Let’s Talk About Jevons Paradox

Reference to a 19th-century insight about coal has, somewhat improbably, become one of the most-cited frames in the current AI conversation. If you have spent any time on the subject over the last 12-18 months, you have almost certainly come across it. It surfaces on earnings calls when an efficiency-driven price cut spooks investors. It appears in trade-press explainers and conference panels on AI displacement. It shows up in the recurring debate over data center electricity demand. It is now also standard fare in bar-association reflections on the future of legal practice, and is referenced in many conversations about AI in professional services.

To any long time reader of Guaranteed Dissent this phenomenon is already familiar. I first referenced it in my post On Silicon Dreams and Carbon Nightmares.

I want to push back, just a little bit. Not on Jevons. On the lazy version of the argument that hides behind Jevons’ observation.

What Jevons actually said

William Stanley Jevons published The Coal Question in 1865. He was watching James Watt’s steam engine become dramatically more efficient at burning coal, and noticed that England was burning more coal, not less. His core line:

It is a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth.

The mechanism is straightforward and we can observe it in many places. Make a resource cheaper to use, and use-cases that were previously uneconomic become compelling. Demand expands so much that aggregate consumption rises, sometimes well past where it started. Take lighting as an example: With the advent of LED lights it became economically feasible to have LEDs in almost anything and more lights were installed. We are building brighter streets, extending business hours, and adding decorative lighting in the most useless places. The rebound exceeds 100% — what energy economists later called backfire.

This is a real, durable, structurally interesting phenomenon. It is also a phenomenon about coal, or more generally fungible resources.

The lazy Jevons

In its current AI-discourse form, Jevons has - explicitly or implicitly - turned into a generic reassurance: whatever the efficiency gain is, demand will absorb it, the market will grow, it’s all good.

Legal is no exception. Take Thomson Reuters analysis of the Future of Professionals report:

When the work that law firms are doing today can be done at a lower price point, basic economic theory kicks in and demand will go up… This theory of demand expansion will more than compensate for the reform — not the repeal — of the billable hour, and the legal industry will continue to grow.

You can find variations of this almost everywhere now. Industry research firms have produced confident-looking forecasts for the industry: $1.05 trillion now, $1.38 trillion by 2030, all AI-driven. These numbers travel from report to report with no disclosed model underneath. [https://www.grandviewresearch.com/industry-analysis/global-legal-services-market]

The strongest version

The possibly strongest version comes from Richard Susskind, who has spent decades arguing that the legal market visible in billable hours is a small slice of the actual legal need in society: AI could crack the problem of unmet legal need by opening legal services up to a new generation of consumers. Harvey CEO Winston Weinberg points at the same gap from the supply side. On Sequoia’s Training Data podcast he observed that the average US lawyer rate is $352 an hour, “so almost no one can afford a lawyer”, and that even if every existing lawyer spent 20 hours a week on access-to-justice work, the gap wouldn’t close.

I don’t think they’re wrong. There is a vast pool of latent legal demand that simply cannot transact at current price levels. Were firms to lower the effective cost dramatically, a chunk of that market would buy.

What I believe deserves closer inspection, however, is whether treating the observation as evidence that ‘the legal pie grows’ is sufficient, as if that resolves the question of what happens to firms, lawyers, billable hours, or any particular segment of the market. It doesn’t. The way lazy Jevons is referenced is mostly meaningless.

Where the analogy is showing cracks

Coal was (analytically) clean. One fungible input, one common substrate — thermal energy — feeding many heterogeneous downstream uses: textiles, rail, smelting, electric power. Efficiency gains propagated uniformly through that single substrate. The market expansion showed up as one aggregate phenomenon, even though the downstream applications were wildly diverse.

Legal services are not coal. The approximate substrate might be legal expert reasoning over text. And while LLMs have most certainly lowered the cost of interacting with long form text, under the microscope their legal reasoning capabilities continue to underwhelm regularly (though they also impress on occasion), not to speak of architectural limitations that are inherent to the current form of the technology itself.

For legal services the heterogeneity sits at a different layer. ‘Legal services’ is not a single input being made more efficient. It is a portfolio of distinct cognitive tasks - contract review, regulatory filing, diligence or discovery, courtroom advocacy, strategic judgment to name a few. Each of these comes with very different elasticities of substitution to AI, very different latent demand pools, and very different efficiency curves under AI. The same model that 10x’s contract review might 50x discovery might barely move courtroom advocacy, which is bounded by court capacity. That equally applies to the demand side: It’s not that SMEs that couldn’t afford a lawyer don’t do business at scale today - they do, just without proper legal guidance and occasionally wading into gray zones. In this way, legal services fully reflect ‘jagged AI frontier’.

