Ask almost anyone what should be done, and the answer is immediate: the rancher caused the damage, so restrain the rancher. The familiar remedies follow—a fence, a tax, or liability requiring the rancher to compensate the farmer.

Ronald Coase’s 1960 paper, The Problem of Social Cost, challenged its intuition. It became one of the most cited papers in economics. It introduced what later became known as the Coase Theorem—a name Coase himself never accepted—and helped launch both law and economics and the new institutional economics. Yet most textbooks present it as a clever solution to externalities. That reading misses its central insight. The paper is fundamentally about legal institutions and the costs of defining, enforcing, and transferring rights.

This is the fourth piece in a series that reads one classic paper at a time with AI at your side, alongside the book I wrote with Raul A. Sosa, AI at Your Side: The Student’s Guide to Smarter Learning, forthcoming from Oxford University Press. The place to start is the analysis Coase was writing against.

The externality and the Pigouvian tax

Consider the rancher’s decision one steer at a time. Each additional steer costs a little more to raise than the previous one because the best pasture is used first, and each causes $100 of damage to the farmer’s crops. The rancher bears the first cost but not the second. That unpaid cost is the externality.

Table 1. Each steer does $100 of damage the rancher never pays for. Society bears both costs.

Suppose a steer sells for $300. The rancher adds steers as long as the next one costs him up to $300 to raise, so he keeps five. From society’s perspective, the cost is the rancher’s production cost plus the crop damage, and by that measure only three steers are worth raising. The fourth destroys $50 of social value and the fifth destroys $100. The private herd is five, the socially efficient herd is three, and the $150 of destroyed value is what economists call a deadweight loss.

Figure 1. The rancher stops where his own cost meets the $300 price: five steers. Society would stop at three. The shaded gap is the deadweight loss, $150.

The standard cure comes from Arthur Pigou, and it is the one economists have traditionally reached for first: tax the rancher $100 per steer, the amount of the damage. The tax folds the farmer’s loss into the rancher’s own cost, and he stops at three on his own. The tax does not distort the market. It removes a distortion that was already there. For forty years this was the answer to an externality: measure the gap between private and social cost, and tax it away. One policy tool for one objective, as Jan Tinbergen taught.

Asking an AI what an externality is returns a definition you already knew. Handing it the table and making it work through the numbers is a different exercise.

Prompt: “Here is a cost schedule for a rancher whose steers damage a neighbor’s crops [paste Table 1]. A steer sells for $300. Find how many steers he raises on his own and how many are efficient for society. Show the steer-by-steer comparison, compute the deadweight loss, and tell me in one sentence why the gap between his cost and society’s cost is exactly $100.”

The response (generated using Claude):

The rancher looks only at his own cost: $100, $150, $200, $250, $300. The fifth steer, at $300, is his last; a sixth would cost $350. He keeps five. Add the $100 damage and a steer costs society $100 more, so the third is the last one worth raising and the efficient herd is three. The waste is $50 on the fourth steer and $100 on the fifth, a $150 deadweight loss. The $100 gap between his cost and society’s is exactly the damage he never pays: erase it, and the two herds become one.

The point of the exercise is not the answer, which you already have. It is to check whether the machine reasons the way you would: whether it sums the two steps of the loss, $50 plus $100, instead of giving one blended number, and whether it connects the gap to the damage. The checking is where the learning happens.

Coase’s argument

Pigou’s analysis treats the rancher as the cause of the harm and the farmer as its victim. Coase rejects this framing. The harm, he argues, is reciprocal: restraining the cattle protects the farmer’s crops, but it also reduces the rancher’s production of meat. “The real question that has to be decided is: should A be allowed to harm B or should B be allowed to harm A? The problem is to avoid the more serious harm.” The choice is between meat and crops, and calling the rancher the culprit does not say which is worth more.

From that reframing Coase derives the result. Suppose the two can talk, a court has assigned the right clearly, and reaching and enforcing a deal cost nothing. Then the herd ends at three steers no matter who holds the right. If the rancher is liable, he pays $100 per steer in damages, faces the full social cost, and stops at three. If he is not liable, the farmer pays him to shrink the herd. The fifth steer sells for $300 and costs $300 to raise, so it earns the rancher nothing. The fourth costs $250, so it earns him only $50. Each does $100 of damage, so the farmer gladly pays the rancher to give both up, and the trading stops where the rancher’s profit on a steer equals the damage it does. Either way, three steers.

That is the Coase Theorem as the world came to state it: when rights are clear and bargaining and enforcement is costless, the allocation of resources, here the size of the herd, does not depend on who holds the rights. Coase shows it with the same cattle example. Such agreements, he writes, “would not affect the allocation of resources but would merely alter the distribution of income and wealth.” The assignment of the right determines who pays whom, not how many steers get raised.

The theorem is best understood by taking the costless bargaining away. Ask the AI to redo the problem with a friction.

Prompt: “Confirm the herd is three under both liability rules, and say who pays whom in each. Then add a friction: reaching a deal costs $200, more than the $150 the efficient herd would save. Redo both rules. In which does the herd fall to three, and in which does it stay at five?”

The response (generated using Claude):

With free bargaining the herd is three either way. Under liability the rancher pays the farmer; under no liability the farmer pays the rancher. Now add the $200. Under liability the rancher owes for damage whether or not they ever negotiate, so he still cuts to three. Under no liability the only fix is a deal worth $150 that costs $200, so no one makes it, and the herd stays at five. The right the rancher started with now decides the outcome.

The important difference between the two cases is this. Under liability, no bargain is required, so transaction costs never become relevant. Under no liability the efficient outcome requires a deal that costs more than it saves, so the deal is never struck. When bargaining and enforcement are free, the assignment of the right changes only the direction of payment. When it is costly, it changes what gets produced. That cost of striking and enforcing a deal is what economists call a transaction cost.

Stigler and the name

The Chicago economists were initially skeptical of the result. Coase had first advanced it in a 1959 article on how the government allocates radio frequencies, and Aaron Director invited him to dinner at his home so that the Chicago economists could show him he was mistaken. About twenty economists came, Milton Friedman and George Stigler among them. Coase asked the room to assume a world without transaction costs and to accept that, in such a world, the assignment of liability would not affect the allocation of resources. Stigler recalled the evening in his memoirs: Friedman “did most of the talking, as usual,” and over two hours of argument “the vote went from twenty against and one for Coase to twenty-one for Coase.” It was Stigler who later named the result the Coase Theorem. (Historian Steven Medema notes that no vote was actually taken. The tally was Stigler’s dramatization, though Coase agreed it captured the spirit of the evening.)

