Over the past 22 years the United States has imported roughly 2.85 billion individual animals belonging to nearly 30,000 species. At the Action for the Planet event this year, Michael Tlusty, professor of sustainability and food solutions at the University of Massachusetts Boston, argued that the binding constraint on policing that flow is not the border but the paperwork, and that language models should be pointed at it. The pitch is narrower than wildlife-crime rhetoric usually is: not identifying animals, but triaging import declarations to surface the records that look wrong.
Some of those animals are sold as pets or sent to zoos. Others go into research, into consumer goods, into medicines. At that volume the three questions a regulator actually cares about — is the shipment safe, is it legal, is it sustainable — stop being answerable shipment by shipment, because there is no realistic amount of human attention that scales to billions of entries.
Two categories carry most of the risk. The first is venomous species, which officials watch for and which still slip through in some consignments. Materials from the Environmental Information Service describe how agencies try to reduce the risk of non-native species establishing themselves; Wesley Daniel, a specialist at the US Geological Survey, makes the case that stopping a shipment before it lands is the most effective and cheapest way to avoid whatever an introduced species does afterwards. The second category is endangered species, already thinned by poaching, fishing and hunting, and the hardest to protect precisely because protection depends on documentation rather than observation.
Both problems reduce to the same shape: rare violations hidden inside enormous volumes of routine filings. And the routine filings are where the real failure lives. Importers are allowed to declare broad categories — "tropical fish" — without naming a species. Tlusty walked through declarations covering bats, seahorses and other animals. Seahorses are listed under CITES, the Convention on International Trade in Endangered Species of Wild Fauna and Flora, which means an import needs a permit and regulators are supposed to know where the animal came from and how it was taken. That information can simply stay buried in the documents.
Tlusty also pushed a definition worth taking seriously: treat as wild any animal that has not been domesticated. Do that and all imported seafood falls inside the regulated category, including the large share arriving from developing countries and small island states, which then need the same scrutiny as a crate of reptiles.
The most concrete technical point he made was about targeting. The existing tools — the ENA approach, detection dogs, X-ray screening — work, but only within limits. X-ray analysis in particular behaves very differently depending on the question asked of it. Ask a system to determine what is inside a box and it burns enormous compute; ask whether the box contains a crocodile or a bird and the problem collapses to something tractable. That is a real insight about where machine inspection earns its keep, and it applies to the document problem too: a model told to find seahorses in a shipping manifest is doing something far more achievable than a model asked to understand the manifest.
Here is where I think the argument gets thinner than it sounds. A model reading declarations can only find what the declaration contains. "Tropical fish" is not a detection failure that better software fixes; it is a reporting standard that permits the omission, and no amount of inference recovers a species name that was never written down. The genuinely useful version of this work is narrower and less glamorous: flagging the statistical oddities — a route, a volume, an importer, a code combination that does not match the rest of the corpus — and handing a human a shortlist. That is worth doing. It is also not the same as knowing what is in the box.
Notably absent from the account of the talk: any accuracy figure, any agency running this in production, any pilot with a number attached. Tlusty describes a capability and a data problem. What is missing is evidence that a model applied to real customs filings catches violations a targeted human review would not — which is the only claim that would move a regulator.
The framing Tlusty ends on is that compliant trade should continue and illegal trade should be found and stopped, which quietly makes the model a regulatory instrument rather than an analytical one. A system that flags the suspicious records is also, by omission, certifying everything it does not flag. Against 2.85 billion animals, the false negatives are not a rounding error in a benchmark; they are the shipments that clear.