Moral hazard and unobserved action in fisheries management

The assessment that was already wrong

In March, a fisheries scientist signs off a stock assessment for a groundfish population — call it the northern shelf cod stock, though the pattern recurs across cod, herring, hake, whatever species runs on a biennial cycle. The assessment draws on the last two years of trawl survey data, aged catch samples, and a population model tuned to those inputs. It concludes the spawning stock biomass is stable. A total allowable catch is set from that conclusion and handed to the quota managers. Vessels fish against it for the next two seasons.

Nine months in, a marine heatwave pushes bottom temperatures up by more than two degrees across the survey grid. Recruitment collapses. Nobody notices in real time, because nobody is meant to: the next trawl survey isn't scheduled until the following spring, and the next full assessment won't be published for another year after that. Catch reports keep arriving weekly, quota filings keep clearing, and every one of those numbers looks unremarkable in isolation. The stock is being fished down against a target that stopped being true the week the water warmed. By the time the next assessment catches up, the biomass estimate drops sharply, the TAC is cut hard to compensate, and the cut lands on a fleet that did nothing differently from what the rules told it to do.

What actually went wrong

No one lied. No vessel misreported its catch, no observer falsified a log, no quota filing was fraudulent. The scientist who signed the assessment used the best data available on the day it was due. The failure sits somewhere else: between the assessment date and the present, the thing the quota was supposed to track — the actual, current state of the stock — was not observed at all. It was inferred, once, from a snapshot, and then treated as current for two years because the institutional cycle had no mechanism for treating it as anything else.

This is not a data quality problem in the ordinary sense. The catch reports were accurate. The survey was competently run. The gap is structural: the action that mattered — the stock's real-time trajectory, and by extension every fishing decision made against a quota derived from a stale trajectory — was unobserved for the whole interval between assessments. Vessels did not need to game anything to exploit that gap. The gap did the work on its own.

Moral hazard, named properly

This is moral hazard in its precise economic sense, not its loose one. The term comes from Kenneth Arrow's 1963 paper on medical insurance, and it does not mean dishonesty. It means one party takes an action the other cannot observe, and the unobserved action is the one that matters. Michael Rothschild, Joseph Stiglitz, James Mirrlees and Bengt Holmström spent the 1970s formalising it as the principal-agent problem: because effort — or in this case, the true state of the resource — is hidden between observation points, a contract can only be written against outcomes, and outcomes are noisy. Holmström's 1979 result, "Moral Hazard and Observability," made the logic exact: the best contract available depends entirely on how informative the available signal is. Mirrlees shared a Nobel Prize with William Vickrey in 1996; Holmström shared a separate one with Oliver Hart in 2016. They solved different corners of the same problem: why sensible, honest parties still end up with inefficient arrangements, with no villain required.

In fisheries management, the assessment cycle is the contract, the TAC is the price, and the stock's actual biomass between assessments is the hidden action. The regulator is not being deceived by the fleet. The regulator is being deceived by time — by the fact that its own instrument for knowing the state of the world updates on a schedule slower than the world changes. Two seasons is not a scandal. It is the width of the blind spot the whole quota system was built around.

Intake as the fix that isn't a slogan

The natural response is more frequent assessment: annual instead of biennial, or continuous stock indices built from satellite temperature anomalies, acoustic surveys, and near-real-time catch composition rather than a single annual trawl. This is not a paperwork change. It is a change in what the regulator is permitted to observe, and therefore in what it can hold anyone to.

A system that only ever sees a frozen stock assessment behaves like a Large Language Model: its knowledge has a cutoff, and nothing that happens after that cutoff can be checked, priced, or acted on until the next cycle arrives. A system that observes the fishery only while a survey vessel is actually on the grounds behaves like a Large World Model: conduct within the surveyed window is visible and can be managed, but the moment the vessel leaves, the stock's trajectory reverts to invisible, and the next two years are managed on inference. A Large Universe Model, in this domain, is the fisheries-management analogue of never letting the boat leave: catch reports, satellite-derived temperature anomalies, acoustic biomass indices, and quota filings all streamed continuously, each carrying a timestamp and a confidence value, each belief about stock size revised as new readings arrive rather than replaced wholesale on a two-year clock.

This does not make the stock knowable in some absolute sense. It makes the unobserved margin — the interval during which a quota can drift out of truth without anyone noticing — bounded and estimable rather than open-ended. That is the entire claim, and it is smaller than it sounds, and still worth having.

Two objections a fisheries scientist would actually raise

"Put cameras and continuous sensors on everything and the fleet will fish to the sensor, not to the stock. You'll get technically compliant catch composition and discard behaviour that looks clean on video and still guts the population."

This is Goodhart's law with a hull number, and it is not hypothetical. Electronic monitoring programmes in New Zealand and parts of the EU moved observer coverage from roughly 2 per cent human placement to near-total video coverage, and reported discard rates rose sharply once cameras went in — not because vessels started discarding more, but because discarding that had always happened became visible for the first time. The objection is right that a single new stream invites gaming around its edges: species misidentification on camera, high-grading just outside the lens angle, timing catches to gaps in satellite pass coverage. But the force of the objection falls as the number of independent streams rises. Gaming needs a gap between what is measured and what matters, and a single camera has exactly one such gap. Cross-checking video against VMS position, catch-report weight, and acoustic biomass trend closes most of the room for any one stream to be played without the inconsistency showing up somewhere else. The residual problem changes from hidden effort to distorted effort. That is a real, permanent problem. It is a smaller one than the one it replaces.

"Fisheries already solved this with individual transferable quotas — give a vessel owner a tradeable property right in the stock and their incentive aligns with conservation without anyone needing to watch them."

ITQ systems, from New Zealand's 1986 Quota Management System onward, are a genuine instance of Holmström and Milgrom's insight: sometimes it is cheaper to align interests than to gather information. But every ITQ is priced against an observed stock assessment, and every trade in quota value is a bet on the next assessment's number. When the assessment itself lags the stock by two seasons, the property right is aligned with a fiction. Incentive design and intake are not competitors here; the quota's incentive properties are only as good as the assessment feeding it. Better streams make ITQs work as intended rather than replacing the need for them.

What is left after the streams run continuously

Nothing in continuous intake reaches the fisheries scientist's private judgement call — the decision, made once, to weight one ageing sample over another, or to flag a survey anomaly as noise rather than signal. That remains genuinely unobservable, and no sensor network changes it. What continuous streaming buys is a dense base rate: a scientist's call can be checked against thousands of comparable calls made under similar conditions, so an assessment that quietly diverges from what the temperature and acoustic streams were already showing becomes visible as an anomaly, even though the reasoning behind it never was. The two-season gap does not close entirely. It shrinks to the width of genuine judgement, which is where the problem belongs, and where no wider intake will ever reach.