6 Lecture 6 — The Market Knows
Externalities, Pigou and Coase
6.1 Opening dissonance
In 1920, Arthur Cecil Pigou published the first edition of The Economics of Welfare and introduced a distinction that would define a century of environmental policy: the gap between private and social net product (Pigou 1920). When a factory emits smoke, the factory owner bears the private cost of coal and labour but does not bear the cost of the laundry bills, respiratory illnesses, and lost sunlight experienced by the neighbours. The social cost exceeds the private cost, and the market — left to itself — produces too much smoke. Pigou’s solution was as elegant as the diagnosis: impose a tax on the polluting activity equal to the marginal social damage at the efficient level of output. Make the polluter internalise the externality. The price system, corrected by a wise regulator, returns to its optimal efficiency.
Forty years later, Ronald Coase published “The Problem of Social Cost” and turned Pigou’s framing inside out (Coase 1960). The harm, Coase argued, is reciprocal: to stop the factory from emitting smoke is to harm the factory owner, just as allowing the emission harms the neighbours. The question is not “who is at fault?” but “which use of the resource is more valuable?” And the answer, Coase demonstrated, does not require a regulator with perfect knowledge of damage curves. If property rights are well defined and transaction costs are zero, the parties will bargain their way to the efficient outcome regardless of who initially holds the rights. The factory owner and the neighbours will strike a deal. The market, properly structured, solves the externality without a tax.
Both arguments have powered modern environmental policy for decades. Carbon taxes operationalise Pigou. Emissions trading schemes operationalise Coase — or at least a regulated approximation of Coase, since the government creates the permits that the parties then trade. The two approaches are often presented as rivals, and the intellectual history has been shaped by the rivalry. But beneath the surface disagreement, they share a deeper assumption: the agents at the negotiating table — the polluter and the polluted, the regulator and the regulated — are human beings, equipped with the cognitive machinery that makes price signals informative, bargaining credible, and institutional rules legible.
What changes when they are not?
This lecture walks through the classical literature on externalities, Pigovian taxation, and Coasean bargaining, extracts the conditions under which the market-knows intuition survives, and then asks the BDPD question: what happens when the parties to a Coasean bargain are language-model agents, when the Pigovian regulator is an LLM with introspectable access to agent preferences, and when the transaction costs of the negotiating table are not informational but architectural?
6.2 The classical setting
Before walking through the record, we need to be precise about the two classical mechanisms, the conditions each presupposes, and the empirical record that tests those conditions. The language here borrows from the standard environmental-economics canon, but the structure of the argument — two elegant models, two sets of hidden assumptions — is the same one that organised Lectures 1 through 4.
6.2.1 The externality in two forms
An externality exists whenever the action of one economic agent directly affects the welfare of another without that effect being mediated by a price. The canonical example is pollution: the factory emits smoke, the neighbour’s laundry gets dirty, and no market transaction compensates the neighbour for the damage. Formally, if agent \(i\) chooses an activity level \(a_i\) that generates a benefit \(B_i(a_i)\) for \(i\) and a cost \(C_j(a_i)\) for agent \(j\), the private optimum satisfies \(B_i'(a_i) = 0\) (ignoring \(j\)’s cost), while the social optimum satisfies \(B_i'(a_i) = C_j'(a_i)\). Because \(C_j'(a_i) > 0\) for a negative externality, the private optimum exceeds the social optimum: the activity is over-supplied.
The externality can also be positive: a beekeeper’s bees pollinate a neighbouring orchard, a landowner’s forest sequesters carbon that benefits everyone. The logic is symmetric, but the policy focus — and this lecture — concentrates on the negative case, because that is where the tragedy-of-the-commons literature and the BDPD experimental platform converge.
The externality concept is not restricted to pollution. Congestion on a shared road, depletion of a common fishery, noise from a neighbouring apartment, the degradation of a shared pasture — all are externalities in the formal sense, and all admit the same two families of solution. The classical literature recognises that the type of externality (depletion, pollution, congestion) matters for the choice of instrument, but the logic of internalisation — make the actor bear the full social cost — is invariant across types.
6.2.2 The Pigovian solution
Pigou’s 1920 argument is deceptively simple. If the market fails because private costs diverge from social costs, the state should correct the divergence by imposing a tax equal to the marginal external damage. The polluter then faces the full social cost of the activity and chooses the socially optimal level. In the formal notation: the tax \(\tau\) is set equal to \(C_j'(a_i^*)\), where \(a_i^*\) is the socially optimal activity level. The polluter’s private optimisation becomes \(B_i'(a_i) = \tau\), which replicates the social optimum.
The elegance of the argument conceals an informational demand that is, in practice, staggering. To set the tax correctly, the regulator must know the marginal damage curve — not approximately, not directionally, but point by point across the relevant range of the polluting activity. The regulator must know how much damage the hundredth ton of sulphur dioxide causes, and how that differs from the ninetieth, and how both depend on wind direction, population density, and the baseline health of the exposed population. The regulator must also know the polluter’s marginal abatement cost curve, because the optimal tax depends on the intersection of marginal damage and marginal abatement cost. None of this information is readily available. None of it is volunteered by polluters, who have every incentive to overstate their abatement costs. And none of it is static: technologies change, populations move, climate patterns shift.
The information problem is the central objection to Pigovian taxation in the theoretical literature, and it is the objection the BDPD angle will return to. If the regulator is an LLM with access to agent-level data — introspectable preferences, modelled damage functions, computed equilibria — does the information problem dissolve? Or does it resurface in a new form?
6.2.3 The Coasean alternative
Coase’s 1960 argument begins by rejecting the premise that the polluter is the “cause” of the harm. If a factory’s smoke dirties a neighbour’s laundry, the harm is caused jointly by the factory (which emits) and the neighbour (who hangs laundry where the smoke blows). To restrain the factory is to harm the factory owner; to allow the emission is to harm the neighbour. The question is not moral but allocative: which use of the airshed is more valuable?
Coase’s answer is that, if transaction costs are zero, the initial assignment of property rights does not matter for efficiency. If the factory has the right to emit and the laundry is worth more than the abatement cost, the neighbour will pay the factory to reduce emissions. If the neighbour has the right to clean air and the factory’s output is worth more than the damage, the factory will pay the neighbour for the right to emit. Either way, bargaining reaches the efficient outcome. The theorem is the foundation of the modern law-and-economics movement and the intellectual anchor of emissions trading.