Coase never accepted the theorem as a statement of his contribution. The frictionless world of zero transaction costs was a device for exposing Pigou’s framework, not a description of any actual economy. He made the point years later, in The Firm, the Market, and the Law: “The world of zero transaction costs has often been described as a Coasian world. Nothing could be further from the truth. It is the world of modern economic theory, one which I was hoping to persuade economists to leave.”

Property rights, transaction costs, and the legal system

The contribution, then, is not a solution to the externality problem. It is a different way of thinking about what determines economic efficiency.

Transaction costs are never zero. Defining and enforcing legal rights is costly, and when those costs are high, efficient bargains are never struck. Then, as Coase wrote, “the initial delimitation of legal rights does have an effect on the efficiency with which the economic system operates.” An economy is a system of rights: to use land, to emit smoke, to make noise, to be free of it. Producing anything requires combining rights, and combining rights requires transacting. The more it costs to define a right, to find its owner, to transfer it, and to enforce it, the more beneficial trades never happen. The efficiency of an economy depends crucially on the transaction costs its property-rights system generates. That observation is the starting point of the new institutional economics. It is also why much of the paper reviews English court cases about smoke, noise, and straying animals. When transacting is costly, a judge who decides who holds which right is deciding how resources will be used.

It also explains why clear property rights alone do not solve the externality problem. When transaction costs are low, a bad assignment gets corrected by bargaining, as the rancher and the farmer showed. When transaction costs are high, no bargaining will fix a bad assignment, and the resources stay where the law put them. The legal system therefore matters in two ways. First, it determines the level of transaction costs through the clarity of rights and the cost of enforcing them. Second, where bargaining cannot work at all, the initial assignment becomes the final allocation, so courts and legislatures should try to assign rights where they do the least damage.

Spain provides a striking example. The Crown gave the Mesta, the guild of migratory shepherds, legal rights of passage and pasture across the farmland of Castile. It was the rancher and the farmer again, now at the scale of a kingdom. Millions of farmers and herders could not bargain with one another, so the allocation of land was determined by the initial assignment of rights. Economists still debate whether those privileges built Spain’s advantage in wool or held back its agriculture (Drelichman, 2009). But no one debates that the law, good or bad, governed the use of the land the whole time.

The implication is not that Pigou’s logic was simply incorrect. In a world without transaction costs, the tax is unnecessary: the parties bargain their way to the efficient herd. In the world with transaction costs, where the tax is actually needed, the tax must be judged like everything else, by comparing its costs with its benefits.

Coase gives a concrete example. A factory’s smoke does $100 of damage a year. The factory can eliminate the smoke with a device that costs $90 a year, while the neighbors can avoid the harm by adjusting on their side for $40 a year. The cheapest solution is the neighbors’ $40 one. But under a $100 tax, the factory buys the $90 device, since $90 beats $100, and the neighbors, who no longer suffer any smoke, do nothing. Society spends $90 where $40 would have done, because the tax puts all the pressure on the factory and none on the neighbors.

The tax also demands a lot from the government. Setting it correctly means measuring the full loss the smoke imposes on everyone affected, information Coase was “unable to imagine” anyone collecting. Running a tax system is itself expensive. The government must administer it, and the taxes it raises and the spending it finances distort other decisions in the economy, so removing one loss creates others. None of this makes taxes useless. It makes them one institutional arrangement among several, alongside private bargains, firms, direct regulation, and, sometimes, doing nothing. The government’s “administrative machine,” Coase warned, “is not itself costless.” “All solutions have costs,” he concluded, and the right question is which arrangement produces the most value in the actual world, not in a frictionless one.

The problem becomes more complex when many parties are involved. Coase’s examples have two neighbors, and later work showed that even costless bargaining can fail with three or more. Whatever two of them agree to pay the third, the third can always offer one of them a better private deal, and the three-way agreement collapses (Aivazian and Callen, 1981). In practice, the failure is even more pronounced: a polluted river involves a large number of parties, and the cost of bringing them to one table is enormous. Those are precisely the cases where bargaining breaks down and institutional design—the laws and organizations that define and enforce rights—does all the work.

Three extensions

Three lines follow from Coase’s central idea.

Three important implications follow from Coase’s central idea. The first is Coase’s own earlier paper, The Nature of the Firm (1937), written when he was in his twenties. If markets coordinate the economy as well as the textbooks say, why is so much production organized inside firms, where a manager gives orders and no prices are used? Firms are, in a phrase Coase took from D. H. Robertson, “islands of conscious power in this ocean of unconscious co-operation.” Coase’s answer is that using the market is costly. “There is a cost of using the price mechanism”: you must find out what things cost and write a contract for every deal. A firm replaces many small bargains, whose enforcement is not costless, with one standing arrangement. A firm stops growing when doing one more task inside the organization costs as much as buying it from the market or, in Coase’s words, when “the costs of organising an extra transaction within the firm are equal to the costs involved in carrying out the transaction in the open market.” That condition still tells economists where the boundary of the firm lies: what a company does itself and what it buys from others. The 1991 Nobel Prize cited the 1937 and 1960 papers together.

The second extension is political economy. Daron Acemoglu (2003) asked why societies do not bargain their way to efficient institutions the way Coase’s neighbors bargain to the efficient herd. His answer is commitment. “Underlying the Coase theorem is the ability to write enforceable contracts.” In Acemoglu’s words, those who hold power “cannot commit to not using their power to renege on their promises,” so the deals that would deliver better institutions cannot be trusted and are never made. In Coase’s terms, the costs of enforcing those deals are prohibitive. Inefficient institutions survive because the groups that would have to give up power cannot be promised compensation for doing so.

The third extension is international. Coase’s insight applies whenever rights are difficult to define and enforce. This is true not only for private disputes but also for political and international ones. Within a country, the state acts as a third-party enforcer of rights. Internationally, there is no sovereign above states, and great powers often settle questions of property and security through power rather than law. Institutional economics largely assumes states that enforce contracts at home and a rules-based international order abroad. Yet that international order often lacks a de jure enforcer. That missing enforcer is one of the field’s blind spots.

