Somewhere along the way, someone got the idea that economics is about money. It is not, at least not in the way people usually mean that. Economics is about choice: what people want, what stands in their way, what they know, what they do not know, and how they behave when the rules around them change.
That is also a large part of what now matters about AI. Not the chips, not the demos, not the theatrical arguments about whether a chatbot is secretly alive. The important part is what these systems cause people and institutions to do. They rank, recommend, allocate, optimize, deny, approve, summarize, prioritize, and act. Those are decision functions. Economics has a great deal to say about them.
This matters because AI governance is still too often treated as a problem for computer scientists, lawyers, and ethicists, with everyone else invited later. That is better than treating it as a purely technical problem, but it still leaves out one of the fields that has spent the most time thinking carefully about uncertainty, incentives, and conflicting preferences. The result is a lot of AI governance that rediscovers old economic problems, often with worse vocabulary and less memory.
It also tends to miss the operational realities that operations researchers have spent decades formalizing: capacity constraints, throughput limits, queues, triage rules, and resource allocation. A model that flags more fraud than your investigators can review, or more packages than your inspectors can open, is not a working fraud detector or a working customs system. It is a demo.
What Economics Actually Is
Economics is the study of how agents make choices under conditions of scarcity. The money part comes later. Money is a coordination mechanism, a way of aggregating preferences and making exchange possible. The more basic questions are about decisions: what agents want, what constraints they face, what information they have, and what they will actually do when you put them inside a system with incentives.
That is not a small domain. Decision theory, which sits at the intersection of economics, mathematics, and philosophy, gives us formal ways to reason about choices under uncertainty. Mechanism design, sometimes called reverse game theory, asks how to build systems that produce desired outcomes even when the agents inside those systems pursue their own interests. Welfare economics asks whose preferences count and how we should think about them when they conflict. Information economics asks what happens when different agents know different things.
Those are not side issues for AI. A recommendation system is a mechanism. A reinforcement learning agent is an optimizer pursuing objectives under constraints. A large language model trained on human feedback is, among other things, an attempt to compress many human judgments into one system. The vocabulary is different, but a lot of the structure is familiar.
What Economists Know That AI Practitioners Keep Rediscovering
Goodhart’s Law, usually summarized as the idea that when a measure becomes a target, it stops being a good measure, comes from Charles Goodhart’s 1975 work on monetary policy. The AI community has been rediscovering it ever since under names like reward hacking, specification gaming, and teaching to the test. Optimize a system for a measurable proxy of what you actually want, and the system will usually find ways to score well on the proxy while damaging the underlying goal. That is not a weird edge case. It is one of the basic lessons of incentive design.
The principal-agent problem is the classic framework for understanding what happens when you delegate a task to someone whose interests are not perfectly aligned with yours, especially when they know things you do not. A large class of AI alignment problems has this shape. The principal may be the person, organization, or public body trying to get a system to behave in some intended way. The agent may be the AI system, the developer, the deploying organization, or some combination of them. The information asymmetry is that modern AI systems can act on patterns in data that humans cannot fully inspect, verify, or even notice. Economists have been studying that basic structure for a long time. AI alignment is not the same problem in new clothes, but it is close enough that ignoring the older literature is wasteful.
Externalities are costs or benefits that fall on people who were not party to the original transaction. Carbon emissions are the standard example. The cost of burning fuel is borne partly by people who had no role in the decision to burn it. AI systems generate externalities all the time. A content recommendation algorithm that maximizes engagement can impose costs on mental health, social cohesion, and information quality that do not appear in the objective function. An autonomous vehicle decision system that optimizes passenger safety can impose costs on pedestrians and other road users. Economists have a well-developed toolkit for analyzing and internalizing externalities. AI governance does not use that toolkit nearly as seriously as it should.
Preference aggregation, the problem of combining different and sometimes conflicting preferences into a collective choice, is one of the oldest and hardest problems in social choice theory. Arrow’s impossibility theorem, proved in 1951, shows that no social welfare function can satisfy a small set of apparently reasonable conditions at the same time. That matters for AI because systems that claim to reflect “human values” or “human preferences” are making an aggregation choice whether they admit it or not. Whose preferences count? How much? What happens when they conflict? These questions have formal treatments, and sometimes formal impossibility results. They should not be treated as vibes.
What Computer Scientists Know That Economists Miss
The critique runs both ways.
Economic models often abstract away implementation details. That is sometimes necessary, but it is also dangerous, because implementation details are often where the real behavior appears. A mechanism that is incentive-compatible in theory can fail in practice because the agents gaming it have computational advantages the theory did not account for. A policy that is optimal in a model can fail badly when deployed in a system with data quality problems the model assumed away.
The pace of capability change in AI is also faster than many economic models are built to handle. Much economic analysis treats technology as a slowly evolving background condition. But the capability gains in large language models and agentic systems since 2020 are abrupt enough to strain that assumption. Economists who are not paying close attention to what the systems can actually do may end up modeling a technology that has already moved on.
Data-generating processes are messier than economic models often assume. Economists are trained to be careful about endogeneity, selection bias, and measurement error. They are less often trained to think about the specific failure modes of large-scale machine learning: distribution shift, spurious correlations that survive in-sample validation, and training data that preserves patterns we should not want to reproduce. The machine-learning literature has developed tools and instincts here that many economists do not yet use routinely.
The Governance Gap
AI governance frameworks designed by computer scientists alone tend to focus on technical properties of systems: accuracy, robustness, fairness as defined by statistical metrics, and explainability. Those properties matter, but they are not enough.
AI governance frameworks designed by economists alone would have the opposite weakness. They would focus on incentives, market design, externalities, and preference aggregation, all of which matter, but none of which substitutes for understanding the system as built.
The point is not that economists should take over AI governance. They should not. The point is that a field that has spent decades asking what agents will actually do inside a system should be present when we design AI systems and the governance frameworks around them. Not at the end, after the objective functions have been chosen and the deployment context has already been fixed. At the beginning, while those choices are still being made.
AI is not just a computer science problem with economic side effects. It is a decision-making problem built out of software, data, institutions, incentives, and constraints. That requires computer science and economics, along with law, ethics, domain expertise, and operations research, from the start.
Somewhere along the way, someone got the idea that AI is the study of computers. That is not quite right either. AI is increasingly the study of decisions made through computers. Economics has been there the whole time.