What counts as IUU fishing, and why has it been hard to measure?
IUU fishing is the umbrella term for catches taken in violation of applicable law, catches that go unreported or misreported to authorities, and catches taken in waters where no conservation regime applies at all. That three-part definition comes straight from the FAO’s framing of the problem, and each branch creates a different measurement gap. Illegal fishing hides behind jurisdictional friction. Unreported fishing is absent from the ledger by definition. Unregulated fishing happens where no ledger exists.
The scale estimates that circulate are large and necessarily soft. IUU catch is commonly pegged at roughly US$23 billion a year in lost economic value, a figure attributed to FAO work and repeated across secondary sources. By share, the same compilation estimates IUU at up to 30% of total catch in some important fisheries. Treat those as order-of-magnitude bounds, not survey-grade counts. The whole problem is that nobody is counting from a complete record.
The people most exposed are the ones with the least capacity to document the exposure. NOAA’s overview is blunt that developing countries dependent on fisheries for food security and export income bear the brunt, because IUU operators exploit weak management regimes and limited monitoring budgets. That asymmetry matters here. The countries losing the most catch are the ones least able to staff the audits that would catch it.
How does the enforcement data layer work today?
Current IUU enforcement leans on a patchwork: self-reported import paperwork, satellite-based vessel tracking, port inspections, and species-specific documentation rules. Each produces documents. None of them, on its own, produces a joined-up picture across the supply chain.
The United States’ contribution to that patchwork is the Seafood Import Monitoring Program (SIMP), a rule that requires importers of species considered especially vulnerable to IUU fishing to hold and report permitting and chain-of-custody records. The State Department describes US reporting and recordkeeping measures as a way to keep IUU-caught fish out of the domestic market. SIMP is a paper-trail rule. It tells importers what to record and hold, and gives regulators a hook to demand those records. It does not by itself read the records, link them across borders, or flag the vessel names that recur across multiple suspicious shipments. That cross-referencing is manual, slow, and capacity-bound.
The deeper structural problem is heterogeneity. A port inspection report in one country, a customs declaration in another, an NGO investigation, a flag-state sanction notice, and a trade-press article about a labor abuse case all describe overlapping reality in incompatible formats and languages. The evidence is fragmented across document types that no single relational schema was built to hold. That is the gap an extraction system is designed to close: treat every document as unstructured text, let a language model read it, and emit records that can be queried together.
What does IUU+DB actually extract?
IUU+DB runs an LLM over heterogeneous documents in stages. It ingests them, classifies whether each one describes a relevant IUU+ incident, extracts structured data elements from the ones that do, and then supports deduplication and trend analysis across the resulting database. That description is from the paper’s abstract; the full methodology, including which model family drives extraction and how classification is prompted, sits behind the PDF.
The extracted fields are where the enforcement value lives. Per the abstract, the system pulls out six kinds of element:
| Extracted element | What it enables |
|---|---|
| Actors | Names companies, captains, and beneficial owners worth tracking across incidents. |
| Locations | Maps where incidents cluster, exposing ports and waters that function as hotspots. |
| Species | Lets importers and regulators run source and species risk assessments. |
| Vessels | Links repeat-offender vessels across separate documents and jurisdictions. |
| Violations | Classifies the type of illicit behavior for pattern and trend analysis. |
| Enforcement outcomes | Records whether an incident produced a sanction, a seizure, or nothing. |
The deduplication step matters more than it sounds. A single IUU incident generates ripples: a flag-state report, a port-state seizure record, a wire story, an NGO brief. Without merging, the same event shows up four times and inflates every trend. Without merging done badly, distinct incidents collapse into one and the database understates the problem. Deduplication against duplicated and adversarial text is a hard information-retrieval problem in its own right, and it sits between the extraction step and any usable trend output.
Who is the database meant to serve?
The paper names three audiences explicitly, and they want different and partly contradictory things from the same records. Academic and NGO researchers want geographic and behavioral hotspots surfaced and evidence organized so patterns become arguable. Industry sourcing teams want source and species risk assessments they can run before buying. Government agencies want support for policy implementation and targeted enforcement. The abstract lists all three as intended beneficiaries.
That breadth is worth pausing on, because each audience imposes a different bar on the same extracted data. A researcher drawing a heatmap can tolerate a precision slip that mislabels a few port calls, because the gradient still points the right way. A government agency building a targeted enforcement action against a named vessel cannot. The same database serves both, but the precision floor for “actionable” is set by the most consequential use case, not the average one. That tension is the setup for the question the abstract leaves open.