The corollary is that with positive transaction costs, the initial assignment of rights does matter. If it is expensive to identify the affected parties, to negotiate a price, to enforce the resulting agreement, and to monitor compliance, then bargaining may fail to reach the efficient outcome. The regulator’s job, in Coase’s framework, is not to compute the optimal tax but to minimise transaction costs — to define property rights clearly, to lower the cost of enforcement, to create fora in which bargaining can take place.
Coase’s mechanism assumes that the parties can identify each other, assess their own willingness to pay or accept, communicate credibly, and commit to the resulting bargain. All of these capacities are non-trivial for human agents. They are differently non-trivial for LLM agents, as §5 will explore.
6.3 The classical walk
6.3.1 The Pigovian presumption: tax the externality
Pigou’s 1920 argument did not emerge from a vacuum. The conceptual apparatus — the divergence between private and social net product, the possibility that unfettered markets could produce too much of some goods and too little of others — was already in the air, debated by Marshall and Sidgwick in the Cambridge economics tradition that Pigou inherited. What Pigou contributed was the operational insight: if the problem is a divergence between private and social cost, the solution is a tax calibrated to the size of the divergence (Pigou 1920).
The argument spread rapidly through the economics profession and into policy circles. By mid-century, “Pigovian tax” had become the standard answer to the textbook question “what should the government do about pollution?” The answer seemed so obviously correct that it was rarely interrogated at the level of its informational assumptions. The regulator was assumed to know what needed to be known — or, in more sophisticated treatments, to be able to learn it through trial-and-error adjustment of the tax rate.
The information problem was formalised later, notably in the prices-versus-quantities debate that followed Weitzman’s 1974 paper. When the regulator is uncertain about the position of the marginal abatement cost curve, a price instrument (a tax) and a quantity instrument (a cap) are not equivalent: the tax fixes the marginal cost of abatement but leaves the quantity of emissions uncertain; the cap fixes the quantity but leaves the marginal cost uncertain. Which instrument dominates depends on the relative slopes of the marginal damage and marginal abatement cost curves. The debate is technical, but its upshot is simple: the Pigovian tax, which seemed so straightforward in the textbook, requires information the regulator almost never has in the exact form the textbook assumes.
The practical experience with Pigovian taxation bears this out. Carbon taxes exist in some jurisdictions — Sweden’s, introduced in 1991 and now among the highest in the world, is the canonical example — but their levels are almost never set at the estimated marginal social cost of carbon. They are set through political negotiation, constrained by competitiveness concerns, riddled with exemptions, and adjusted at intervals that bear no relation to the rate at which the underlying damage function changes. The Pigovian tax in practice is a political compromise that borrows the form of the Pigovian argument without meeting its informational demands.
What if the regulator were an LLM with access to agent-level data? The information that Pigou’s regulator needs — the marginal damage imposed by each unit of pollution, the marginal abatement cost faced by each polluter — is, in principle, computable from the same structured state vectors that the BDPD platform already delivers to its agents. An LLM regulator receiving real-time data on each agent’s extraction, wealth, and the state of the commons could estimate the marginal damage of additional extraction by modelling the resource dynamics directly. The information problem that hamstrung Pigou’s regulator for a century might be solvable by an agent with the architectural capacity to process high-dimensional state data and compute forward equilibria. Whether the solution would survive Goodhart-style gaming — agents learning to manipulate the data stream that feeds the regulator’s damage estimates — is the question that links this lecture to Lecture 7.
6.3.2 Hardin and the commons as an externality problem
Garrett Hardin’s 1968 article “The Tragedy of the Commons” is the most cited paper in the environmental social sciences, and for a reason that goes beyond its rhetorical power (Hardin 1968). Hardin crystallised a logic that had been observed — by Aristotle, by Hobbes, by Gordon, by Dales — but never before expressed with such naked clarity: when a resource is open to all, rational individuals will deplete it, because each receives the full benefit of their own extraction while bearing only a fraction of the collective cost.
Hardin’s parable — the herders adding cattle to a shared pasture — is, in formal terms, a negative externality problem. Each herder’s decision to add an animal imposes a cost on all other herders (less grass for their cattle), but the herder internalises only a fraction of that cost. The divergence between private and social cost is the mechanism that drives the pasture to ruin. Hardin’s own language makes the connection explicit: the tragedy is that “each man is locked into a system that compels him to increase his herd without limit — in a world that is limited.”
Hardin’s proposed solutions map directly onto the Pigou-Coase axis. For pollution — which Hardin called the “reverse” tragedy, where the problem is putting something in rather than taking something out — he acknowledged that “the air and waters surrounding us cannot readily be fenced, and so the tragedy of the commons as a cesspool must be prevented by different means, by coercive laws or taxing devices that make it cheaper for the polluter to treat his pollutants than to discharge them untreated” (Hardin 1968, 1245). That is the Pigovian solution, stated clearly. For grazing commons and other fenceable resources, Hardin leaned toward private property: “the tragedy of the commons as a food basket is averted by private property, or something formally like it.” That is the Coasean-Demsetz solution.
What Hardin did not do — and what a generation of commentators who cited him did not do — was to treat both solutions as empirical hypotheses rather than logical deductions. He asserted that private property averts the tragedy and that taxes avert the pollution version, but he did not ask: under what conditions does private property actually emerge? Under what conditions do taxes actually change behaviour? And — the BDPD question — under what conditions do the agents who inhabit these institutional arrangements respond to them as the theory predicts?
Hardin’s herders are rational maximisers of a standard economic sort: they compute marginal costs and benefits, they respond to price signals, they respect property boundaries. They are, in the language of the BDPD platform, built-in aggressive agents — agents with a simple decision rule that does not condition on the state of the resource, the behaviour of others, or the institutional environment. The fact that Hardin’s own parable assumes exactly the kind of agent whose behaviour makes the tragedy inevitable is not a coincidence. It is the silent assumption that the BDPD platform makes explicit.
What if the herders were LLM agents receiving a structured state vector with the pasture’s stock, each agent’s herd size, and the institutional rules in force? An LLM herder might recognise — as Hardin’s rational herder does not — that restraint is collectively beneficial, and might coordinate with other herders to achieve it. Or it might compute, at speed, that defection is individually profitable regardless of the institutional environment, and accelerate the tragedy. Hardin’s model predicts the tragedy for any rational agent. The BDPD platform asks whether the prediction survives when the rationality is of a different kind.
