Artificial intelligence systems, and large language models (LLMs) in particular, are increasingly embedded in the infrastructure of cross-border commerce: customs classification, trade documentation, multilingual communication, contract drafting, and regulatory compliance screening. A well-documented limitation of these systems is "hallucination" — the generation of fluent, confident output that is factually incorrect, internally inconsistent, or entirely fabricated. This paper argues that, in the trade context, hallucination is not merely a technical nuisance but is beginning to function as a de facto non-tariff barrier (NTB): it raises compliance costs, introduces legal liability and documentation risk, erodes trust in AI-mediated trade facilitation tools, and creates asymmetric burdens that fall disproportionately on smaller exporters and firms in developing economies with weaker capacity to audit AI output. Using a doctrinal and case-based qualitative method, the paper synthesises secondary evidence — including the Moffatt v. Air Canada ruling, customs and trade-compliance industry reports, and recent World Trade Organization and International Monetary Fund publications on AI in trade — to build a conceptual framework describing four mechanisms through which hallucination distorts trade: compliance-cost inflation, information asymmetry, liability displacement, and trust erosion in digital trade infrastructure. The paper concludes with policy recommendations for regulators, firms, and international bodies, and identifies primary-data gaps that future empirical research should address.
This paper adopts a qualitative, doctrinal, and case-based design rather than a primary-data survey design. Three considerations informed this choice. First, the phenomenon under study is recent — most of the institutional literature on AI in trade facilitation dates from 2025 [4], so large-sample primary data on hallucination-driven trade friction specifically does not yet exist in a form that could be independently verified. Second, the paper's contribution is conceptual: building a defensible framework connecting two established literatures (AI hallucination and NTB theory) that have not previously been synthesised in this way. Third, case-based analysis of a documented legal dispute [18] provides a verifiable, citable instance of hallucination-driven commercial and legal cost that can anchor the conceptual claims in a real, adjudicated fact pattern rather than hypothetical scenarios.
Sources were drawn from three tiers:
Building on the literatures reviewed above, this paper proposes that AI hallucination functions as a non-tariff barrier through four overlapping mechanisms.
As customs administrations and private logistics firms increasingly deploy AI for Harmonized System (HS) code classification, rules-of-origin determination, and documentation drafting [19-20], hallucinated classifications or invented regulatory requirements generate rework, penalties, and shipment delays. Industry sources note that incorrect HS coding carries direct legal and financial consequences, including misapplied duties [20], and that inaccurate shipping documentation remains one of the most persistent sources of friction in global trade even before AI is introduced [19]. Where AI tools are trusted without adequate human verification, hallucinated outputs convert what should be a productivity gain into a new source of cost — functionally equivalent to a technical barrier to trade, since the exporter must now spend additional resources verifying AI-assisted documentation against authoritative sources.
NTB theory has long recognised that regulatory complexity disadvantages parties with fewer resources to interpret and verify requirements [15]. AI hallucination reproduces and potentially deepens this asymmetry. The WTO–ICC business survey referenced in the World Trade Report 2025 found that only 41 per cent of small firms’ report using AI in their trade operations, compared with over 60 per cent of large firms [6]. Smaller exporters that do adopt AI tools for compliance or translation, but lack the in-house legal or technical capacity to independently verify AI output, are structurally more exposed to hallucination-driven errors than large firms with dedicated compliance teams. This mirrors the long-standing finding that TBT and SPS measures disproportionately burden developing-country and small-firm exporters [15] AI hallucination risks becoming a digital-era analogue of the same dynamic.