As a consequence, some segments in legal services will ‘Jevons hard’, and others will not. In this sense reference to Jevons is more akin to a mirage in the desert: the life-saving oasis is just past the dune, until the seemingly clear directions turn out to be rather fateful.

The aggregate effect on the legal market is much more the weighted sum of all of this. Without a cleaner understanding of the weights inside, the statement ‘AI will grow the legal pie’ is not wrong so much as it is instruction-free. It collapses a sectorally uneven, structurally complex reshaping into a single reassuring number.

Jevons really is a question, not an answer

Jevons in 1865 was not making a terminal claim. He wrote a 400-page book in which he modelled coal consumption curves under multiple scenarios, projected reserve exhaustion, engaged with policy responses, and made specific quantitative predictions about the economy. The Coal Question is the opening of a serious inquiry about what to do.

The current use of Jevons feels to me as attempting the opposite. Once invoked, the speaker has done what they came to do: place a respectable economic name on top of a hunch and walk away. A serious engagement with the framework, applied to AI in legal services, would at minimum have to take a position on five questions:

1. What input is being made more efficient? ‘Legal services’ is a heterogeneous bundle of cognitive operations, each ‘Jevons-ing’ different aspects at different rates.
2. What is the magnitude of the efficiency gain? Watt’s engine roughly doubled coal efficiency. Is the relevant AI-driven gain 2x, 10x, 100x? At which task, and measured how?
3. What is the elasticity of demand (and which one specifically) at the new effective price? Coal’s elasticity drove the rebound. Is legal demand actually elastic in the segment under discussion? Is lacking legal access hindering business? What is the size of the latent pool, and across which prices does it transact? And what if AI tools make legal knowledge so accessible to the demand side that they don’t actually need a service provider any longer?
4. What is the time horizon? Jevons’ coal rebound played out over decades. Are we talking about a two-year reshaping, ten years, thirty? Different timelines imply radically different decisions today.
5. What are the binding supply-side constraints? Coal had to be mined. What is the binding constraint as legal services scale — number of tokens, number of qualified reviewers, regulatory licensing, court capacity, malpractice insurance? Without a good grasp of the constraints, we cannot say where the rebound actually lands.
6. What if the underlying models’ coal - its tokens - become drastically more expensive because the foundation labs are running out of runway to support the current token economics?

I don’t profess to have answers to these questions. I don’t think anyone actually does. That is the honest answer. The dishonest version is to skip the questions, lean on Jevons as if it provided any answers, and treat the resulting reassurance as if it were earned.

I’m sorry, it is not earned. It is borrowed.

The situation follows structurally the same arc that ‘disruption’ went through in the 2010s. Framed in his 1997 book ‘The Innovator’s Dilemma’ Clayton Christensen’s original framework made falsifiable predictions about low-end market entry under specific conditions.¹ By the end of the decade ‘disruption’ had drifted into ‘anything new and threatening I want to gesture at without having to explain’. Same fate, same loss of analytical content under the pressure of needing a sophisticated-sounding shorthand. Jevons in the 2020s is going through the same decay, and the legal-AI conversation is one of the places it is most visible.

The systems lens

The aggregate can be directionally instructive, but it also hides where the action is. Most institutional confusion about complex systems shares a common shape. Someone notices a real phenomenon at the aggregate level - the market will grow - and treats it as if it answers a question about behaviour at the level where decisions actually are made: which workflows, which practice areas, which services, served by what kind of configuration of human and machine? The aggregate is only a headline number, and almost every instructive takeaway that matters is inside that number.

The interesting work that produces decisions you can actually act on, sits one level down. That is not a Jevons question. It’s an instrumentation and observability question. And an institutional one of ‘are we asking the right questions and do we have the tools to answer them?’ And on this level the reality is that most of the market doesn’t have the first idea of what this break down looks like even today, let alone how to reshape it for tomorrow.

So what

The lazy claim — AI will grow the legal pie, see, Jevons — is approximately as correct as it is operationally useless. It says nothing about which lawyers, which firms, which practice areas, which clients, which workflows, or which price points are about to be reshaped, eliminated, or born.

The honest version is more messy. The market will expand, but unevenly. Some segments will balloon. Some will be hollowed out. Some will barely move. The total dollar figure five years from now might be larger. But the composition will look almost nothing like the composition today, and the people best positioned to navigate the shift will be the ones who stopped asking whether the pie grows and started measuring the components that make up their slice.

Aggregates are reassuring. They are also where the most instructive detail goes to die.

Footnotes

1. Empirical tests have also shown that the theoretical foundations were less stringently applied than most people would have assumed. ↩