Perfectly Readable, Perfectly Printed, Perfectly Wrong: The Limits of AI and Country of Origin Determination

Country of Origin sits at the heart of customs processing. It determines duty rates, trade measures, admissibility, and ultimately whether goods move quickly across borders or become costly exceptions.

For some goods, origin is straightforward. Agricultural products, minerals, and live animals take their origin from where they are grown, extracted, or raised, but for most manufactured goods, origin is a legal determination based on where the last substantial transformation took place. That assessment depends on bills of materials, value thresholds, supplier declarations and much more under specific trade agreements.

This is where a growing misconception is taking hold.

A new wave of solutions are promoting AI tools that claim to determine country of origin from a product image or a single document. It is an appealing proposition. It is also fundamentally flawed.

This is not a criticism of AI. At Hurricane, we build and deploy self-learning AI systems every day. We understand their strengths. We also understand their limits. The idea that a model can infer legal origin from a photo fall squarely into the latter category, not because the technology is weak, but because the premise is incorrect.

Country of Origin is a legal conclusion, not a visual feature.

It is not what is printed on a label, nor is it a probabilistic guess. It is derived from codified rules applied to supply chain data. A camera cannot access bills of materials, manufacturing steps, or value-added calculations.

Consider a smartphone labelled “Assembled in China.” Its components may originate from dozens of countries. Or a garment labelled “Made in Vietnam,” where the cotton was grown elsewhere, processed in another country, and only assembled in Vietnam. The label indicates, but it does not prove. Under audit, origin must be evidenced through documentation and traceable supply chain data.

The key point is simple, the information that determines origin is not visible on the product surface. It exists in the data behind it.

The problem becomes more serious when fraud enters the equation.

Origin fraud is a major enforcement priority globally. Recent settlements illustrate the scale and intent behind misrepresentation. Ceratizit agreed to a $54.4 million settlement involving Chinese-origin goods transshipped through Taiwan. Evolutions Flooring settled for $8.1 million over similar practices involving Malaysia. Univar paid $62.5 million in a case involving saccharin routed through Taiwan to avoid steep duties.

In each instance, the goods were accompanied by documentation and markings that appeared consistent and credible. They were also incorrect.

An image recognition model would not have detected the deception. It would likely have reinforced it, returning a high-confidence prediction based on what it could see.

This leads directly to another issue, confidence scores.

AI outputs often present a confidence percentage that appears authoritative. In reality, this is a probability, not a measure of correctness. Models are also known to be overconfident, with reported confidence frequently exceeding actual accuracy.

In a Customs context, that distinction matters. A confidence score is not a defence. Legal responsibility for origin declarations remains with the declarant, regardless of how the result was generated.

That brings us to risk.

For importers and merchants, incorrect origin declarations result in underpaid duties, penalties, shipment delays, or seizure. For carriers, forwarders, and postal operators, the consequences include increased inspections, operational delays, customer service costs, and reputational damage. With ICS2 and evolving EU low-value consignment regimes tightening data requirements, the cost of poor data quality will only increase.

None of this suggests AI has no role in origin analysis. On the contrary, it has a significant one.

AI is highly effective when applied to risk targeting, anomaly detection, and document triage. It adds value when combined with structured product data, supply chain information, and rules of origin logic. Crucially, effective systems surface uncertainty rather than masking it.

This is the direction both industry and regulators are taking. HMRC’s 2025 Transformation Roadmap emphasises AI in compliance and targeting, while retaining human decision-making for complex determinations. US CBP follows a similar model in practice. AI supports, it does not adjudicate.

The idea of “photo in, origin out” oversimplifies a legal and data-intensive process into something it is not.

If you work in cross-border trade, customs, or logistics, you are likely already seeing these tools appear in procurement discussions. The question is not whether to use AI, but how to use it responsibly.

Because when it comes to Country of Origin, perfectly readable and perfectly printed can still be perfectly wrong.

At Hurricane, we believe origin determination should be evidenced, not estimated. If you are navigating ICS2, low-value consignment changes, or evaluating where AI genuinely adds value in your cross-border operation, we are happy to talk.
Link to book demo contact.

Country of Origin is a legal conclusion. Treat it like one. At Hurricane, we help carriers, marketplaces, postal operators and merchants determine, evidence and declare origin with the data and rules logic the job actually requires. Get in touch to find out more.