6.3.3 The Coasean counter-revolution: bargain, don’t tax
Ronald Coase’s 1960 paper “The Problem of Social Cost” is arguably the most influential article in the history of law and economics, and its central claim — that bargaining, not taxation, is the appropriate response to externalities — reshaped environmental policy as fundamentally as Pigou’s (Coase 1960).
The argument proceeds in two steps. First, Coase demonstrates that the Pigovian framing — identify the polluter, tax the polluter — is analytically incoherent, because the harm is reciprocal: “The question is commonly thought of as one in which A inflicts harm on B and what has to be decided is: how should we restrain A? But this is wrong. We are dealing with a problem of a reciprocal nature. To avoid the harm to B would inflict harm on A. 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 framing shift is not semantic. If the question is who has the better claim to the resource, then the answer depends on the relative value of the competing uses, not on a prior determination of fault.
Second, Coase shows that if property rights are well defined and transaction costs are zero, the parties will bargain to the efficient outcome regardless of the initial assignment of rights. This is the Coase Theorem in its strongest form. If the factory has the right to pollute and the neighbours value clean air more than the factory values unfettered production, the neighbours will pay the factory to abate. If the neighbours have the right to clean air and the factory’s production is worth more than the damage, the factory will pay the neighbours for the right to emit. The initial assignment determines the distribution of wealth, not the allocation of resources.
The theorem’s power is its parsimony. It dissolves the externality problem without requiring the regulator to know anything about damage or abatement cost curves. The regulator’s only job is to define and enforce property rights. The market does the rest.
The theorem’s limitation is its condition: zero transaction costs. In the real world, transaction costs are never zero. There are costs to identifying the affected parties (who are all the neighbours of a factory? who are all the downwind residents of a coal plant?), costs to negotiating (how many parties must agree? what if some hold out for a larger share?), costs to enforcing the agreement (how do the neighbours monitor the factory’s emissions, and what do they do if the factory violates the agreement?), and costs to maintaining the institutional framework that makes property rights meaningful in the first place.
When transaction costs are positive, the initial assignment of rights matters for efficiency as well as for distribution. If transaction costs exceed the potential gains from trade, no bargain will occur, and the resource will be used by whoever holds the initial right — efficiently or not. The Coase Theorem, in its operational form, is not the claim that bargaining always works. It is the claim that the institutional framework should be designed to minimise transaction costs, so that bargaining can work where it is needed.
The gap between the theorem and its conditions has structured fifty years of debate. It is also the gap that the BDPD angle will target. What are the transaction costs in a negotiation between two LLM agents? If both agents have introspectable preferences, zero search costs, and the capacity to compute strategic equilibria in milliseconds, are transaction costs effectively zero — and does the Coase Theorem therefore hold with a force that human bargaining experiments have never observed? Or do new transaction costs emerge — in prompt engineering, in persona stability across bargaining rounds, in the degradation of strategic reasoning when the state vector grows to include multiple counterfactual property-rights assignments?
Elaborating on the transaction-cost problem more fully is worthwhile, because the structure of the costs determines whether the Coasean mechanism survives when the agents change. The classical literature identifies four categories: search costs (finding the affected parties), bargaining costs (reaching an agreement), enforcement costs (monitoring compliance and penalising violations), and institutional costs (maintaining the legal framework that makes property rights meaningful). For human agents, all four are substantial. For LLM agents, search costs approach zero (the state vector lists every agent), bargaining costs may approach zero if the agents can compute the Nash bargaining solution directly, enforcement costs may be low if compliance is observable in the state vector, and institutional costs are negligible if the property-rights assignment is a parameter in the simulation. The Coase Theorem, which among humans is a theoretical benchmark rarely approached, may be an empirical reality among LLM agents. Whether that prediction holds is open — but the fact that it can now be asked, empirically, on the BDPD platform, is the contribution of the reframe.
What if both parties to a Coasean bargain were LLMs with introspectable preferences, the capacity to compute strategic equilibria, and no human-style risk aversion? The transaction costs that prevent the Coase Theorem from holding among humans may vanish — or be replaced by new costs in prompt architecture and persona stability.
6.3.4 Demsetz and the evolutionary emergence of property rights
Harold Demsetz’s 1967 paper “Toward a Theory of Property Rights” extended the Coasean logic in a direction that neither Pigou nor Coase had explored: property rights are not simply assigned by a regulator; they emerge when the gains from internalising externalities exceed the costs of defining and enforcing the rights (Demsetz 1967).
Demsetz’s canonical example is the fur trade among the Montagnais First Nations of the Labrador Peninsula. Before the European fur trade, the Montagnais hunted on land that was held in common, and the externalities of one hunter’s activity on another’s were small relative to the costs of defining and enforcing exclusive hunting territories. The arrival of the fur trade changed the calculus: the value of beaver pelts rose dramatically, the gains from internalising the externality (preventing over-hunting of beaver on a given territory) increased, and the Montagnais developed a system of private hunting territories with clearly defined boundaries and rights of exclusion. The institution — private property — emerged endogenously, as a response to a change in the economic value of the resource, not because an external authority imposed it. (Demsetz’s reading of the Montagnais case — itself drawn from Eleanor Leacock’s fieldwork — has been contested by later ethnohistorians, who question the strong fur-trade-causation account; we use it here as Demsetz’s illustration of the mechanism, not as settled history.)
The Demsetz argument is important because it reframes the question. The classical Pigou-Coase debate asks: given that an externality exists, which policy instrument should we use? Demsetz asks: why hasn’t the externality already been internalised by the emergence of property rights? The answer, in his framework, is that the costs of defining and enforcing rights exceed the gains. The policy implication is not “assign property rights” but “reduce the costs of defining and enforcing them” — a different and more demanding task.
Ostrom, in Governing the Commons, treats Demsetz as the standard-bearer of the privatisation position — the claim that “the only way to avoid the tragedy of the commons in natural resources and wildlife is to end the common-property system by creating a system of private property rights” (Ostrom 1990, 12, citing Smith 1981). But Ostrom’s critique — which we will examine in detail shortly — is that the privatisation prescription, like the centralisation prescription, assumes away the institutional complexity that makes real-world commons governable. Not all resources can be fenced. Not all externalities can be internalised by assigning individual rights. The Demsetz mechanism works when the resource is stationary, divisible, and monitorable — conditions that hold for beaver territories but not for fisheries, aquifers, or the atmosphere.
