Halibut commission staff propose letting AI read half of the fish earbones behind catch limits

A Pacific halibut's age is written in its earbone. The otolith adds a ring each year, and for a century scientists have read those rings the same way — break the bone, bake it, count under a microscope. Those ages feed the stock assessment that sets Alaska's halibut catch limits.

The International Pacific Halibut Commission has 1.6 million otoliths in its archive, collected since 1925. Reading one by hand costs $3 to $6. On Wednesday, commission staff asked its Scientific Review Board to let a computer read half of them.

The model, a ConvNeXt-Small neural network, comes within a year of a human reader on more than 80 percent of the images it was tested on. Staff trained it on about 9,000 otolith images and held back 2,000 to check it against. Imaging an otolith costs 40 to 50 cents, and building the model ran about $300.

Staff were direct about the limits. They described the mixed approach as slightly less precise than hand reading and slightly biased against it, based on fewer than 1,600 paired reads. Older fish come out younger than they are. How the model holds up across multiple years is unproven. The proposal keeps a 10 percent quality-control re-read by hand.

The board has watched this develop since a 2024 report on the image database and an update last year on the mixed workflow. One member asked Wednesday whether the split could go further, to 90 percent machine and 10 percent human. Another, calling aging error "more insidious than people give them credit for," pressed staff on what the change would do to the stock assessment. Staff said any test would have to match the assessment's timeline and data sources.

A day earlier, Annemarie Hoang had put it plainly: "That AI work needs to be verified."