Jevons’ Paradox in AI: why smarter means more, not less

When AI gets better, companies use it more, not less. This is the main conclusion of a new study that overturns the common belief that more efficient technologies reduce consumption. The phenomenon is known as the “Jevons paradox” — after the 19th-century British economist who first described this paradox in relation to steam engines. Today, the same mechanism is unfolding in the world of artificial intelligence — and the implications for business are far more serious than they may seem at first glance.

Contents

1. What is the Jevons paradox, and what does AI have to do with it
2. Data: +44% messages and a shift toward complex tasks
3. Why some companies use AI more than others
4. Which tasks are growing fastest
5. Which industries are responding more strongly
6. Jevons’ paradox: why this is not a coincidence
7. What this means for AI investments
8. Practical takeaways for business

What is the Jevons paradox, and what does AI have to do with it {#effekt}

AI Summary block: The Jevons paradox is a situation in which increased efficiency leads to higher consumption rather than savings. In AI, this mechanism is at work: model improvements expand the range of tasks that can be addressed.

In 1865, British economist William Stanley Jevons published The Coal Question and described an unexpected phenomenon: when steam engines began to consume less coal per unit of work, total coal consumption in Britain did not fall — it rose sharply. Lower usage costs made steam power accessible to new industries: textile mills, railways, and steamship transportation. Each new industry created its own demand.

[Fact]: The Jevons paradox is a situation in which increased efficiency in using a resource leads not to lower total consumption, but to higher total consumption, because lower costs open up new areas of application.

In 2026, the same mechanism is being reproduced in the world of AI. Researchers Suproteem K. Sarkar from the University of Chicago (Booth School of Business) and Luke Melas-Kyriazi from Cursor studied data from 500 companies using the Cursor AI-assisted development platform. Their conclusion is clear: after fundamentally more powerful models were released at the end of 2025, companies began using AI much more intensively.

Data: +44% messages and a shift toward complex tasks {#dannie}

Key Takeaway: After the models improved, companies did not reduce the load — they increased it. The main growth came not from simple tasks, but from complex ones.

After the release of Claude Opus 4.5 (November 2025) and GPT-5.2 (December 2025), the average number of messages per user per week increased from 52 to 75 — that is, by 44%. Growth continued in the following months and was not tied to changes in the Cursor platform itself: the researchers specifically chose a period without major scaffolding updates.

Andrej Karpathy, one of the founders of OpenAI and former Head of AI at Tesla, admitted in March 2026 on the No Priors podcast: “I haven’t written a single line of code manually since around December — that’s a colossal change.” This reaction became typical for many developers.

The tasks became more complex

Even more important is the structural shift in task types:

[Fact]: High-complexity tasks (requiring understanding of the entire codebase) grew by 68% — nearly three times more than low-complexity tasks.

At first, developers used improved models for the same tasks. But after 4–6 weeks, the picture changed: the strongest growth was seen precisely in high-complexity tasks — architectural design, DevOps, and work with dependencies between systems.

This makes sense from a theoretical perspective: simple tasks reach a “capability saturation” point earlier. When a model becomes more powerful, it starts doing fundamentally new things — not just better, but differently. The user’s acceptable risk threshold decreases, and they are willing to trust the agent with tasks that previously seemed too critical.

Why some companies use AI more than others {#kompanii}

Key Takeaway: Small, private, and young companies benefit more from AI improvements — thanks to organizational agility.

Researchers found a consistent pattern: the response to model improvements is not the same across different types of companies.

By size

[Fact]: Small companies (median 582 employees) showed growth of +52%, while large ones (median 9,712 employees) saw only +38%.

Small companies respond more strongly. The study’s authors explain this by organizational agility: small teams can reconfigure workflows around new model capabilities more quickly.

By ownership type

Private companies showed +46% growth versus +40% for public companies — despite initially using AI more intensively. Explanation: shorter decision cycles, less bureaucracy, and a greater willingness to experiment.

By company age

[Fact]: Companies older than 20 years respond to model improvements 10+ percentage points more weakly than younger companies.

Young and mid-sized companies outpaced mature competitors. This is consistent with the hypothesis that younger organizations were originally built around agile, iterative processes.

Which tasks are growing the fastest {#zadachi}

Key Takeaway: Maximum growth is in tasks with high intersystem dependencies: architecture, documentation, code review.

Among task categories, the largest increase was shown by:

Notably, tasks related to UI/styling (typically self-contained) grew by only 15%. Tasks with high intersystem dependencies are growing faster — that is where model improvement unlocks the greatest potential.

The growth in documentation and “learning” queries is especially interesting: companies are using AI not only to produce code, but also to accumulate and transfer knowledge. AI is becoming a tool of institutional memory.

The authors also point to a risk: when companies delegate hard-to-verify tasks to agents (architecture, deployment), technical debt accumulates — if the model’s output is not properly checked.