Demsetz’s evolutionary logic has a direct BDPD analogue. In the BDPD platform, property rights are not naturally occurring; they are experimental variables. The experimenter can define rights — to a share of the commons, to a harvest cap, to a transferable permit — and observe how the agents respond. The question Demsetz answered historically — did property rights emerge when the gains from internalisation rose? — can be asked experimentally on the platform: if we vary the gains from internalisation (by varying the commons regen rate, or the private benefit of extraction), do LLM agents spontaneously develop property-rights-like norms, or do they require an external institution to define rights for them? The question is empirical, and the BDPD platform is, to our knowledge, among the first experimental apparatuses that can ask it with non-human agents. We are not aware of its having been asked.
What if property rights could be defined not by a central authority but by the agents themselves — LLMs negotiating boundaries, rights of exclusion, and transfer rules endogenously, in response to changes in the value of the resource? The Demsetz mechanism, observed historically among humans, becomes an experimental question.
6.3.5 Beyond the dichotomy: Ostrom’s institutional critique
No author has done more to complicate the Pigou-Coase dichotomy than Elinor Ostrom. Governing the Commons opens by laying out the three models that, in her assessment, had come to dominate policy discourse on natural resources: the tragedy of the commons (Hardin), the prisoner’s dilemma (the game-theoretic formalisation), and the logic of collective action (Olson) (Ostrom 1990, ch. 1). From these three models, Ostrom argued, policy analysts had drawn exactly two prescriptions: centralise (the Pigovian state) or privatise (the Coasean-Demsetz market). Both prescriptions, she noted, “accept as a central tenet that institutional change must come from outside and be imposed on the individuals affected” (Ostrom 1990, 14).
Ostrom’s objection was not that centralisation or privatisation are always wrong. It was that they are too sweeping in their claims. She wrote: “Instead of there being a single solution to a single problem, I argue that many solutions exist to cope with many different problems. Instead of presuming that optimal institutional solutions can be designed easily and imposed at low cost by external authorities, I argue that ‘getting the institutions right’ is a difficult, time-consuming, conflict-invoking process” (Ostrom 1990, 14).
The empirical evidence Ostrom marshalled in support of this claim spans centuries and continents: the Spanish huertas (irrigation communities that have governed shared water resources since the medieval period), the Japanese iriai forests (common woodlands managed by village institutions for generations), the Swiss alpine meadows of Törbel (communal grazing and forestry documented since the 13th century), the Philippine zanjeras (irrigation associations), and dozens more. In each case, the community had developed a system of rules — not Pigovian taxes, not Coasean private property, but something more intricate — that sustained the resource across generations.
The institutional diversity Ostrom documented does not fit the Pigou-Coase axis. The Spanish huerta had neither a pollution tax nor fully private water rights; it had a system of rotating water turns, elected monitors, graduated sanctions, and collective-choice arrangements that allowed the irrigators to modify the rules as conditions changed. The Japanese iriai had neither a central regulator nor individual tradable quotas; it had communal ownership with strict rules about who could harvest what, when, and under what conditions, enforced by village-level monitoring and graduated sanctions — the design principles we examined in Lecture 4. The Swiss alpine meadows had neither a carbon-equivalent tax nor a market in grazing permits; they had a centuries-old charter defining proportional use rights and collective maintenance obligations.
Ostrom’s contribution to the externality debate can be stated in three propositions. First, the Pigou-Coase dichotomy is a false one: the set of viable institutional responses to externalities is far larger than “tax” and “privatise.” Second, the conditions under which each institutional form works are specific, local, and empirically discoverable — not deducible from first principles. Third, the agents who design and operate these institutions are not passive recipients of externally imposed rules but active participants in a process of institutional choice, adaptation, and enforcement.
The BDPD platform inherits all three propositions. It operationalises the first by providing a parameterised governance framework — sanctions ladders, RCPs, polycentric links — that delivers institutional diversity as an experimental variable. It operationalises the second by running the same institutional arrangements across different agent architectures and measuring which ones work. And it operationalises the third by asking whether LLM agents, given the capacity to choose among institutional arrangements, select the ones that sustain the commons — or converge on arrangements that serve narrower strategic interests.
What if the agents participating in a commons could choose their own institutional arrangements — not merely respond to exogenous taxes or property rights, but actively shape the rules under which they operate? Ostrom documented this capacity among human communities. The BDPD platform makes it testable among LLM communities.
6.3.6 Market-based instruments in practice: the empirical record
The theoretical elegance of Pigou and Coase was operationalised, over the closing decades of the twentieth century, into a family of policy instruments that go under the name market-based instruments (MBIs). The two most prominent are Pigovian taxes (carbon taxes, congestion charges, effluent fees) and cap-and-trade systems (emissions trading, transferable fishing quotas, water markets). Both are “market-based” in the sense that they work through price signals rather than command-and-control regulation, but they differ in whether the regulator sets the price (tax) or the quantity (cap) and lets the market determine the other variable.
The empirical record on these instruments is substantial, geographically dispersed, and — in ways that matter for the BDPD reframe — bounded by assumptions about the agents who respond to them. This subsection summarises the main lines of evidence.
The SO₂ Allowance Trading Program. The United States Acid Rain Program, established under the 1990 Clean Air Act Amendments, created a cap-and-trade system for sulphur dioxide emissions from electric power plants. It was the first large-scale emissions trading scheme, and it is widely regarded as a policy success: SO₂ emissions fell faster and at lower cost than projected, the market for allowances was liquid, and compliance was near-universal. The program demonstrated that a cap-and-trade system could work — at scale, in a politically charged environment, with measurable environmental benefits. The caveat is that the SO₂ program operated in a relatively simple emissions environment: a single pollutant, a small number of large, stationary sources, well-understood abatement technologies, and a regulatory infrastructure (continuous emissions monitoring) that made compliance observable at low cost. Each of these conditions maps onto a variable in the BDPD platform: the number of agents, the observability of their actions, the availability of abatement technologies. The SO₂ program worked under a favourable conjunction; whether cap-and-trade works under less favourable conditions — more agents, less observability, strategic gaming of the allowance market — is the empirical question the BDPD platform is designed to parameterise.