Coase in the age of AI

The costs of transacting are precisely what AI is now reducing. Drafting a contract, monitoring compliance: each is a task these systems already perform. As the cost of using the market falls, the boundary of the firm shifts, and work that was kept inside a company because contracting it out was too costly can move to the market.

Lower transaction costs do not, however, resolve the question of who owns the rights. When a self-driving car injures a pedestrian, or a model is trained on disputed data, the question is the one Coase posed in 1960: the harm is reciprocal, and the task is to assign a clear right and enforce it. This is the argument I made in Let Coase Drive Us Home: what will limit the adoption of AI is not the technology itself but the institutions that govern it—the definition and enforcement of rights over data, decisions, and liability. Nor do the commitment problem in politics or the absence of an enforcer between nations disappear simply because technology improves. The bottleneck for AI will be institutional, not computational.

The worst way to read Coase today is to ask an AI to explain the Coase Theorem. The answer will be the familiar frictionless interpretation, the very reading Coase spent thirty years correcting. The better way is the one this essay has illustrated: build the example, work the bargain under both rules, add a transaction cost, add a third player. AI can help, but only as a sparring partner. That is the habit of mind this series is trying to cultivate.

Exercises

1. Reconstruction (no AI). Explain, in your own words, why Coase thinks “how do we restrain the rancher?” is the wrong question. What does the reciprocal nature of the problem mean, and how does it change what the law should aim for?

2. Invariance by hand (no AI). Using Table 1 and a $300 steer, show that the herd is three under both liability rules, and work out who pays whom in each. State in one sentence what is invariant across the two regimes and what changes.

3. Compare the arrangements (AI). Give an AI the smoke example: $100 of annual damage, a $90 prevention device for the factory, and a $40 adjustment available to the neighbors. Ask it to compare a $100 tax on the factory, a clear assignment of rights with costless bargaining, and costly bargaining. Which arrangement minimizes social cost, and why does the tax choose the wrong solution in this case?

4. Break the bargaining (AI). Have an AI build the simplest example of three parties who would all gain from an agreement, but where every proposed deal can be undone by a side deal between two of them. Then ask what this implies for externalities that involve millions of parties.

5. The political theorem (AI). Ask an AI why societies do not simply adopt efficient institutions, then rebut its answer using Acemoglu’s commitment problem: a bargain no one can be forced to keep is not a bargain.

6. Coase on a machine (AI). Choose one current issue—liability when a self-driving car injures a pedestrian or ownership of a model’s training data—and analyze it as Coase would. Identify the reciprocal harm, the parties, the right in dispute, the transaction costs, and where enforcement breaks down.

References

Daron Acemoglu (2003). “Why Not a Political Coase Theorem? Social Conflict, Commitment, and Politics.” Journal of Comparative Economics 31(4): 620–652.

Varouj A. Aivazian and Jeffrey L. Callen (1981). “The Coase Theorem and the Empty Core.” Journal of Law and Economics 24(1): 175–181.

Ronald H. Coase (1937). “The Nature of the Firm.” Economica 4(16): 386–405.

Ronald H. Coase (1960). “The Problem of Social Cost.” Journal of Law and Economics 3: 1–44.

Ronald H. Coase (1988). The Firm, the Market, and the Law. University of Chicago Press.

Mauricio Drelichman (2009). “License to Till: The Privileges of the Spanish Mesta as a Case of Second-Best Institutions.” Explorations in Economic History 46(2): 220–240.

Sebastian Galiani and Gustavo Torrens (2025). “The Economic Approach to Geopolitics.” Journal of Economic Behavior & Organization 237.

Steven G. Medema (2020). “The Coase Theorem at Sixty.” Journal of Economic Literature 58(4): 1045–1128.

Arthur C. Pigou (1920). The Economics of Welfare. Macmillan.

George J. Stigler (1988). Memoirs of an Unregulated Economist. Basic Books.

Sebastian Galiani and Raul A. Sosa (forthcoming). AI at Your Side: The Student’s Guide to Smarter Learning. Oxford University Press.

***If this essay resonated with you, you can subscribe to my Substack for regular reflections on economics, politics, geopolitics, AI, and literature.

*** Disclaimer: I used ChatGPT-5.5. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.

Most people answer that it happens to you. You did not choose your genes, the town you grew up in, the driver who ran the red light, or the virus going around the office. Health can feel like weather, something that arrives and has to be dealt with.

In 1972 Michael Grossman, one of the founders of health economics, wrote down a model with a different and slightly uncomfortable answer. To a large degree, he argued, you produce your own health. You begin adult life with a stock of it, it wears down a little each year, and every year you decide how much to spend keeping it up. The checkup you scheduled or skipped, the hour at the gym or on the couch, the sleep you protected or spent, the vaccine you took or did not: all of these are inputs, and health is what they produce. Seen this way, health is closer to capital than to weather. It is a stock you maintain, like a house or a skill, and the interesting question stops being what happened to you and becomes how much of it you choose to hold.

That idea launched the modern economics of health capital and, like most foundational contributions, it has since hardened into a single diagram in most textbooks (a notable exception is Health Economics by Bhattacharya et al.). Students learn to draw it, reproduce it under exam pressure, and move on. Almost none of them can say where it comes from or what it is really made of.

We take that memorized diagram and rebuild it from the decisions underneath, one at a time, so that by the end the finished curve cannot possibly look as though it fell from the sky. You will have watched it assembled out of your own choices, which is the only version of the model worth carrying around.

This is the third piece in a series that reads one classic paper at a time with AI at your side, in the spirit of the book I wrote with Raul A. Sosa, AI at Your Side: The Student’s Guide to Smarter Learning, forthcoming from Oxford University Press. Here the method earns its keep. Ask a chatbot to explain the Grossman model and it will hand you the last page: the diagram, fluent and confident and of no real use for understanding. A summary of a model you cannot yet derive is only the slogan at higher resolution. Used the other way, as something you build alongside and argue with, the same tool becomes the best sparring partner a student has ever had. I will flag the moments where reaching for it pays off.

