What the Ocean Knows

The Atlantic cod collapse is one of the best-documented ecological and economic catastrophes in the history of food production. Canadian cod stocks in the Grand Banks had supported fishing communities for five hundred years. By 1992, NOAA and Fisheries and Oceans Canada estimated that the stock had fallen to roughly 1% of its historic levels. The Canadian government imposed a moratorium on cod fishing. An estimated 30,000 fishers and plant workers lost their jobs overnight. The ecosystem changes were extensive enough that, thirty years later, the stock has not recovered to levels that support commercial fishing.

The collapse happened despite stock assessment. Scientists were measuring the population. They were making recommendations. The recommendations were not followed with sufficient rigor, or followed too late, or based on models that underestimated the rate of decline. The politics of fishing communities and short-term economic interests overrode the precautionary science.

This is the context for AI in fisheries management: a history of models being defeated by a combination of biological complexity, data limitations, and political economy.

Modern fisheries stock assessment depends on a combination of survey data (research vessels systematically trawling transects and counting what they catch), acoustic surveys (sonar-based estimation of fish biomass), and commercial catch data (logbooks, observer programs, landing records). The data is expensive to collect, temporally sparse, and subject to systematic biases—particularly in the commercial catch data, where underreporting is endemic.

Machine learning is improving every stage of this process. Computer vision applied to underwater video—from survey vessels, from baited remote underwater video stations, from commercial trawl cameras—can identify and count fish species with accuracy that is competitive with trained human observers. This matters because human observers are expensive, operationally limiting, and can only be deployed on a small fraction of fishing trips.

The startup Pelagic Data Systems has been deploying GPS tracking devices on small-scale fishing vessels in Southeast Asia and West Africa since 2016, building the first spatially and temporally continuous record of fishing effort in regions where official monitoring barely existed. The data feeds into stock assessment models that previously had no reliable fishing mortality estimates for these regions. You cannot manage what you cannot measure; these systems are creating a measurement layer in places where fisheries management had been operating essentially blind.

Hydroacoustic data interpretation—extracting fish density and distribution estimates from echo sounder recordings—is another area where deep learning has shown clear improvement over traditional processing. Echoview software with ML components can now classify acoustic targets (distinguishing fish schools from zooplankton, from gas bubbles, from the seafloor) faster and with higher accuracy than manual analysis, enabling more extensive survey coverage with existing survey vessels.

Aquaculture is the other half of the ocean food production story, and it’s where AI is moving from research novelty to commercial standard fastest. Global aquaculture production overtook wild-catch fisheries in terms of food fish supplied in 2023—the first time in history that more fish was farmed than caught. The productivity frontier in seafood is now in controlled production systems, not in managing wild stocks.

Salmon farming—particularly in Norway, which produces roughly 60% of global Atlantic salmon—operates at a scale where individual fish welfare and production optimization interact in ways that require continuous data collection. Norwegian salmon farms are monitored by underwater cameras with computer vision systems that track individual fish size, feeding behavior, sea lice infestation levels, and behavioral indicators of stress. The regulatory requirements in Norway mandate sea lice counts; the economic imperative to minimize sea lice treatment costs (treatments are expensive and stressful to fish) made computer vision for continuous lice counting economically compelling before it was required.

The precision feeding application is particularly valuable. Salmon feeding from pellet dispensers, calibrated by feeding cameras that watch for uneaten pellets sinking through the water column, can reduce feed conversion ratios—the amount of feed required per kilogram of fish produced—by 10-15% compared to fixed feeding schedules. Given that feed represents roughly 50% of salmon farming operating costs, that’s a direct and substantial margin improvement.

More sophisticated systems integrate feeding behavior, water temperature, dissolved oxygen levels, and fish density to make real-time feeding rate adjustments. Akva Group, the Norwegian aquaculture technology company, reports that their sensor-integrated feeding systems achieve feed conversion ratios of around 1.1-1.2 kg of feed per kg of fish produced—compared to industry averages closer to 1.3-1.4 without optimization. The difference is both an economic win and an environmental improvement (unconsumed feed that sinks to the seafloor creates oxygen depletion and habitat damage).

The disease forecasting application is where AI is trying to address what is arguably aquaculture’s most acute risk. Infectious salmon anemia, pancreas disease, and amoebic gill disease can devastate a farm site when they appear. The transmission vectors are partly environmental (infected wild fish, water temperature and oxygen conditions that favor pathogens) and partly operational (fish density, stress, prior exposure). Early warning systems that identify elevated disease risk before clinical signs appear could substantially reduce the losses—which can be total at a site when a severe outbreak hits.

Cermaq, one of the major salmon farming companies, has been developing disease early warning models that integrate environmental monitoring, fish behavior tracking, sea lice data, and historical outbreak records. The models are not yet at the accuracy levels that would allow confident preventive action, but they’re improving year over year as the training datasets grow.

The wild fisheries monitoring application has an intelligence dimension that is increasingly relevant: illegal, unreported, and unregulated (IUU) fishing. The Global Fishing Watch initiative, launched with support from Google, SkyTruth, and Oceana, uses AIS vessel tracking data and satellite imagery to monitor fishing vessel behavior globally. Machine learning classifiers identify when vessels are behaving like they’re fishing versus transiting, flag vessels operating in protected areas or in jurisdictions where they’re not authorized, and detect the “dark vessel” behavior of ships that turn off their AIS transponders.

The 2024 publication of the first comprehensive global fishing effort dataset in Nature demonstrated that AI-processed satellite data revealed twice as much fishing activity in some regions as official vessel monitoring systems had captured. This is the measurement layer that makes any serious international fisheries governance possible.

The honest assessment of AI’s limitations in ocean systems comes back to the cod collapse context. The tools are better. The data is richer. The models are more sophisticated. None of this is sufficient if the political economy of fishing management continues to prioritize short-term catch over long-term sustainability.

The Norwegian salmon industry has been willing to accept regulatory monitoring requirements and transparency because the industry is highly concentrated (a handful of large companies dominate), the Norwegian regulatory environment is relatively functional, and the economic interests of the industry are at least partly aligned with sustainable management. West African or Southeast Asian small-scale fisheries, with thousands of individual operators, minimal regulatory capacity, and immediate subsistence pressures, face a fundamentally different governance problem.

AI that provides excellent fisheries monitoring data and delivers it to a regulator with the capacity and political will to act on it is genuinely useful. The same monitoring data delivered to a regulator without enforcement capacity, or ignored by fishing communities with no economic alternatives to over-fishing, changes nothing.

The ocean knows quite a lot. The question is whether the humans managing it are in a position to act on what they learn.