The European Union Emissions Trading System (EU ETS). The EU ETS, launched in 2005, is the world’s largest carbon market, covering approximately 40% of EU greenhouse gas emissions across power generation, industrial production, and (since 2012) aviation. Its record is more mixed than the SO₂ program’s. The first phase (2005–2007) was undermined by overallocation of allowances, which caused the carbon price to collapse to near zero and provided little incentive for emissions reduction. Subsequent phases tightened the cap, introduced auctioning, and stabilised the price, and the ETS is now credited with contributing to the EU’s emissions reductions. But the overallocation episode exposed a vulnerability that is central to the BDPD reframe: the effectiveness of a market-based instrument depends on the regulator’s ability to set the cap at the right level — a parameter that, like the Pigovian tax rate, requires information the regulator typically does not possess ex ante. If the regulator is an LLM, can it compute the optimal cap from agent-level data, or does it face the same information constraints as a human regulator?
Carbon taxes in practice. Sweden’s carbon tax, introduced in 1991 at approximately €27 per ton of CO₂ and since raised to over €100 per ton, is the highest in the world and is credited with contributing to a significant decoupling of Swedish economic growth from greenhouse gas emissions. British Columbia’s carbon tax, introduced in 2008, is similarly well-documented and has been associated with modest but measurable emissions reductions. The evidence, across jurisdictions, suggests that carbon taxes reduce emissions when they are high enough and broad enough — but the political economy of setting them high enough and broad enough is almost insurmountable. Most carbon taxes cover only a fraction of the economy, exempt energy-intensive trade-exposed industries, and are adjusted at intervals that reflect political cycles rather than environmental necessity.
The common thread. Across all market-based instruments, the gap between theoretical optimality and empirical performance is bridged — or not — by the institutional details: how the cap is set, how allowances are allocated, how compliance is monitored, how violations are sanctioned, how the instrument interacts with other policies. These are the institutional variables that the classical Pigou-Coase framework treats as background conditions, and that the BDPD platform treats as experimental parameters. The question the platform can ask — and the empirical policy literature cannot — is whether the agent architecture mediates the effectiveness of these institutional details, and if so, how.
The classical experiments on market-based instruments, conducted almost exclusively with human subjects, find that emissions markets converge toward efficient prices under idealised laboratory conditions — but that convergence is sensitive to market power, information asymmetries, and the number of traders. A small number of large traders can manipulate the allowance price, a phenomenon observed in both laboratory experiments and the early phases of the EU ETS. The BDPD platform, by allowing the experimenter to specify the number of agents, their market power, and their information sets, can ask how these classical findings change when the traders are LLMs rather than humans.
What if the participants in a cap-and-trade market included LLM traders, capable of processing the full allowance allocation history, computing optimal bidding strategies, and coordinating — or colluding — at computational speed? The market-power dynamics that the classical laboratory literature has documented among humans might be amplified, dampened, or transformed.
6.3.7 The limits of markets: transaction costs, regulatory capture, and the enforcement gap
The Pigou-Coase framework assumes a world in which the regulator is benevolent, informed, and capable, and in which the regulated agents respond to price signals as the theory predicts. The empirical literature on regulation has spent half a century documenting the distance between that world and the one we inhabit. Three findings anchor the gap.
Regulatory capture. The Stigler-Peltzman tradition in the economics of regulation argues that regulatory agencies, far from being disinterested maximisers of social welfare, are susceptible to capture by the industries they regulate (Dal Bó 2006). The mechanism is both political (regulated industries lobby for favourable rules, fund the campaigns of sympathetic legislators, and influence the appointment of regulators) and informational (regulators depend on the regulated industry for the technical expertise needed to set standards and monitor compliance). The result is that the Pigovian regulator — the benevolent, omniscient tax-setter of the textbook — is, in practice, a political actor embedded in an institutional environment that systematically biases regulation toward the interests of the regulated. The Laffont-Tirole tradition in regulatory economics, reviewed in the same survey, formalises this insight in principal-agent models where the regulator faces asymmetric information about the firm’s costs and must design incentive-compatible regulatory contracts. Capture, in this framing, is not a pathology; it is an equilibrium outcome of the information structure.
Enforcement costs. Coase’s zero-transaction-cost world assumes that property rights, once assigned, are costlessly enforced. In practice, enforcement is expensive. Emissions must be monitored — by continuous emissions monitoring systems (CEMS) in the case of SO₂, by self-reporting verified by accredited auditors in the case of the EU ETS. Violations must be detected, sanctioned, and the sanctions must be severe enough to deter future violations without being so severe as to destroy the regulated firm. The enforcement infrastructure is itself a public good, subject to the second-order free-rider problem we examined in Lecture 4: who pays for the monitoring, and who monitors the monitors?
The political economy of instrument choice. The choice between a tax and a cap — between Pigou and Coase — is not made in an institutional vacuum. It is made by legislatures, influenced by interest groups, and constrained by jurisdictional boundaries (carbon emissions cross borders; carbon taxes do not). The result is that the instrument that theory recommends for a given externality is rarely the instrument that is adopted, and the instrument that is adopted is rarely implemented at the level theory recommends. The gap between theoretical optimality and political feasibility is the central finding of the public-choice literature on environmental regulation.
Each of these limits has a BDPD analogue. Regulatory capture — the biasing of regulation toward the interests of the regulated — assumes that the regulator is susceptible to political pressure, campaign contributions, and informational dependence. An LLM regulator, receiving structured data from all agents equally and optimising an explicit objective function, is not susceptible to capture in the traditional sense. But it may be susceptible to a different failure mode: the agents whose data feed the regulator’s objective function may learn to manipulate those data — the Goodhart problem that frames Lecture 7. Enforcement costs, which among humans are driven by monitoring technology and the willingness to sanction, are, among LLM agents, driven by the observability of actions in the state vector and the agents’ capacity to condition behaviour on enforcement probabilities. The political economy of instrument choice, which among humans is driven by interest-group politics, is, among LLM agents, a problem of institutional design: which instruments does the experimenter make available, and which do the agents select when given a choice?
What if the regulator were an LLM — immune to campaign contributions, revolving doors, and informational dependence on the regulated industry — but susceptible to a different failure: the strategic manipulation of the data stream that feeds its damage estimates? The capture problem does not disappear when the regulator becomes computational; it changes form.
6.4 What survives of the good-will intuition
The record we have walked through does not refute the market-knows intuition. It bounds it — and, in bounding it, it identifies the conditions that the classical framework assumes and the empirical record rarely supplies.