Health is capital

Everything that follows rests on one reframe. In the model, health is a stock, and Grossman has it doing three jobs at once. You value it directly, because feeling well beats feeling sick, so it enters your utility, the economist’s term for how well off you are. It is also an input, because a healthy body produces something scarce and valuable, namely time that is not lost to illness. And it is a durable asset, one that lasts, wears down, and can be built back up. Most goods do one of these jobs. Health does all three, which is what makes the model powerful and what makes it easy to get lost in. By the end all three will be pressed onto a single curve, the standing bargain economists make with dimensions to keep a diagram on the page.

The central move in the paper is to separate two things that everyday language runs together: health and health care. Nobody enjoys a doctor’s visit. Nobody wants a colonoscopy, a blood draw, or a statin for its own sake. What you want is health, and medical care is one of the inputs you buy to produce it. The demand for care is a derived demand, in the same way the demand for flour is derived from the demand for bread. That is the first screw to come loose. What looked like the good you were buying is only an input, and the thing you actually care about is something you make yourself.

You do not buy health, you produce it

If health is produced, there must be a production function for it, a recipe that turns inputs into output, and the model writes one down. Each period you add to your health, in what the model calls gross investment, by combining medical care and other market goods, call them M, with your own time, Tₕ. Grossman writes this as I = I(M, Tₕ; E). The E at the end is your education, and it sits slightly apart from the other inputs because it governs how efficiently you turn those inputs into health, a point we will come back to.

Time is where the model bites. You have a fixed number of hours in a year, exactly 8,760 of them, which is 365 days times 24 hours. Every hour goes into one of four uses, working for money, doing the things you enjoy, tending to your health, or lying sick. That last use is pure waste. Sick time is taken away from everything else, and the only way to reduce it is to be healthier. So health repays you in a concrete currency, the hours it keeps you from losing.

This is also the point where the money constraint and the time constraint stop being separate. You buy medical care out of income, income comes from the hours you work, and the hours you can work depend on how many of them illness takes from you. Improve your health and the whole system loosens at once.

One feature of that technology does most of the work later, so let me state it plainly now. When you are very sick, one more unit of health buys back many hours. When you are already well, it buys back almost none. So the extra healthy time you get from one more unit of health, a quantity the model calls G, shrinks as your health grows. Hold on to that fact. It is what will make the curve slope downward once we draw it.

The decision, period by period

Every period, then, you face a choice, and it is an optimization like any other. Given the prices in front of you, your wage w, the price of medical care, the interest rate r, you decide how to divide your time and your money. Two separate decisions live inside that choice, and both matter for what comes next.

The first is about cost. To produce any given amount of health, you can lean on medical care or lean on your own time. A busy surgeon buys the trainer and the concierge doctor. A retiree with time to spare and less money walks an hour a day and cooks from scratch. You settle on the mix where the last dollar spent on care and the last dollar’s worth of your own time buy the same amount of health. That mix is the cheapest way to produce a unit of health, and we can call its price π. Notice that π already has your wage and the price of care folded into it, and already reflects a decision you made.

The second is about quantity. How much health should you hold? To answer it you set what a unit of health gives you against what it costs to produce. A unit of health returns the value of the extra healthy time it creates, which is your wage times the extra hours G, so the benefit is w·G. Set that against π, the cost of the unit, and the whole model is about to fold into a single line. For now we price only this investment payoff, the healthy time a unit of health reclaims, and set aside the direct pleasure of feeling well that was the first of its three jobs. Economists call this the pure investment version of the model, and Exercise 4 puts the missing piece back.

The curve is the compression

Divide the benefit by the cost, and you have the rate of return on a unit of health:

rate of return = (w · G(H)) / π

Economists give this rate a name, the marginal efficiency of health capital, or MEC. Read the pieces slowly. w is the value of an hour of your time. G(H) is the extra healthy time one more unit of health provides. π is what it costs to produce that unit. Plot this rate against your health stock H, and you get the curve in Figure 1.

Figure 1. The MEC curve. At each level of health H, it gives the return on one more unit of health.

Now look closely at what sits inside that fraction, because this is where the argument lives. Every term in it has a history. w is what your labor market already pays for an hour of your time. π is the cost-minimized price of health, which came out of the first decision in the previous section. G(H) is the marginal product of your health stock, the extra healthy time one more unit of it buys back at the level of health you are holding. π is the output of the optimization you just ran, w is the price that optimization takes as given, and G(H) is the technology they act on. So every point on the curve answers a conditional question. If your health stock were H, what would one more unit of it return, given that you had already chosen your medical care, your own time, your hours of work, and everything else as well as you could at today’s prices? Ask that question at every level of H, and the answers, taken together, are the curve you are looking at.

So the curve is no primitive. It is the compressed record of your own optimizing, one height for every possible level of health, each height already assuming you got everything else right.

Why the curve slopes down? Of them, w and π do not depend on your health stock at all. Only G(H) does, and G shrinks as H grows. The entire downward tilt of the curve comes from one plain fact: the sicker you are, the more a unit of health gives back, and the healthier you are, the less. That is all “diminishing returns” has ever meant here.

Where your health settles

That was the demand side. The cost side is shorter.

Holding health is not free, and two things charge you rent on it. First, the money you sink into producing health could have earned the market interest rate r somewhere else, so every unit you hold carries an opportunity cost of r. Second, health depreciates. It wears out at a rate we call δ, and faster as you age, so you have to keep investing simply to stay level. Add the two and the cost of holding a unit of health for a period is r + δ. That is a flat line, the same at every level of health, because it is only the price of tying your wealth up in your own body rather than somewhere else.

Set the return equal to that cost, and the model is finished:

(w · G(H)) / π = r + δ

In plain terms, you invest in health until the return on the last unit equals the cost of holding it. Below that level, health pays more than it charges, so you buy more. Above it, health charges more than it pays, so you let it slide. The level of health where the two are equal is the one you choose, and it finally has a name you understand from the inside: H*.

Only now does the diagram earn its place. The downward curve is your own optimizing, compressed. The flat line is the cost of capital. Where they cross is the amount of health that a person with this wage, this efficiency, and this rate of decay will rationally decide to hold.

Figure 2. The optimal health stock. The downward curve is the MEC from Figure 1. The flat line, r + δ, is the cost of holding health. You hold H*, the level where the two cross.

