What a machine can and cannot do here. The topside count and identification is the slow part of the method, and a computer-vision model can accelerate it: modern detectors can box the organisms in an image and propose species identifications, turning an hour of tallying into minutes of review. But a model inherits the method's limits exactly, and it is worth being clear-eyed about this. A single 360° frame carries no distance to any organism (Module 5), and a model cannot recover from the pixels what the pixels do not contain — so a detector can count what is visible in the frame, but it cannot tell how far away a fish was, cannot bound the plot by distance, and cannot size anything. And the conspicuousness limit survives: a cryptic organism hidden from a diver is usually hidden from the model too. AI makes the count faster, not deeper.
The one rule: the machine proposes, you decide. Because of those limits — and one more, below — an AI count is never taken as-is. The model produces proposals: boxes and tentative identifications. The demographer then does two things, and both matter: verify — accept the correct proposals and reject the false ones — and augment — add the organisms the model missed and correct the identifications it got wrong. The human-confirmed set, not the model's raw output, is the count of record. This is not a courtesy check; it is the design. The AI sits inside the image count (the second observer, N₂, of Module 5) as a first pass the human corrects — it does not become a new, unchecked observer, and its raw number is logged but never recorded as the result.
Why the human cannot be removed — model drift. The method's change-detector rests on the counting bias being roughly constant over time, so that it cancels when the same place is compared with itself. A trained diver's bias is stable across seasons. A model's bias is stable only until its version changes — and models are retrained and updated silently, on a schedule set by someone else. Run last year's footage through this year's model and the count can shift for reasons that have nothing to do with the reef. If the raw model output were the record, that drift would masquerade as change in the reef — the one thing the whole method exists to measure truthfully. The human in the middle is the fixed reference that absorbs the drift: because a person verified both years, the comparison is between two human-confirmed counts, and the model's drift shows up only in the logged raw numbers, where it is a useful diagnostic rather than a false signal.
Keeping AI-assisted data trustworthy — provenance. Three things must be recorded whenever a model helps produce a count, and together they are the AI extension of the limitations statement: the model and its version, so a later reader knows what produced the proposals; the raw model count alongside the human-confirmed count, which is that survey's measured model error against human ground truth; and the retained footage, so any count can be re-derived when the model changes. A count without this provenance cannot be trusted across time, because there is no way to tell reef change from model change.
Which tool — a website matter, not a course matter. This module deliberately names no specific model. The durable principle — an assistive detector, model-agnostic, human-verified, and provenance-recorded — is what belongs in a course document. The specific engine, whether iNaturalist's identifier, a general object-detector, or whatever is best next year, is perishable and lives on the course website, named and updated there, exactly as the calculator does (Section Four). The method does not depend on any one product, and nothing in this guide is tied to a tool that will change.