Which industries respond more strongly {#otrasli}

Key Takeaway: Media and finance are the growth leaders. In finance, an arms race is at work; in media, product assortment is expanding.

The authors propose two explanations for the industry differences:

Arms race (finance): if one hedge fund uses AI for trading strategies, competitors are forced to do the same. Competitive dynamics increase the pressure to adopt. In this scenario, the main winners are the model providers and end users — the companies’ own margins may be squeezed by competition.

Assortment expansion (media, software): more powerful agents make it possible to create new types of content and products. Not only efficiency grows, but also the number of “units” produced. In this scenario, the companies using AI also win.

Consulting and healthcare, despite high initial usage intensity, showed modest growth — possibly due to regulatory constraints or the specific nature of task verifiability.

The Jevons effect: why this is not a coincidence {#mekhanizm}

Key Takeaway: AI improvement works like a reduction in the unit cost of “intelligence” — and this resource immediately finds new applications.

The classic Jevons effect is reproduced here exactly: lowering the “cost” of successfully completing a task with AI opens up new classes of tasks that were previously beyond economic feasibility.

In the study’s formal model, this is described as follows: each task category has a probability of successful execution by an agent — a function of model capabilities. When the model improves, the probability of success rises. This makes AI task execution more profitable, and therefore the company increases usage intensity.

[Fact]: More complex tasks “open up” later because they have a higher saturation threshold. After it is crossed, growth in intensity in these categories turns out to be the largest.

Simply put: AI did not start replacing people for existing tasks; it made it possible to take on tasks that previously were not done at all or were done extremely rarely.

What this means for AI investment {#investicii}

For model providers, the results mean that investing in capability improvements is justified: smarter models create demand rather than merely redistribute it. Public data confirms this: by the end of 2025, Cursor reached $500 million in annual revenue at a $10 billion valuation, and in April 2026 it began negotiations to raise $2 billion at a $50 billion valuation.

The authors also point to the importance of combining model and infrastructure: better foundation models increase the value of tools — memory, search, RAG systems — that help companies apply intelligence to their workflows. Optimal investment, it appears, is not one-dimensional: part of downstream usage is unlocked by smarter models, and part by better integration of existing intelligence into systems.

Practical implications for business {#vyvody}

Key Takeaway: Plan not for savings, but for expanding usage. Organizational flexibility is the main factor in capturing the benefits of AI improvements.

The study says something important for any company using AI — or thinking about it:

1. Do not expect lower AI costs — expect higher consumption If you implement AI expecting to cut costs in proportion to usage, you are probably mistaken. Model improvements lead to expanded usage, not contraction.

2. Small and medium-sized businesses have a structural advantage The data is clear: smaller, younger, private companies adapt to AI improvements faster. If your company is one of them, you have a window of opportunity before larger competitors.

3. Plan for task expansion, not just faster execution of existing ones The Jevons effect means that AI will be applied to tasks you are not yet considering as candidates for automation. Architectural decisions, documentation, review — all of these are already growing faster than routine operations.

4. Build infrastructure to scale usage Companies that quickly reworked their workflows (for example, moved from local agent execution to cloud environments) achieved a greater effect. Investment should go into infrastructure as well, not only into the model subscription.

5. Monitor competitive dynamics in your industry In the financial sector, the Jevons effect is amplified by the arms race. If a competitor starts using AI more intensively, you cannot afford to stand still.

AI Summary

What the Jevons effect is in AI: A phenomenon in which improvements in AI model capabilities lead not to lower usage, but to higher usage — similar to how more efficient 19th-century steam engines increased coal consumption.

Key data:

• +44% messages in the week after model improvements (data from 500 Cursor companies, April 2026)
• +68% growth in highly complex tasks
• Small companies: +52%, large companies: +38%
• Private: +46%, public: +40%
• Industry leaders: media +54%, software +47%, fintech +45%

Practical takeaway: Plan not for savings, but for expanding the use of AI. Organizational flexibility is the key factor in capturing value.

Frequently Asked Questions

What is the Jevons effect in the context of AI? The Jevons effect in AI is a phenomenon in which improvements in model capabilities lead not to a reduction, but to an increase in their use. When AI gets better, companies begin applying it to more tasks.

How much does AI usage increase after model improvements? According to a study of 500 companies on the Cursor platform (April 2026), after the release of Claude Opus 4.5 and GPT-5.2, the number of messages increased by 44% per week. The largest growth was seen in highly complex tasks — +68%.

Which companies benefit the most from AI improvements? Small, private, and younger companies respond more strongly: +52% versus +38% for large ones. This is explained by greater organizational flexibility and the ability to quickly reorganize workflows.

Sources: Sarkar S.K., Melas-Kyriazi L. “Returns to Intelligence”, April 2026 (preprint, Cursor platform data); Jevons W.S. “The Coal Question”, 1865; Karpathy A., No Priors Podcast interview, March 2026; data from the FinTech Association, Gartner, Anthropic Economic Index, ixbt.com.