This is the first natural place to bring in AI, and how you do it decides whether it helps at all. Asking it what the MEC curve is returns a paragraph you could have found anywhere. Handing it your own derivation and telling it to find the leak is a different exercise.

Here is my attempt to get from the two first-order conditions to the condition w·G/π = r + δ. Check each step. Tell me the first place my algebra or my economics goes wrong, and why it matters.

What comes back is not a lecture on Grossman but a finger on the exact line where you confused a stock with a flow or dropped a term, and that is the only kind of correction that teaches you anything.

Reading the world with the curve

The reward for all this machinery is that real patterns become easy to read. Once the diagram is two curves that can move, every fact about health becomes a question of which curve moved and why.

Start with education, the model’s most famous prediction. Educated people, Grossman argues, are more efficient producers of health. The same hour of exercise, the same prescription, the same piece of advice about what actually helps all convert into more health in the hands of someone who knows more. That makes each unit of health cheaper to produce, so π falls and the return w·G/π rises at every level of health. The curve shifts up and out, and the optimal stock rises with it. The educated hold more health, which is intuitive enough. The surprising part is what the model then says about medical care. Because they produce health more efficiently, the educated can reach that higher health with fewer inputs rather than more. The model’s headline prediction is that they are both healthier and spend less on medicine, and it holds unless that added efficiency leads them to want so much more health that they buy more inputs after all. The health half of that prediction is among the most robust findings in all of health economics. A result that looks like a paradox falls straight out of a curve that shifted.

The wage works through the same curve but in its own way. A raise makes an hour of your time more valuable, so a day lost to illness costs you more and health is worth more to you. A raise also lifts π, since your own time is one of the inputs into producing health, but the value of the healthy time you gain rises faster, so the return still climbs and you hold more health. The one contrast with education is on the spending side. Because your time has become expensive, you produce that extra health by buying more medical care and using less of your own time. Education and higher wages both raise health, yet they likely push medical spending in opposite directions, and the model tells you exactly why.

Age is the last of the three. As the years pass the depreciation rate δ climbs, health wears out faster, the flat cost line rises, and the optimal stock falls year after year. The model even writes its own ending. Once the health you would rationally choose to hold falls below the minimum a body needs to keep functioning, a level the model calls Hₘᵢₙ, life stops. Death here is not bolted on from the outside but a consequence of the very same optimization that governed everything before it.

Figure 3. The same picture under two shifts. Left: education makes you a more efficient producer of health, pushing the MEC curve outward, so the optimal stock rises from a lower H* to a higher one. Right: as you age your health depreciates faster, lifting the cost line r + δ, so the optimal stock walks down toward Hₘᵢₙ.

The life cycle is a good place for a second round with AI, because it is exactly where intuition tends to fail. Ask what happens to medical spending as a person ages, and then make the model argue against your first guess. The obvious one is that spending falls, since the optimal stock is falling too. The model says the opposite can hold. As depreciation accelerates, people often spend more on their health even as the stock itself declines, because they are fighting a stronger current just to slow the fall. A good prompt makes AI defend that counterintuitive result and lay out the condition under which it holds, so that you come away with the mechanism instead of the wrong hunch.

The right and wrong way to do this with AI

Look back at what just happened. We took a diagram that students memorize, pulled it back into the decisions it stands for, and somewhere along the way it stopped being a picture to reproduce and became something you could reason with.

It could easily have gone the other way. Ask a chatbot to explain the Grossman model and you get the finished diagram and three neat bullet points about what shifts what, correct and fluent and hollow, the last page again with nothing behind it. What made the difference was using the tool in the opposite spirit. You derived the equilibrium condition and had it catch the step you botched. You asked why the curve slopes down and refused to accept “diminishing returns” as a magic phrase until you could say why one more unit of health does less and less as you grow healthier. You predicted a pattern over the life cycle and let it prove you wrong. The model did not become clear because AI explained it to you. It became clear because AI was the thing you argued with while you explained it to yourself.

That is the whole argument of the book I wrote, and health capital may be the perfect case for it, precisely because the payoff is a single, memorable diagram. A memorable diagram is the easiest thing in economics to carry around without understanding. Taking it apart is the difference between having seen the Grossman model and being able to think with it.

So, back to the question we started with. Is health something that happens to you, or something you choose? The honest answer is both, and the model is unusually precise about the split. You do not choose your genes, your accidents, or the rate at which you depreciate. But given all of that, at every age, you are choosing how much health to hold, and the famous curve is nothing more than the running ledger of those choices. Health is not quite weather. It is a decision made under constraints you did not pick, which is true of almost everything worth understanding.

Final comment

Every great model leaves something out. Grossman’s does too. It largely treats individuals as isolated decision makers, paying relatively little attention to social interactions, behavioral biases, or uncertainty. Later work expanded the model in all of these directions. But none of those extensions make sense until you first understand the original.

Exercises

1. Derive the condition w·G/π = r + δ yourself, starting from the idea that the return on a unit of health should equal the return on any other investment. Then hand your derivation to an AI and ask it to find the first step that is wrong or hand-wavy, and why that step matters. Do not ask it to derive the condition for you.

2. The model predicts that more educated people hold more health but buy less medical care. Ask an AI to construct the assumption under which that prediction reverses. Then decide for yourself whether that assumption is realistic.

3. Take a real decision you made this month: skipping sleep in exam week, buying or skipping the gym membership, walking instead of driving. Name the health stock, the investment, the depreciation, and the return in healthy time. Where on your own MEC curve were you sitting?

4. In the version we built we priced health only through the useful time it produces, setting aside the direct pleasure of feeling well that we opened with. Ask an AI to put that consumption value back in, and have it show you which conclusions survive and which ones, like the clean predictions for education and the wage, get muddied once feeling well is something you value directly.

References:

Bhattacharya, J., Hyde, T., & Tu, P. (2014). Health Economics. Palgrave Macmillan.

Grossman, M. (1972). On the Concept of Health Capital and the Demand for Health. Journal of Political Economy, 80(2), 223–255.

Galiani, S., & Sosa, R. A. AI at Your Side: The Student’s Guide to Smarter Learning. Forthcoming. Oxford University Press.

***If this essay resonated with you, you can subscribe to my Substack for regular reflections on economics, politics, geopolitics, AI, and literature.