The 2024 ruling in Moffatt v. Air Canada is instructive here. A Canadian small-claims tribunal held Air Canada liable for its chatbot's fabricated bereavement-fare policy, rejecting the airline's argument that it could not be responsible for its own chatbot's output; the tribunal found the company had a duty of care to ensure the accuracy of information presented through any channel, including an AI system, and awarded damages accordingly [18-19,22]. Trade-press analysis of subsequent disputes — including a 2025 case in which an AI support agent at a software company invented a device-limit policy — suggests courts and regulators are converging on the principle that firms "own" what their AI systems say, regardless of intent [23]. For cross-border trade, this creates a distinctive uncertainty: a hallucinated customs requirement, shipping term, or regulatory claim communicated by an AI system to a foreign buyer or supplier could expose the deploying firm to liability across multiple jurisdictions with different legal standards for negligent misrepresentation, converting a technical error into a cross-border legal-risk premium that functions much like a hidden compliance cost.
Finally, hallucination risk affects the pace and depth of AI adoption in trade facilitation itself. The WTO notes that AI is already used to break down language barriers in sourcing and dispute resolution and to auto-generate first-pass import/export declarations [6]. If firms cannot trust AI-generated translations, tariff advice, or compliance summaries without exhaustive human re-verification, the efficiency gains motivating AI adoption are partially or wholly offset — and the resulting caution itself becomes a form of friction, since firms revert to slower, more resource-intensive manual verification processes that a well-functioning AI tool was meant to replace.
While systematic, large-sample data on hallucination-driven trade disputes does not yet exist publicly, several documented cases illustrate the mechanisms described above.
After his grandmother's death, a passenger consulted Air Canada's website chatbot, which incorrectly advised that he could apply retroactively for a bereavement discount within 90 days of ticket purchase — a policy that did not exist and contradicted the airline's own published rules [18]. Air Canada refused to honour the discount and argued in tribunal proceedings that it could not be held responsible for information generated by the chatbot, effectively treating the AI system as an independent actor. The tribunal rejected this argument outright, holding that a company is responsible for all information presented on its website regardless of whether it originates from a static page or an interactive AI tool, and ordered Air Canada to pay damages and fees [19,14]. Although this dispute concerned domestic consumer travel rather than cross-border merchandise trade, its legal reasoning generalises directly to any customer-facing or partner-facing AI system used in international commerce: firms cannot outsource accountability to the AI system itself, and hallucinated statements about policy, pricing, or eligibility carry the same legal weight as statements made by a human representative.
Major logistics providers have moved quickly to embed generative AI into customs workflows. FedEx's 2025 rollout of an AI-powered customs documentation assistant across several Asia-Pacific markets was explicitly framed as a response to the persistent problem of inaccurate shipping documentation, with each AI-suggested tariff code linked back to the official schedule "to ensure full transparency and verification" [20]. A design choice that implicitly acknowledges hallucination risk and builds in a human-verification safeguard. Trade-compliance technology vendors similarly market AI tools for HS-code classification while cautioning that incorrect classification carries legal and financial consequences [21]. The International Monetary Fund's 2025 technical guidance on generative AI in customs and tax administration likewise frames the technology's value in terms of augmenting, rather than replacing, human risk analysts, reflecting institutional awareness that unverified AI output is not yet reliable enough for autonomous compliance decisions. The World Customs Organization's global survey of AI/ML adoption among customs administrations similarly notes a significant digital divide, with many member administrations lacking the technical capacity to deploy — or to independently audit — AI systems (WCO, n.d.), reinforcing the information-asymmetry mechanism described in Section 4.2.
The WTO's World Trade Report 2025 documents the rapid uptake of AI-powered translation and generative document-summarisation tools in sourcing, after-sales support, and dispute resolution, describing these as improving "first-time-right rates" in import and export declarations [6]. The same report, however, situates this progress within a broader observation that AI adoption remains highly uneven — concentrated in large, urban, digitally connected firms — and that regulatory fragmentation around AI-related goods has intensified, with WTO-notified trade concerns related to AI rising sharply and quantitative restrictions on AI-related goods increasing from roughly 130 in 2012 to about 500 in 2024 [23, 6]. This growing regulatory complexity, layered on top of hallucination risk in the tools meant to help firms navigate it, compounds rather than resolves the compliance burden facing exporters.
