The AI Journal

When a container ship runs aground and its owner orders containers thrown overboard to refloat it, the law does not ask which cargo owner was simply unlucky. It totals everything saved — ship, fuel, and every container that stayed aboard — and bills every owner a share of what was lost, in proportion to what each of them had at risk on the voyage. This is general average, and it may be the oldest continuously invoked rule in commercial law: a version of it appears in the Digest of Justinian, attributed to a Rhodian statute older than Rome’s own maritime power, and shipping still applies something recognisably the same rule under the York-Antwerp Rules today. Longevity of that order tends to be read as a verdict. If a mechanism kept its basic shape while marine insurance, actuarial science and admiralty courts grew up around it, the reasoning goes, it must have been doing something right economically — pooling an unpredictable loss efficiently among the people exposed to it, much as insurance does. A 2024 paper by Luca Anderlini and Joshua Teitelbaum tests that assumption directly, by asking a narrower question that survival alone cannot answer: whether the mechanism actually gives owners an incentive to report their cargo’s value honestly, and whether the allocation that results is efficient. Under the model economists ordinarily use for decisions made under risk, the answer to both is no.

General average asks each owner, after the loss, to declare what their cargo was worth, and it charges contribution in proportion to that declared figure while compensating whoever’s goods were actually sacrificed on the same basis. Where an ordinary insurance contract fixes the payout and the premium in advance of any claim, general average sets both only afterwards, from numbers the interested parties supply themselves. Anderlini and Teitelbaum model this as a problem in mechanism design: does an owner who reasons in the way expected-utility theory assumes — weighing outcomes by their probabilities and maximising the resulting average — do better by declaring truthfully? Under that assumption they show the answer is no. An expected-utility maximiser has an incentive to misstate value, and the allocation that results is not Pareto efficient. The ancient rule fails the test that a genuine mutual-insurance mechanism is supposed to pass.

It does not fail across the board. The same paper shows that general average is both truthful and efficient once owners are assumed to reason differently: not by weighing outcomes against probabilities they can actually state, but by judging each option on its worst case, a rule decision theorists call maxmin. A maxmin owner does not ask how likely a storm severe enough to force jettison happens to be; unable or unwilling to put a number on it, they plan for the possibility and judge a contract by how badly it could go rather than by an average weighted by likelihoods nobody aboard could credibly supply. Under that description of the decision-maker, honest reporting becomes each owner’s best strategy, and the resulting contribution is efficient. The efficiency story usually told about general average’s survival is not exactly wrong, then; it is aimed at the wrong kind of rationality. The mechanism was never solving the problem expected-utility theory poses. It was solving the one that arises when probabilities cannot be assigned at all — a distinction economists trace to Frank Knight, who separated calculable risk from an uncertainty that resists any probability figure — and general average sits in Knight’s second category, not his first.

That reattribution answers a puzzle that efficiency-as-such cannot. Actuarial marine insurance, with premiums set from tabulated loss experience across many voyages, existed alongside general average from at least the seventeenth century, and by the following one underwriters in London and Amsterdam were pricing risk from genuine statistical records rather than guesswork. If general average were simply an early, crude attempt at the probabilistic risk-pooling that insurance performs formally, the arrival of proper actuarial pricing should have made it redundant, or at least pulled its terms toward an insurance contract’s. Neither happened. General average kept its own procedure — a jettison still has to be shown reasonably necessary against an actual peril, adjusted after the event from what actually occurred rather than from a table of odds fixed in advance — running alongside insurance rather than dissolving into it. Ron Harris’s account of early modern maritime risk-sharing treats that persistence as evidence the two were never doing the same job: insurance priced a calculable hazard ahead of the voyage; general average adjudicated an actual event nobody could have priced, because the specific peril, and the master’s specific decision to sacrifice this cargo rather than that, could not be reduced to a rate table beforehand.

The strongest objection to reading the mechanism this way is that it credits medieval and early modern shipowners with a decision-theoretic sophistication nothing in the record shows they possessed. Nobody aboard a fourteenth-century cog was reasoning explicitly about maxmin preferences; they were following an inherited custom, codified in the Rolls of Oléron and enforced by admiralty courts that cared about precedent, not about which axioms of choice under uncertainty a merchant happened to satisfy. Guido Rossi’s recent study of the doctrine’s Roman-law ancestry presses the same worry from a different angle: the apparent unbroken continuity from the Rhodian sea-law to the modern rule, often invoked as proof of a self-evidently correct solution independently converged on across legal traditions, turns out to be considerably more reconstructed and discontinuous than that story needs, stitched together retrospectively by later jurists rather than transmitted whole. If neither the actors nor the doctrine’s own history supports a decision-theoretic reading, the maxmin result risks being read backwards into a practice with a simpler explanation: contribution in proportion to declared value was a workable convention because it was easy to administer and hard to contest after the fact, not because it satisfied any particular axiom of rational choice. That objection is fair as far as it goes, but it answers a different question from the one the formal result poses. Anderlini and Teitelbaum are not claiming that fourteenth-century merchants derived maxmin preferences from first principles; they are identifying which assumption about how a rational actor handles genuine uncertainty makes a three-thousand-year-old rule’s design intelligible rather than perverse. A convention can be sustained by custom and still turn out to be exactly the rule a maxmin-reasoning population would have needed, for the same reason a trait can arrive by drift and still be one that a fitness calculation would also justify: the result explains why the rule was fit for purpose once it existed, without having to explain how anyone arrived at it.

What general average actually demonstrates, across three thousand years spent running alongside instruments built on the opposite assumption, is not that ancient markets stumbled onto modern insurance early. It is that a rule engineered for agents who cannot price a storm will keep outperforming one built for agents who can, for exactly as long as the storm stays unpriceable — and that shipping, for most of its history, never stopped being a business conducted in that condition.