*** Disclaimer: I used ChatGPT-5.5. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.

Milton Friedman, borrowing from Leonard Read’s 1958 essay I, Pencil, used to hold one up and make a simple point: nobody knows how to make it. Not really. The wood, the graphite, the rubber, the metal — each part comes from a different place, through chains of production no single mind could reconstruct. Thousands of people cooperate to produce it, and no one is in charge.

That simple observation is the starting point of Friedrich Hayek’s 1945 paper, The Use of Knowledge in Society. It is one of the most cited and least understood papers in economics.

The slogan is familiar: knowledge is dispersed, prices transmit information, markets coordinate.

The problem is that slogans are cheap. Understanding is not.

Hayek’s real move is subtler and more uncomfortable. Instead of solving the problem of allocating resources given “the data,” he questions the premise that such data exist in any centralized, usable form. The relevant knowledge in an economy is fragmented, local, often tacit, and constantly changing. It lives in the heads of individuals, embedded in context, frequently not even articulable. In Hayek’s own words, it exists “solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.”

That is the problem.

Once you see it that way, prices are no longer just a clearing device. They are a communication system. Their function is not to tell us everything, but to tell us enough.

A coffee shop owner does not need to know whether a drought hit Brazil or a port shut down. She only needs to know that coffee has become more expensive. That signal is sufficient for her to adjust. Multiply that across thousands of decisions, and coordination emerges, not because anyone understands the whole, but because no one needs to.

This is Hayek’s claim: markets work not because they are efficient in some abstract sense because they use knowledge that cannot be centralized.

At this point, the standard comparison with central planning becomes almost trivial. The issue is not whether planners are intelligent. It is whether they can access, aggregate, and update the relevant knowledge fast enough. In many cases, the deeper question is whether that knowledge can be collected at all.

And here is where most readings of Hayek go wrong.

Hayek’s actual claim is more subtle than “markets are perfect.” Any alternative institution must be judged by its ability to use dispersed knowledge.

You see the force of this immediately when prices are prevented from doing their job.

Take rent control. The intention is straightforward: make housing more affordable. But the instrument interferes with the signal. A rising rent is not just a burden — it is information. It tells us that demand exceeds supply.

Suppress the price, and you suppress the message.

The empirical literature is remarkably consistent on this. Rebecca Diamond, Tim McQuade, and Franklin Qian (2019) show that San Francisco’s rent control expansion led landlords to withdraw a significant fraction of rental units, ultimately pushing rents up. David Autor, Christopher Palmer, and Parag Pathak (2014) find that removing rent control in Cambridge increased property values broadly, including for units never controlled. Edward Glaeser and Erzo Luttmer (2003) show that rent control generates massive misallocation: people occupy the wrong apartments in the wrong places.

When the price system is muted, coordination breaks down.

None of this implies that distribution does not matter. It implies that distribution should be addressed with instruments that do not destroy the informational role of prices. Transfers, not price controls. One instrument per objective.

Now, the interesting question, and the reason to revisit Hayek today, is whether AI changes any of this.

There is a tempting view: Hayek’s argument was a product of limited data and weak computation. Today, we collect information at massive scale and process it in real time. Perhaps the problem of dispersed knowledge has simply shrunk.

Hayek’s problem was never just computational. It was epistemic. It concerns what information exists in the first place, and in what form.

Much of the relevant knowledge in an economy is generated only in action. It is local, contextual, and often fleeting. A shopkeeper noticing a shift in demand, an engineer improvising a substitution, a worker adjusting effort in response to incentives: these are not always recorded, and when they are, they are often recorded too late.

AI can aggregate observed data extraordinarily well. What it cannot do, at least not in any general way, is replace the process through which much of that knowledge is created.

So the boundary shifts. In some domains, coordination can be more centralized than before. In others, Hayek’s argument becomes even stronger. What emerges is a hybrid economy, with different coordination mechanisms operating in parallel and the boundary between centralization and decentralization shifting case by case.

That is the real update.

There is also a deeper misunderstanding worth clearing up.

Students often misread Hayek’s argument as requiring prices to be “right.” All it requires is that prices convey information.

The stronger version of the claim goes one step further, and this is where the argument becomes non-trivial. Prices do not need to be general equilibrium prices to generate coordination. They only need to move in the right direction and be locally informative.

Think of the economy as a process, not a solution. At any point in time, prices are “wrong” in the sense that they do not clear all markets simultaneously. But they still contain signals about relative scarcity. When a price rises, even if it overshoots, it pushes behavior in the correct direction: it induces substitution away from the good, encourages supply responses, and triggers search for alternatives. These adjustments do not require that the price be exactly right, only that it not be completely disconnected from underlying conditions.

In that sense, prices work like gradients rather than solutions. They guide decentralized adjustments without requiring global consistency at every moment. In this Hayekian sense, efficiency is the property of a process that uses dispersed information to move the system in the right direction over time.

This is exactly the same logic behind the efficient-markets intuition in finance. Efficiency does not mean prices are always correct. It means that there is no systematic free arbitrage. Prices can be volatile, even wrong in levels, and still perform their informational role.

It also clarifies why suppressing price movements is so damaging. We never had a perfectly correct signal. What we lose, when prices are suppressed, is the direction of adjustment itself.

So where does this leave us?

Hayek’s insight is not that markets are optimal. It is that coordination in a complex economy depends on institutions that can make use of dispersed knowledge. Prices happen to be extraordinarily effective at doing this.

The problem persists. AI becomes one more participant in the system that handles it.

How to Read This Series

This is the first of a series of thirty papers to be read alongside the book AI at Your Side: The Student’s Guide to Smarter Learning, by Sebastian Galiani and Raul A. Sosa, forthcoming at Oxford University Press. The goal is to read each one slowly and carefully, in the way these classics deserve to be read, rather than rush through a syllabus.

Each paper will be approached in the same way: as something to work through, step by step.

The worst way to read Hayek today is to ask an AI for a summary. You will get a polished version of something you already knew, and mistake the fluency for understanding.

The only way to understand a paper like this is to reconstruct the argument yourself, apply it, and push against it.

AI can help, but only if it is used as a sparring partner.

That is the discipline this series is trying to build.

Exercises

1. Reconstruction (no AI).
   In your own words, explain why Hayek thinks the “data” of the economy do not exist in a centralized form. Be precise. Avoid slogans.

