Principles of AI Economy

Reading John Stuart Mill’s Principles of Political Economy, I found myself thinking about a worker whose abilities had not changed, but whose economic position had. The worker still knew the trade. Still put in the hours. Still came home tired. Yet the reward for that work bought less of what the household needed.

The history behind this thought is Europe’s Price Revolution. During the sixteenth and early seventeenth centuries, rising prices reshaped economic life. Silver arriving from the Americas contributed to this transformation, although inflation began before the largest inflows and had other contributing causes. It unfolded unevenly over generations. John Munro’s research explores these origins.

Its relevance today lies in a difficult possibility: an economic system can change the value of your position without any deterioration in your effort or ability. Artificial intelligence makes that possibility worth examining again.

1. Wealth must be measured by what it enables

Mill opens his discussion by challenging the identification of wealth with money and precious metals. A country’s stock of bullion is an inadequate account of its prosperity. What matters includes the useful things people can obtain and the capacity to produce them. This argument appears in the preliminary remarks of Principles of Political Economy.

Applied to AI, this suggests a practical test. Counting models, chips, subscriptions, or generated tokens tells us something about resources and activity. It does not tell us how much people’s lives have improved.

A business might generate ten times as many reports without making better decisions. A country might build substantial computing capacity while struggling to translate it into better services.

AI infrastructure can be productive capital. Its economic value depends on what that capital enables, after accounting for its costs. The meaningful measures are things such as reliable services, useful discoveries, affordable products, and time returned to people.

2. Abundance changes what earns a premium

When something becomes easier to supply, its price can fall, depending on demand, competition, and quality.

We have already seen sharp reductions in the price of some AI capabilities. Stanford’s 2025 AI Index reported that the inference price of performance comparable to GPT-3.5 on a particular benchmark fell more than 280-fold between November 2022 and October 2024. This measures a specific capability, rather than the full cost of delivering reliable work. The figures appear in the Stanford AI Index.

The economic question extends beyond cheaper software. If an acceptable first draft becomes easy to produce, what will customers still pay a premium for? Perhaps understanding the problem, accessing relevant information, verifying the result, or accepting responsibility when something goes wrong.

These are possibilities, not permanent protections. AI may change those activities too. The principle is that value depends partly on what remains difficult to obtain as other things become abundant.

3. Greater productivity does not guarantee greater income

Consider a hypothetical freelancer who can now produce twice as much work with AI. If prices remain stable and enough customers want the additional work, income may rise. If competitors gain the same capability and prices fall sharply, income may stagnate or decline. The outcome depends on demand and who captures the savings.

There is evidence that AI can improve productivity. A 2023 study of customer-support agents found a 14% average increase in issues resolved per hour, with larger benefits for less experienced workers. That is evidence from a particular workplace, not a universal estimate for the economy. The study, Generative AI at Work, was conducted by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond.

Producing more per hour and receiving more income per hour are separate outcomes. For households, the relevant question is what their earnings can buy. Cheaper digital services may help, but they may offer little compensation for lost income if housing, food, and other necessities remain expensive.

4. The distribution of gains is part of the technology’s economic impact

AI can help workers perform tasks, replace some tasks, and create demand for others. Economic research distinguishes the displacement of labour from the productivity gains that can generate additional demand. Neither effect alone tells the whole story. Daron Acemoglu and Pascual Restrepo examine these effects in Artificial Intelligence, Automation and Work.

A company’s savings might become lower customer prices, higher wages, greater profits, shorter working hours, or investment in expansion. Which outcome emerges depends on competition, ownership, bargaining power, institutions, and business decisions.

The technology creates possibilities. The arrangements around it influence who benefits. This is why an AI economy should be judged both by how much it produces and by how widely people can participate in its prosperity.

5. The speed of adjustment matters

Even a beneficial transformation can impose serious costs on people caught between an old source of income and a new one. Learning another skill takes time. Building a customer base takes time. Moving into a different occupation may require money that someone losing income does not have.

“People will adapt” is therefore an incomplete economic answer. We need to ask how quickly, at whose expense, and with what support.

For businesses, that means examining how employees can move into valuable responsibilities as tasks change. For governments, it means treating access, education, competition, and support during transitions as part of economic development. For individuals, it means understanding which parts of their work customers value—and how that value may change.

What struck me in Mill was the distance between possessing something associated with wealth and enjoying the substance of prosperity. AI invites us to examine that distance again.

We may acquire extraordinary capacity to generate answers, software, images, and analysis. The achievement will be turning that capacity into useful outcomes and a better standard of living.

The question I want to carry into the AI economy is simple: what does an ordinary person’s working day enable them to afford?