Halibut commission staff propose aging half of otoliths by AI

International Pacific Halibut Commission staff proposed Wednesday that its Scientific Review Board adopt a workflow aging half of Pacific halibut earbones by artificial intelligence and half by hand, a split that would touch the data behind Alaska catch limits. A new ConvNeXt-Small model matches manual break-and-bake reads within one year on more than 80% of test images. Staff developed it on about 9,000 of 11,107 otolith images, holding out 2,000 for testing; the commission's archive holds 1.6 million otoliths collected since 1925. The board has followed the effort since a 2024 report on the image database and CNN testing, and a 2025 update on the mixed workflow.

The design keeps a 10% quality-control re-read. Staff put manual aging at $3 to $6 per otolith, versus 40 to 50 cents for imaging plus about $300 per model build. They called the mixed approach slightly less precise and slightly biased against break-and-bake, based on fewer than 1,600 paired reads, with older fish underestimated and multi-year performance still unproven.

One board member asked, "can you go to 90/10?" Another member, calling aging error "more insidious than people give them credit for," pressed staff on how the change would affect the stock assessment; staff said any test would need to match the assessment's intended timeline and data sources. Annemarie Hoang said a day earlier that "that AI work needs to be verified."