2. Prices without equilibrium.
   Consider a market where demand suddenly increases and prices rise sharply, overshooting the eventual equilibrium.
   What information is still conveyed by the price change?
   What adjustments does it trigger?
   In what sense can this still be “efficient” in Hayek’s sense?

3. Application.
   Take both policies: rent control and housing vouchers.
   Using Hayek’s framework, describe what information each policy preserves or destroys, and predict the behavioral responses.

4. AI challenge.
   Ask an AI: “Does large-scale data and machine learning invalidate Hayek’s argument about dispersed knowledge?”
   Then write a one-paragraph rebuttal to its answer.

References

David H. Autor, Christopher J. Palmer, and Parag A. Pathak (2014). “Housing Market Spillovers: Evidence from the End of Rent Control in Cambridge, Massachusetts.” Journal of Political Economy 122(3): 661–717.

Rebecca Diamond, Tim McQuade, and Franklin Qian (2019). “The Effects of Rent Control Expansion on Tenants, Landlords, and Inequality: Evidence from San Francisco.” American Economic Review 109(9): 3365–3394.

Milton Friedman (1980). “The Power of the Market.” Free to Choose, Episode 1. PBS.

Edward L. Glaeser and Erzo F. P. Luttmer (2003). “The Misallocation of Housing under Rent Control.” American Economic Review 93(4): 1027–1046.

Friedrich A. Hayek (1945). “The Use of Knowledge in Society.” American Economic Review 35(4): 519–530.

Leonard E. Read (1958). “I, Pencil: My Family Tree as Told to Leonard E. Read.” The Freeman, December 1958.

***If this essay resonated with you, you can subscribe to my Substack for regular reflections on economics, politics, geopolitics, AI, and literature.

*** Disclaimer: I used ChatGPT-5.2. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.

Most students encounter this model early in their training. Many can write down the full-income constraint on an exam. Far fewer can explain what it really means. And fewer still can translate the first-order conditions into a story about why wealthier societies eat more fast food, why dual-income families outsource childcare, or why Americans are, as Becker once put it, simultaneously “wasteful” of goods and “economical” with time. The algebra is not the hard part. The hard part is building the bridge between the math and the economics.

Building that bridge is one of the goals of the book I am writing with Raul A. Sosa, AI at Your Side: The Student’s Guide to Smarter Learning, forthcoming from Oxford University Press. The book explores how AI can be used not to solve problems for students but to help them reason through models step by step, explaining each result, modifying assumptions, and defending conclusions on their own. One of the chapters develops this idea through formal economic models, and Becker’s framework is a central example.

In this post, I revisit Becker’s paper, show what learning it with AI looks like in practice, and argue that this approach can change how we teach formal economics.

The Model

Recall the setup. In traditional consumer theory, households maximize utility over market goods subject to a budget constraint. Becker added a simple but powerful observation: consumption takes time, and that time could have been spent earning money.

Instead of consuming goods directly, households produce what Becker called “commodities,” things like a home-cooked meal, a night at the theatre, or a commute to work. Each one requires both market goods xᵢ and time Tᵢ:

Zᵢ = fᵢ(xᵢ, Tᵢ)

The household maximizes utility U = U(Z₁, …, Zₘ) subject to two constraints. A goods constraint:

∑ pᵢxᵢ = V + T_w w

where pᵢ are goods prices, V is non-labor income, T_w is working time, and w is the wage rate. And a time constraint:

∑ Tᵢ = T − T_w

where T is total available time. But these two constraints are not independent. Time can be converted into money by working more. Substituting one into the other, and writing inputs in per-unit terms (bᵢ for goods per unit of Zᵢ, tᵢ for time per unit), yields a single full-income constraint:

∑ (pᵢbᵢ + wtᵢ) Zᵢ = V + wT

This is where the elegance of the framework becomes clear. Each commodity now has a full price, πᵢ = pᵢbᵢ + wtᵢ: the cost of the goods it requires plus the foregone earnings from the time it absorbs. And the right-hand side, S = V + wT, is full income: the value of non-labor income plus what the household could earn if every available hour were allocated to market work.

The first-order conditions follow naturally:

∂U / ∂Zᵢ = λπᵢ for all i

At the optimum, the marginal utility per dollar of full price is equalized across all commodities. It is the standard equal-bang-for-the-buck rule, but with a redefined “buck” that includes time.

And this reframing has real bite. A rise in wages does not just increase income. It changes the relative price of everything you do. Time-intensive activities become relatively more expensive. Goods-intensive ones become relatively cheaper. The substitution effect pushes consumption toward goods-intensive commodities. The income effect, since full income has risen, could push in either direction depending on the income elasticity of time-intensive activities. The net outcome is empirical, but the framework tells you exactly what forces are at play and why.

From this single structure, Becker derived predictions about hours of work, the secular decline in the work week, cross-sectional income elasticities, the demand for quality, transportation choices, the division of labor within families, and even the economics of queuing. The paper’s lasting influence comes not from any one application but from the framework itself: a unified way to think about every use of time.

What Students Miss

The algebra is well within reach of an undergraduate course. Where students struggle is interpretation. They memorize πᵢ = pᵢbᵢ + wtᵢ without internalizing that a home-cooked meal and a delivered meal can have very different full prices even when the grocery bill is the same. They write down the first-order condition without seeing that it explains why a busy lawyer hires a housekeeper while a graduate student cleans the apartment. They follow the comparative statics of a wage increase without clearly distinguishing the substitution effect from the income effect, and without recognizing that these two forces can push in opposite directions for the same activity.

The gap is between manipulating the model and thinking with it. That is where AI comes in.

Learning Becker with AI

In our book, we demonstrate several techniques for using AI as a learning partner with this model, each targeting a different skill: checking your own mathematical setup, interpreting equilibrium conditions in plain language, explaining a result and getting corrected, translating algebra into diagrams, and stress-testing assumptions. Let me show two of them here.

Making the first-order condition crystal clear

Consider a student who has correctly derived MUᵢ / πᵢ = λ but struggles to say what it means in words. Here is a prompt we use in the book:

Prompt: “I solved the first-order conditions for Becker’s model and got MUᵢ / πᵢ = λ for each commodity. I understand the math but not the economics. Explain what this condition means as if I am choosing between cooking dinner (time-intensive, cheap in money) and ordering delivery (expensive in money, saves time). Specifically: what is πᵢ for each option, what does equality across activities mean for my behavior, and what is λ in a sentence I could say out loud?”