Does extraction precision hold when the documents are adversarial?
This is the open question, and the public abstract does not settle it with numbers.
The abstract reports “case studies and validation results” showing the system can organize fragmented evidence and surface hotspots, but it does not disclose F1, precision, or recall figures in the abstract text. The angle the whole paper invites, cheap structured extraction across a full document population, depends entirely on those numbers being high enough for the most demanding of the three audiences. A preprint that validates on clean port reports is a different proposition from one that validates on deliberately vague trade filings or documents written to mislead.
Adversarial text is the specific worry. An LLM extracting actors and violations from a document whose author intended to obscure the actor or minimize the violation is not doing ordinary named-entity recognition. It is doing extraction against a writer who is, in effect, attacking the extractor. Grey literature and trade records are full of euphemism, shell-company indirection, and species mislabeling. The same property that makes LLMs attractive here, tolerance for messy unstructured prose, is the property that makes them vulnerable to confident misreads when the messiness is intentional rather than accidental. General background on how LLMs process text describes the attention-based mechanism that underlies this kind of extraction; it does not, and cannot, guarantee precision on adversarial input.
The honest framing is that IUU+DB demonstrates a pipeline architecture and claims validation through case studies, while the precision question, the one that decides whether this becomes an enforcement tool or a research aid, is exactly what readers should check the full paper for before relying on any extracted record.
What changes if the bottleneck becomes document coverage?
If extraction precision clears the enforcement bar, the limiting factor in supply-chain accountability moves from how many auditors an agency can hire to how many documents an agency can feed the model. That is a real shift. Auditor capacity is linear in headcount and budget. Extraction throughput is a function of compute and document access. The countries that lose the most IUU catch are the ones least able to staff manual audits, but they are not, in principle, the least able to run a model over a document corpus. That inverts part of the asymmetry NOAA describes.
The second-order effect is legibility. Fleets, processors, and transshipment operators that have lived in unstructured records, legible in principle to anyone who read every port report, become legible in practice to anyone who can run extraction over them. A vessel name that recurs across five separately published incident reports in three languages is the kind of signal a human analyst almost never assembles and a deduplication step almost always can. That is the mechanism by which an extraction database becomes an enforcement lever. It is also the mechanism by which the same database becomes a liability if precision is low: a confidently wrong linkage of a clean vessel to an abuse case is a defamation risk, not just a database error.
The author list mixes fisheries-domain experts with computer scientists, and the work is filed under information retrieval, AI, and computers and society per the arXiv classification. That composition is the right tell. The hard part of this system is not the model. It is the domain knowledge needed to define what counts as an incident, what a violation category should be, and when two records describe the same event. The LLM is the extraction engine; the fisheries expertise is the schema.
Frequently Asked Questions
Which fisheries show the highest reported share of IUU catch?
West Africa is a documented hotspot: estimates there run as high as 37% of fish caught being unreported or illegal. That is above the broader ‘up to 30%’ figure often cited for important fisheries globally, and it lines up with the article’s point that countries with the weakest monitoring capacity lose the most catch.
How does document extraction differ from satellite vessel monitoring as an enforcement tool?
Satellite systems such as AIS track where a vessel broadcasts that it is. Extraction tools like IUU+DB read port reports, customs filings, sanctions notices, and news text instead. AIS can be switched off or spoofed; documents can be falsified or incomplete. Each tool sees a different slice of the supply chain, so they complement rather than replace each other.
What does a regulator actually need to deploy an extraction pipeline like IUU+DB?
Beyond a language model, the regulator needs document access rights, a schema agreed on by fisheries domain experts, and a deduplication workflow. The paper’s mixed author list of fisheries scientists and computer scientists signals that the hard part is defining what counts as an incident and a violation category, not just running inference.
What is the enforcement risk if extraction links a clean vessel to an incident by mistake?
A false vessel linkage is not only a database error; it can become a defamation or due process liability if it feeds a seizure, blacklist, or import denial. That means agencies need a documented audit trail and human review before acting on any extracted link, no matter how confident the model looks.
Could better extraction push illegal operators to produce cleaner, more misleading paperwork?
Yes. Once operators know documents are being mined, the attack shifts from hiding events to seeding plausible deniability through euphemisms, shell-company indirection, and species mislabeling. The same tolerance for messy prose that helps LLMs read fragmented records also makes them vulnerable to intentional noise, so precision metrics must be revalidated against evolving document tactics.