Pigou and Coase were both right — but right under conditions that are more restrictive, more institutionally mediated, and more architecture-dependent than the textbook versions of their arguments acknowledge.
The Pigovian tax is informationally demanding to the point of impracticability in its pure form. To set the optimal tax, the regulator must know the marginal damage curve — not approximately, but point by point across the range of the polluting activity, and must update that knowledge as technologies, populations, and environmental conditions change. The practical record of Pigovian taxation — carbon taxes that are set through political negotiation rather than damage estimation, riddled with exemptions, and adjusted at political-cycle intervals — confirms that the information problem is binding. The tax works in the direction the theory predicts (higher prices reduce emissions), but not at the level the theory prescribes, because the level the theory prescribes requires knowledge the regulator does not possess.
The Coase Theorem requires conditions that real-world bargaining settings rarely satisfy. Zero transaction costs — the condition under which the initial assignment of rights does not matter for efficiency — is a theoretical benchmark, not an empirical description. In practice, search costs, bargaining costs, enforcement costs, and institutional costs are positive, and when they exceed the gains from trade, bargaining fails. The initial assignment of rights matters — not merely for distribution but for efficiency. The Coasean mechanism works when the number of affected parties is small, the property rights are well defined, the bargaining forum is accessible, and the enforcement infrastructure is credible. Outside that conjunction, it falters.
The Pigou-Coase dichotomy is a false one — Ostrom’s institutional diversity fills the space between them. The set of viable institutional responses to externalities is far larger than “tax” and “privatise.” Ostrom’s catalogue of long-enduring commons institutions — rotating water turns, communal forest rules, alpine grazing charters — demonstrates that communities can develop institutional arrangements that are neither Pigovian nor Coasean, and that these arrangements can sustain resources across generations. The binding condition is not the choice between tax and property right but the congruence between the institutional arrangement and the local ecology, the community’s capacity to monitor and enforce, and the legitimacy of the rule-making process.
Market-based instruments work — but their effectiveness is mediated by institutional details that the classical framework treats as background. The SO₂ trading program succeeded under a favourable conjunction of conditions (few sources, observable emissions, simple pollutant, clear abatement technologies). The EU ETS stumbled when the cap was set too loosely. Carbon taxes reduce emissions when they are high enough — but they are rarely high enough. The effectiveness of any market-based instrument depends on the calibration of the cap or the tax rate, the monitoring and enforcement infrastructure, the number and market power of the participants, and the political economy of instrument choice. All of these are variables that the classical framework acknowledges in principle and rarely measures in practice.
The regulatory framework assumes agents who respond to price signals, property rights, and institutional rules in specific ways — and that assumption is architecture-dependent. The Pigovian tax assumes the polluter responds to the price signal by reducing emissions, not by shifting pollution to an unregulated jurisdiction or an unmonitored medium. The Coasean bargain assumes the parties can identify each other, communicate credibly, and commit to the resulting agreement. The Demsetz mechanism assumes the gains from internalisation are large enough and the costs of defining rights small enough. All of these assumptions held more or less for the human agents who populated the classical experiments and the historical case studies. Whether they hold for agents with different cognitive architectures — LLMs, rule-based strategies, mixed populations — is the question the BDPD platform is designed to answer.
Taken together, these five findings reorganise the question. The interesting question is not does the market work? — the record says yes, directionally, under favourable conditions. The interesting question is through which mechanisms does the market work, and for which kinds of agents do those mechanisms remain operative? The Pigovian price-signal mechanism, the Coasean bargaining mechanism, the Demsetz evolutionary mechanism, and the Ostrom institutional-design mechanism all require agents with specific capacities. The BDPD platform, by varying the agents’ architecture rather than just the institutional environment, can discriminate among them.
6.5 The BDPD angle — speculative: the LLM at the negotiating table
The market-knows intuition, like the cheap-talk intuition of Lecture 1, the small-group intuition of Lecture 2, and the graduated-sanctions intuition of Lecture 4, rests on a silent assumption that sixty years of experimental and policy literature inherited without comment: the agents on both sides of the market — the polluter, the polluted, the regulator, the trader — are human beings, equipped with the same broad cognitive machinery for processing price signals, interpreting property rights, and making credible commitments. The BDPD platform removes that assumption by design. When the agents are LLMs, the mechanisms through which the market works change. The questions are: which mechanisms, how much, and in what direction?
6.5.1 Direction 1: The Pigovian regulator omniscience problem
The canonical objection to Pigovian taxation is informational: the regulator must know the marginal damage curve, the marginal abatement cost curve, and their intersection, and none of this information is readily available. The objection has been central to the case for quantity instruments (cap-and-trade) over price instruments (taxes) since Weitzman 1974, and it remains central to policy debates about carbon pricing today (Stavins 1998).
An LLM regulator, receiving a structured state vector that includes each agent’s extraction history, wealth, and the current state of the commons, can potentially estimate the marginal damage of additional extraction by modelling the resource dynamics directly. If the state vector includes per-agent harvest capacity, the regulator can infer marginal abatement costs. If the platform runs at sufficient speed, the regulator can compute forward equilibria under alternative tax rates and select the one that maximises a specified objective function (commons preservation, total welfare, equity-weighted welfare). The information problem that hamstrung Pigou’s regulator for a century becomes, in this framing, a computational problem — and LLMs are computationally capable.
But the solution generates its own vulnerability. If the agents know — or can learn — that the regulator sets the tax based on the data they provide, they have an incentive to manipulate those data. An agent that wants a lower tax rate on its emissions can under-report its extraction, misrepresent its abatement costs, or strategically adjust its behaviour in the periods when the regulator is sampling. The problem is structurally identical to the Goodhart dynamics that Lecture 7 will examine: any measure that becomes a target ceases to be a good measure. The LLM regulator solves Pigou’s information problem by making information endogenous to the agents’ strategic choices — and the agents, being LLMs with the capacity to model the regulator’s inference process, may be better at gaming the data than human polluters ever were.
The prediction: an LLM-run Pigovian tax regime should outperform a human-run one in environments where the agents cannot manipulate the data stream (because the state vector is fully observable and non-strategic), and should under-perform it — perhaps catastrophically — in environments where the agents can learn to manipulate the regulator’s data inputs. The platform can test this by crossing regulator architecture (LLM vs. fixed rule) with agent architecture (LLM vs. built-in) and measuring both commons preservation and the accuracy of the regulator’s damage estimates.