The response (generated using Claude):

Start with πᵢ, the full price. Cooking costs little in market goods, say $8 in groceries, but takes an hour. If your wage is $20/hour, the full price of cooking is $8 + $20 = $28. Ordering delivery might cost $22 but takes only 5 minutes: the full price is $22 + $20 × (5/60) ≈ $23.67. So delivery, which looks more expensive in dollar terms, is actually cheaper once you count time.

The condition MUᵢ / πᵢ = λ says: at your best choice, the satisfaction you get from spending one more dollar of full price must be the same across all activities. If cooking gives you more satisfaction per dollar-of-full-price than ordering, you should cook more (and order less) until the marginal utilities adjust and the ratios equalize. It is the same logic as the standard equal-bang-for-the-buck rule, but now “buck” includes time.

And λ? In one sentence: λ is how much happier you would be if you had one more dollar of full income, meaning one extra dollar or the equivalent in freed-up minutes.

The student chose the context (cooking versus delivery). The AI supplied concrete numbers that reveal a surprise: the option that looks expensive in money is actually cheaper in full-price terms. That inversion is the core of Becker’s insight, and it becomes vivid in a way that the abstract condition alone never achieves.

Figure 1. The full price of dinner. Cooking looks cheaper in market prices ($8 vs. $22) but is more expensive in full-price terms ($28 vs. $23.67) once foregone earnings are included.

Explaining the model back and getting corrected

The second technique draws on what psychologists call the protégé effect: you learn more deeply when you explain something to someone else (Chase et al., 2009). Here the “someone else” is the AI, instructed to catch mistakes.

Prompt: “I am going to explain to you what happens in Becker’s model when the wage rate increases. Correct me if I get anything wrong and tell me what I am missing.

Here is my explanation: When w goes up, all activities become more expensive because time is more valuable. So the consumer substitutes toward activities that use less time and more market goods, like eating out instead of cooking, taking an Uber instead of walking, or buying pre-recorded lectures instead of attending long seminars. That is why richer people buy more convenience services. The model predicts that economic growth should shift consumption toward goods-intensive, time-saving alternatives.”

The response:

Good overall. Your substitution effect story is correct and your examples are well chosen. Two things to refine.

First, a correction: you said “all activities become more expensive.” That is true in absolute terms, but what drives substitution is relative prices. Time-intensive activities become more expensive relative to goods-intensive ones. A meal that takes an hour rose in full price by much more than an Uber ride that takes 10 minutes. It is the change in relative full prices, not the level, that triggers substitution.

Second, an omission: you described only the substitution effect. But the wage increase also raises full income (wT + V), which creates an income effect. If some time-intensive activities are normal goods (say, a long home-cooked meal with family), the income effect pulls you toward them even as the substitution effect pushes you away. The net result is ambiguous for these goods. Your story holds cleanly for activities where both effects point the same direction, but it is incomplete for activities that people value more as they become richer. This is why very wealthy people sometimes have personal chefs who cook elaborate meals. Goods expenditure replaces their own time, but the time-intensive experience is a luxury they can afford precisely because of the income effect.

The student’s explanation was mostly right. But the AI identified a subtle imprecision (relative versus absolute prices, exactly the distinction Becker emphasized) and surfaced a missing piece (the income effect) with a concrete example that makes it memorable. No generic textbook summary could produce this kind of feedback. The correction is targeted to what the student actually said.

Figure 2. A wage increase rotates and shifts the full-income budget line. The consumer moves from A to B: less of the time-intensive commodity (cooking), more of the goods-intensive one (delivery). The slope steepens because the full price of cooking rose more than the full price of delivery.

Why This Matters

What these examples illustrate is a different way of learning formal economics. The student does the thinking. The AI sharpens it. Instead of receiving a pre-packaged explanation, the student attempts an interpretation or formulates an argument, and the AI responds to what was actually said rather than delivering a generic lecture.

Becker’s paper works particularly well for this approach because the difficulty is not algebraic. The math is within reach of any undergraduate. What is hard is making the leap from equations to economic reasoning: the kind of understanding that lets you modify an assumption, apply the model to a new setting, or recognize it at work when you choose between cooking and ordering dinner tonight. That is exactly the kind of understanding that a well-structured conversation with AI can help develop.

In our book, we apply the same method to other models and other fields. The specific mathematics changes, but the learning approach is the same: bring your own work to the conversation, be specific about what you know and what you do not, and treat every AI response as the beginning of a dialogue rather than the end of one.

If you teach economics, you have probably seen students copy derivations without understanding them, or paste models into a chatbot and submit whatever comes back. Neither approach builds understanding. There is a better way, and it starts with treating AI as a thinking partner that meets students at their specific point of confusion. That is what the book is about.

Becker’s great insight was not simply that time matters. It was that once time has a price, almost every ordinary choice looks different. The household is no longer just choosing how to spend money. It is choosing how to spend a scarce life. That is why this model travels so far: from meals to commuting, from childcare to leisure, from the demand for convenience to the demand for children. And that is also why it is such a powerful model to learn with AI. Once a student sees how the logic works in one case, the real learning begins: not repeating Becker’s equations, but using them to think through new ones.

Exercises

Once students see the logic of the model, the natural next step is to push it beyond meals, commuting, or convenience and ask what it implies for some of the most important choices people make.

First: if raising children is relatively intensive in time, what should happen to the demand for children when other goods become much cheaper relative to the wage?

Second: why did Becker believe that the scarcest resource in modern economies is not goods but time?

References

Becker, G. S. (1965). A theory of the allocation of time. The Economic Journal, 75(299), 493–517.

Chase, C. C., Chin, D. B., Oppezzo, M. A., and Schwartz, D. L. (2009). Teachable agents and the protégé effect: Increasing the effort towards learning. Journal of Science Education and Technology, 18(4), 334–352.

Galiani, S. and Sosa, R. A. (forthcoming). AI at Your Side: The Student’s Guide to Smarter Learning. Oxford University Press.

***If this essay resonated with you, you can subscribe to my Substack for regular reflections on economics, politics, geopolitics, AI, and literature.

*** Disclaimer: I used ChatGPT-5.2. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.