineers who own architecture, system design, product judgment, and verification.
-
Algorithmic-hiring recruiter Market: Application volume routed through algorithmic screening platforms has exploded. Discovery in Mobley v Workday surfaced 1.1 billion applications processed through a single vendor's pipeline.
Wage: Recruiters increasingly operate the AI scoring system rather than evaluate candidates directly. The tier is compressing fast as the platform owner - Workday and similar - captures the margin that used to flow to in-house recruitment.
-
Citation-audit paralegal Market: After Mata v. Avianca in June 2023 - where ChatGPT-generated case citations led to sanctions - every AI-assisted legal brief now requires a human substance-validation pass. Demand for citation auditing has grown across firms.
Wage: Compensation has risen because the substance test is non-delegable. The work cannot be done by AI without recreating the original problem, so the seat catches the Jevons expansion rather than falling under it.
8 of 9 plotted roles sit in the Jevons-misunderstanding zone. The market needs the skill more than ever. The wages are lower than ever before.
The diagnosis
Why does the expansion bypass workers?
With AI and platforms, efficiency gains do not just make the old production system larger. They can re-architect the production system. The algorithm can capture the customer, route demand, set prices, define quality, observe performance, modularize tasks, and use worker output to improve the system. So expansion no longer has to pass through the old labor bundle.
Expansion
Worker bypassed
Route 1
The platform sits between you and the customer.
Mechanism
When a delivery / matching layer aggregates fragmented workers, the layer captures the customer-of-record + price-discovery + reputation signal. Each worker is priced as commodity throughput against the AI floor.
Case · Performance marketing
Meta Advantage+ generated $20B/year advertiser revenue at $4.52/$1 ROAS. Google PMax reached 1M+ advertisers; 71% global / 75% US adoption. The platform owns bidding, targeting, and creative-variant selection - the work that used to be the marketer's premium.
Falsifier
Refuted if workers retain direct customer relationships AND price discovery happens outside the aggregator AND reputation portability across aggregators is real.
The capture stack
Where did the expansion go?
If workers didn't catch the expansion, where did the dollars go? Pick a role. Drag the year slider. The data-platform and customer-channel layers grow. The worker layer thins. The producing role doesn't receive the value the expansion created.
$100 of localization spend
Worker sliverMTPE per-word is what the translator gets.
- Data platformData platform32%
The platform that owns the data infrastructure (Westlaw for legal, Bloomberg for finance, customer-data platforms for support) captures the productivity gain from the role's accumulated knowledge.
- Automation inferenceAI-capital rent15%
Foundation-model or platform-compute layer; the AI capital owner takes its cut.
- Customer channelPlatform capture25%
The platform that owns the buyer relationship - matching, pricing, reputation - captures the channel rent the worker used to capture.
- WorkerWorker feeds AI28%
What the human doing the work actually keeps after the platform, inference, and channel layers take their cuts. This share thins as the others grow.
Leading indicator
Which roles enter this pattern next?
In the short term, AI-mention share of job postings is a leading indicator of pushing workers into roles that train the model but fail to gain leverage from it. Analysis of job postings for the jobs mentioned below showed a high concentration of jobs below the algorithm rather than above it.
Postings change before hiring changes. When the share doubles inside a year, the role is being absorbed into AI-mediated work at the hiring margin. In 2025, HR and Marketing doubled fastest. Source: Indeed Hiring Lab , January 2026 update.
+130%
AI-mentioning postings vs Feb 2020 baseline
while total postings sit at only +6% above baseline · YoY −5.2%
Detail · Marketing
What this tells us
"Marketing's AI-mention share nearly doubled in 2025 - from 8.4% to 14.9% - entering the bundle-absorption phase visibly."- Indeed Hiring Lab
Performance marketing is hiring new seats into platform-execution work. Meta Advantage+, Google PMax, and TikTok Smart+ now own the optimisation; the marketer operates the platform's controls rather than holding the optimisation themselves. The seats grow; the per-seat pay compresses.
The four positions
How can workers benefit from the expansion?
Four positions are reachable as individual practice. They divide cleanly along two questions: how much capital does reaching the position take, and where does the premium come from - delivering value to a buyer, or refusing the algorithmic alternative? Click a quadrant to see who holds the position today.
Sovereign capital
Own the AI stack inside the offering you own.
Customer ownershipCapital tierRecombination premium
A practitioner who owns the customer's new problem and the AI stack inside the offering - not as a vendor's seat, but as their own equity.
Holding it today
- Practitioner-turned-founderBuilt the AI platform for the craft they used to do; sells to the same buyers
- Bootstrapped vertical-SaaS founderSingle-function tool, no VC, reinvests revenue, owns the customer relationship
- Category-defining AI product founderBuilt a new product category powered by AI as the core ingredient
- Indie maker of a focused AI toolSmall team, one product, distributes direct to the buyer without a platform middle
Three more positions catch the expansion - vendor seat, data platform, regulatory body - but require platform, capital, or institutional backing rather than being reachable as individual practice.
Self-diagnostic
Where do you personally sit?
Five questions about your work. Each answer scores along four axes - customer ownership, constraint holding, capital tier, and how the buyer pays. Watch your position move as you answer. Once all five are in, change any answer to see what would shift you to a different quadrant.
All 5 answered. Your position is below.
The bars show your axis scores from your answers. Drag the dot on the spectrum to explore what other positions look like - the bars reset because they only describe your answers. Click any number above (or the ✓) to bring your scores back.
Who is your buyer-of-record?
Who does the money flow from? Whose contract or invoice carries the work?
Where does your output go after you produce it?
Who or what is the next reader, validator, or consumer of the work?
What does the buyer actually pay you for?
The unit of value the engagement transacts on - outcome, output, or time.
Do you hold a constraint the system can't replicate?
A signature, seal, regulatory accountability, institutional authority, or risk-bearing seat.
Where does the AI stack sit, relative to what you sell?
Who owns the model, the inference, the training data, the platform that runs your work.
Your position so far
Answer to see your position Above the algorithm Below the algorithm
Your computed position -
- Customer ownership
- -
- Constraint holding
- -
- Capital tier
- -
- What buyers pay for
- -
What you have -
What you'd need to add -
Click any number above to revisit a question. Change an answer to see what would shift your position.
Conclusion
Three key points
The paradox correctly predicts that cheaper work expands the market for the work. The wage outcome for coal - workers riding the expansion - required structural conditions that no longer apply in the algorithmic and AI age.
Three layers absorb the gains produced by the cost collapse: the data platform, the inference layer, and the customer channel. The worker layer thins as the market grows. This is a structural shift.
Capturing the new value happens by holding one of four seats - the customer, the credential, the equity, or the accountability counterweight. Move toward a position. Don't defend a category.