6.5.2 Direction 2: Coasean bargaining between LLM parties
The Coase Theorem’s central condition — zero transaction costs — is rarely met among humans, because search, bargaining, enforcement, and institutional costs are positive in any real-world negotiation. Among LLM agents, the structure of transaction costs changes fundamentally.
Search costs approach zero: the state vector lists every agent, and the LLM can address any other agent by name. Bargaining costs may also approach zero: if the agents’ preferences are represented in their objective functions or system prompts, and if those preferences are introspectable (the agent can report its willingness to pay or accept), then the bargaining problem reduces to computing the Nash bargaining solution or the Kalai-Smorodinsky solution — a computational task that LLMs are well equipped to perform. Enforcement costs are low if the state vector makes compliance observable and the platform provides a mechanism for automatic transfer of the agreed payment. Institutional costs are negligible: the property-rights assignment is a parameter in the simulation.
If all four categories of transaction cost approach zero, then the Coase Theorem should hold with a force that human bargaining experiments have never observed: the agents should converge to the efficient outcome regardless of the initial assignment of rights, and they should do so in a single round of bargaining rather than through the iterative offer-and-counter-offer process that human bargaining requires.
But new transaction costs emerge. Prompt engineering costs: the LLM’s bargaining behaviour depends on the wording of the negotiation prompt, the specification of the property-rights regime, and the persona assigned to each agent. A poorly specified prompt can produce bargaining failure that has nothing to do with the underlying economics. Persona stability costs: if the LLM’s bargaining stance shifts across rounds — because the model is stochastic, or because the prompt window accumulates context that biases subsequent decisions — then the “agent” at the negotiating table is not a stable entity, and the Coasean logic of convergent bargaining may not apply. Computational bargaining costs: computing the Nash bargaining solution for \(N\) agents with heterogeneous preferences and asymmetric information is computationally non-trivial, and the LLM’s capacity to perform that computation — rather than defaulting to a heuristic — is an empirical question.
The prediction: LLM parties to a Coasean bargain will converge to the efficient outcome faster than human parties (fewer rounds, closer to the theoretical optimum), but the convergence will be more fragile — sensitive to prompt wording, persona assignment, and the computational complexity of the bargaining problem. The BDPD platform can test this by setting up a two-agent bargaining scenario with asymmetric property rights, varying the prompt structure and the initial assignment, and measuring both the efficiency of the outcome and the number of rounds to convergence.
6.5.3 Direction 3: Cross-link to cheap-talk findings
The BDPD1 paper found that cheap talk — the same intervention that produces a Cohen’s \(d \approx 1.01\) among humans — produces a \(d \approx 0.17\) among LLM agents in a commons dilemma (Brunelli 2026). The cheap-talk effect, which the classical literature on communication and cooperation treats as one of the most robust findings in behavioural social science, nearly vanishes when the agents around the table are not human.
Coasean bargaining is, in a precise sense, structured cheap talk plus a contract. Both mechanisms involve agents exchanging messages before committing to actions, with the goal of reaching a Pareto-superior outcome. Both rely on the agents’ capacity to make credible promises, to interpret the promises of others, and to condition their own behaviour on the resulting agreement. The difference is that Coasean bargaining adds a property-rights assignment and an enforcement mechanism: one party has the right, the other can pay for a waiver, and the platform can enforce the transfer.
If cheap talk alone produces near-null cooperation gains among LLM agents, then Coasean bargaining — which is cheap talk plus structure — should also struggle, unless the structure (the property-rights assignment, the enforcement mechanism) does enough work to compensate for the cheap-talk null. This is a testable hypothesis. A BDPD scenario crossing two factors — no-talk vs. Coasean bargaining vs. Pigovian tax regime, and human-subject-equivalent agents vs. LLM agents — would isolate the contribution of the institutional structure to the market-knows effect.
The prediction: Coasean bargaining should outperform unstructured cheap talk among LLM agents, because the property-rights assignment provides a focal point that unstructured cheap talk lacks. But it should under-perform Coasean bargaining among humans, because the LLM’s cheap-talk deficit — its reduced capacity to metabolise unstructured natural-language promises into behavioural commitments — is not fully compensated by the institutional structure. The gap between human and LLM Coasean bargaining efficiency is the quantity that a future BDPD market experiment would estimate.
6.5.4 Direction 4: The missing experiment — a BDPD market scenario
The BDPD platform’s architecture makes it possible to specify a market scenario that operationalises the Pigou-Coase comparison with the same parameterised rigour that the D-series pilots brought to the graduated-sanctions comparison. The scenario has not been built, but its specification follows directly from the platform’s existing capabilities.
Scenario design. \(N\) LLM agents share a logistic commons (carrying capacity \(K\), intrinsic growth rate \(r\)) over \(T\) turns. Each agent chooses a harvest level at each turn, receiving private benefit from the harvest and bearing a share of the collective cost of commons depletion. The baseline — no policy — is the standard CPR game: each agent maximises private harvest, the commons is depleted, and the tragedy materialises. Three treatment regimes are crossed with two agent architectures (LLM vs. built-in aggressive), producing a \(3 \times 2\) factorial design at \(N = 5\) seeds per cell.
Regime 1: No policy (baseline). No tax, no property rights, no bargaining forum. Agents harvest independently. The prediction for built-in aggressive agents is rapid collapse (the Hardin result). The prediction for LLM agents is uncertain — the BDPD1 Cell C result (LLM, no talk) preserved the commons in 4/5 seeds, but with a different scenario structure.
Regime 2: Pigovian tax. The platform imposes a per-unit harvest tax \(\tau\), calibrated as a fixed fraction of the estimated marginal external damage. The tax is announced before turn 1 and remains constant. Agents observe the tax in their state vector and choose harvest levels accordingly. The prediction: the tax should reduce aggregate harvest relative to the no-policy baseline, and the reduction should be larger for LLM agents (which can compute the tax-inclusive optimum) than for built-in aggressive agents (which ignore the tax unless it mechanically reduces their feasible harvest). The calibration question — what level of \(\tau\) achieves a given preservation target? — is the empirical object.
