What Is AI Fishing? A Plain-English Guide

The definition, before the marketing

Start with what the term legally-adjacent bodies actually mean by it, because the marketing rarely says. The National Institute of Standards and Technology defines an AI system as an engineered or machine-based system that, for a given set of objectives, generates outputs such as predictions, recommendations or decisions influencing real or virtual environments.

Read that twice and notice what it contains: predictions, recommendations, decisions. And notice what it does not contain: any promise that those outputs are right. The definition describes a category of software, not a level of performance.

AI fishing, then, is that category of software applied to fishing questions. When should I go? What fish is in this photo? Is anything under the boat? What should I tie on? Four questions, four different technologies, one label covering all of them — which is exactly why the label alone tells you so little.

Four technologies share the one label

The umbrella covers four distinct tools, and the honest summary of each fits in a sentence.

Bite forecasts score conditions — pressure, wind, light, moon — and the measured effects are modest: the largest study available, 341,959 muskellunge catch records, put the biggest lunar effect near 5%. Most apps, ours included, compute these scores with rules rather than trained models; our forecast article separates the two properly.

Photo identification is trained image classification, and it works impressively inside its training list: 94.3% on 16 species in the benchmark study, against trained human staff misidentifying fish at 17% in a USGS field program. Outside its list, it fails silently — the full story is in our photo-ID article.

Sonar reading applies the same classification to acoustic returns, where published detectors reach F1 scores of 0.87–0.94 on single species and the field still confirms species with nets. Our sonar article covers what that means for consumer fish finders.

Chat assistants generate answers from language models, and generate is the operative word: on legal citation tasks, models fabricated 58–88% of the time. Fishing has its own versions of that failure, catalogued in our chatbot article.

How government fisheries science uses it

The clearest picture of what this technology genuinely does comes from agencies, because they publish their numbers and their checks.

NOAA Fisheries runs camera surveys in the Gulf of Mexico that produce, in a typical year, about 2,000 deployments, 1,000 hours of video and 30 terabytes of data — and turning that into fish counts by hand takes about a year, a bottleneck the agency describes plainly. Its AI project, trained on 600,000 annotations across roughly 140 species, is expected to cut processing to a few months. The same agency’s first machine-learning system for beluga whale acoustics in Alaska classified detections at more than 96% agreement with a human expert, processing overnight what took 10–15 days by hand.

USGS applies the same pattern across the Great Lakes: models that sharpen annual prey-fish abundance estimates from sonar, trials of machine learning to age otoliths — the ear bones that grow yearly rings like a tree — and underwater drones with machine vision that detect and size invasive fish and mussels.

None of this is about catching fish. All of it is about reading instruments faster.

The pattern in every verified use

Look across those projects and one pattern repeats, and it is the most useful thing this page can tell you.

First, the technology’s real product is speed. A year of video becomes months; 10–15 days of acoustic classification becomes a night. Nothing in the verified record shows AI discovering things instruments had not already recorded — it processes what cameras, hydrophones and sonar collected, faster than people can.

Second, the speed is never trusted raw. NOAA states that scientists will still verify the automated counts against manual processing. Sonar surveys are still ground-truthed with nets. And the training itself rests on human labour at a scale worth knowing: one USGS dataset for a single species — the round goby — contains more than 37,000 images in which every fish was traced by hand, with the labels inspected by three independent observers, before any model could learn from it.

Acceleration plus verification. That is what AI fishing looks like when the people using it publish their methods — and it is the standard against which every consumer claim deserves to be held.

What this means when an app says AI

A product saying “AI” has told you its software category, and nothing else yet. The question that converts the label into information is the one agencies answer voluntarily: what was it measured against?

A trained model can answer — it has held-out data and a score. A rule-based system cannot, because nothing was ever measured; that is not damning, but it means any accuracy percentage would be invented. Our own app draws this line in public: the BiteScore page states it is rule-based and unmeasured, our forecast-accuracy article sets out the test it has not yet passed, and our data page lists exactly what goes in.

Do you need any of it to fish? No — and the evidence keeps that answer honest. In the best long-term catch dataset, who was fishing outweighed every condition a forecast can score. Where software and hard-won local knowledge disagree, our comparison article sides with the water. The technology’s fair role is the one the agencies gave it: a fast reader of slow data, checked by someone who knows what they are looking at.