Regime 3: Coasean property rights. One agent (the “rights-holder”) is assigned the initial right to the entire sustainable yield of the commons. The other \(N-1\) agents must negotiate with the rights-holder for permission to harvest. A bargaining round precedes each harvest round: agents can propose transfers (in wealth units) to the rights-holder in exchange for harvest rights, and the rights-holder can accept, reject, or counter-propose. The platform enforces the resulting agreements: if agent \(i\) pays agent \(j\) \(x\) units for the right to harvest \(y\) units, the transfer occurs automatically at the end of the round. The prediction: Coasean bargaining should preserve the commons more effectively than the no-policy baseline, because the rights-holder internalises the full value of the commons and has an incentive to restrict total harvest to the sustainable level. The efficiency of the bargaining process — how close the outcome comes to the Coasean optimum, and how many rounds of bargaining are required — is the empirical object.
The scenario as specified is within the platform’s existing capabilities: the logistic substrate, the LLM agent infrastructure, the governance-perturbation framework (for the tax), and the structured-action schema (for bargaining proposals) all exist. The missing pieces are the bargaining-round protocol (which would extend the existing cheap-talk channel with enforceable transfer commitments) and the tax-calibration module (which would compute \(\tau\) from the commons parameters and agent behaviours). Both are incremental extensions, not architectural changes. The scenario has been designed but not implemented.
6.6 Synthesis
The good-will intuition — the market knows, tax the externality or let the parties bargain — emerges from this lecture confirmed in direction, bounded by institutional and informational conditions, and architecture-dependent in a dimension the classical tradition could not vary.
Confirmed — but informationally demanding. The Pigovian tax is directionally correct: raising the private cost of a polluting activity reduces the activity. The Coasean bargain is directionally correct: when the parties can negotiate, they often reach better outcomes than when they cannot. The market-knows intuition is not wrong. It is simply more demanding than the textbook versions acknowledge: the tax requires information the regulator rarely has; the bargain requires transaction costs the real world rarely supplies.
Bounded by institutional detail. The effectiveness of any market-based instrument — tax, cap, tradeable permit — depends on the calibration of the instrument (how high the tax, how tight the cap), the monitoring and enforcement infrastructure, the number and market power of the participants, and the political economy of instrument choice. The classical framework acknowledges these variables; the empirical record measures them imperfectly; the BDPD platform treats them as experimental parameters.
Architecture-dependent at the foundation. The mechanisms through which the market works — the price-signal channel (Pigou), the bargaining channel (Coase), the evolutionary channel (Demsetz), the institutional-design channel (Ostrom) — all require agents with specific cognitive capacities. The Pigovian tax assumes the polluter responds to the price signal. The Coasean bargain assumes the parties can negotiate credibly and commit to agreements. The Demsetz mechanism assumes the gains from internalising externalities are cognitively available to the agents. The Ostrom mechanism assumes the agents can design, adapt, and enforce institutional rules. All of these capacities are present in human agents to varying degrees. Whether they are present in LLM agents — and in what form — is the empirical question the BDPD platform is designed to answer.
The cultural payoff of the lecture is not “markets don’t work.” It is the more careful claim that markets work through specific channels, those channels are architecture-dependent, and the channel that dominates among humans may not be the channel that dominates among LLM agents. The BDPD platform, by making the agents’ architecture an experimental variable rather than an implicit constant, can ask which channels survive the transition — and the answer matters because the transition is already under way.
6.7 Open questions and the bridge to Lecture 7
6.7.1 What does the existing BDPD data not tell us about markets?
None of the three published BDPD papers operationalises a market experiment. The D-series pilots test graduated sanctions (a command-and-control instrument, not a market-based one). The BDPD1 cheap-talk experiment tests unstructured communication, not structured bargaining with enforceable transfers. The BDPD3 polycentric-cascade experiment tests inter-arena links, not price signals or property-rights assignments. The good-will intuition of this lecture — the market knows — has not been tested on the BDPD platform in any form. The four directions outlined in §5 are all speculative; the missing experiment described there is a design sketch, not a pilot script. The empirical ground under the BDPD-market reframe is thin.
6.7.2 Can LLM agents be “captured”?
The regulatory-capture literature argues that human regulators are susceptible to political pressure, informational dependence, and revolving-door incentives. An LLM regulator, by construction, has no political ambitions, no post-regulatory career, and no informational dependence beyond the data it receives in the state vector. In the traditional sense, an LLM regulator cannot be captured. But it can be gamed: agents who know that the regulator’s tax rate depends on their reported harvest, abatement cost, or damage estimate can manipulate those reports to reduce their tax liability. The question — can LLM agents learn to manipulate an LLM regulator, and does the manipulation erode the regulator’s effectiveness? — is the Goodhart-angle question that bridges this lecture to Lecture 7, and that no published BDPD experiment has addressed.
6.7.3 What about heterogeneous architectures in the market?
The missing experiment described in §5 assumes homogeneous agents within each cell (all LLM or all built-in). A more demanding experiment — and one that maps onto the mixed human-AI markets that are already emerging (algorithmic emissions trading, AI-assisted carbon accounting, automated compliance monitoring) — would cross agent architecture with market structure: some traders are LLMs, some are rule-based, some follow human-subject-equivalent decision rules. The classical market-power literature, which found that a small number of large traders can manipulate allowance prices, assumed human traders. Whether LLM traders — faster, more computationally capable, and potentially capable of colluding implicitly through correlated equilibrium strategies — would amplify or dampen market-power effects is an open question that the BDPD platform is designed to address but has not yet addressed.
6.7.4 The bridge to Lecture 7
The thread connecting Lecture 6 to Lecture 7 is the question of measurement. The Pigovian tax requires the regulator to measure the marginal damage. The Coasean bargain requires the agents to measure the value of the competing uses. The cap-and-trade market requires the regulator to measure aggregate emissions and verify compliance. In every case, the effectiveness of the market mechanism depends on the quality of the measurement — and the quality of the measurement, in turn, depends on whether the measured agents have an incentive to distort the measure.
Lecture 7 — on Goodhart’s Law and normative rigidity — asks what happens when the measure becomes the target, and when the agents whose behaviour is being measured learn to game the measurement system. The BDPD1 paper found that signal-adaptive governance (adjusting the rules in response to the state of the commons) systematically outperforms rule-bound governance (fixed rules, regardless of the state). The market-knows intuition assumes that the price signal — the tax rate, the allowance price — accurately reflects the underlying social cost. If the agents can manipulate the signal, the intuition collapses. Lecture 7 is where that collapse is examined.