Taken together, the case evidence and institutional reporting support the paper's central claim: AI hallucination is beginning to function as a de facto non-tariff barrier, not through deliberate government policy, but through the interaction of legal liability rules, uneven AI adoption, and the structural unreliability of generative AI output. This is a meaningfully different kind of NTB from those classified in the UNCTAD taxonomy. Traditional NTBs are imposed by governments (technical regulations, licensing, inspection regimes) or emerge from market structure (private standards, certification requirements). AI-hallucination friction is distinctive in that it is imposed by neither party to a trade transaction — it is an emergent property of the technology both parties may be relying on to reduce, not increase, trade costs. This makes it harder to regulate through conventional trade-policy instruments such as the TBT and SPS Agreements, which presume an identifiable regulatory measure that can be notified, challenged, or harmonised [2].
At the same time, the mechanisms identified in Section 4 suggest hallucination-driven friction will not be evenly distributed. Firms and economies with stronger legal, technical, and compliance capacity are better positioned to catch and correct AI errors before they cause commercial harm, echoing the WTO's own finding that AI adoption and its benefits remain concentrated among large, well-resourced firms [6]. If unaddressed, this suggests AI hallucination could reinforce rather than narrow the divide between economies well-positioned to capture AI's trade-boosting potential and those exposed mainly to its risks — a concern the World Trade Report 2025 raises about AI adoption generally, which this paper extends specifically to the reliability dimension of that adoption.
The Moffatt v. Air Canada precedent also suggests that liability frameworks are evolving faster than trade-specific regulation. Courts are already assigning responsibility for AI-generated misstatements to the deploying firm, without waiting for AI-specific trade rules. This creates a practical, near-term incentive for exporters, freight forwarders, and platforms to treat AI-hallucination risk as a compliance and legal-exposure issue today, rather than an emerging concern to be addressed once formal regulation catches up.
Several implications follow for firms, regulators, and international institutions.
For firms engaged in cross-border trade, the Moffatt precedent implies that any AI system customer- or partner-facing — chatbots, automated quote generators, AI-assisted customs documentation tools — should be treated as a source of binding representations, with human-in-the-loop verification built into workflows involving pricing, eligibility, regulatory claims, or tariff classification, mirroring FedEx's practice of linking every AI-suggested HS code back to the official tariff schedule [20].
For customs authorities and trade-facilitation bodies, the IMF's emphasis on using generative AI to augment rather than replace human risk analysts offers a workable interim standard: AI-assisted classification and documentation should remain subject to defined human review thresholds, particularly for high-value or first-time shipments, until hallucination rates in domain-specific customs applications are independently measured and published.
For international institutions, the WTO's existing role as a forum for AI-related trade concerns — 80 specific concerns had already been raised at the WTO as of the World Trade Report 2025 [4], could be extended to develop shared standards or disclosure norms for AI reliability in trade-facilitation tools, analogous to the transparency and notification obligations under the TBT Agreement [2]. Given the capacity gaps documented by the WCO among member customs administrations (WCO, n.d.), technical assistance and capacity-building programmes should explicitly include AI-verification capability, not just AI-deployment capability, so that adoption does not outpace the ability to audit it.
Finally, given the uneven adoption documented in the WTO–ICC survey, capacity-building efforts aimed at small and medium-sized exporters should treat AI-verification literacy as important as AI-adoption support, since the smaller firms most likely to benefit from AI-driven cost reduction are also the least equipped to independently catch hallucinated output [6].
This paper is explicitly conceptual and case-based, not a quantitative empirical study, and this is its principal limitation. No large-sample, verifiable primary dataset currently exists measuring the frequency, cost, or trade-volume impact of AI-hallucination-driven friction specifically in cross-border commerce; the case evidence assembled here, while genuine and well-documented, is illustrative rather than statistically representative. Future research should prioritise:
AI is already reshaping the infrastructure of international trade, and institutional forecasts suggest its role will only grow — the WTO projects AI could raise global trade by roughly a third by 2040 if adoption gaps are bridged [4]. But the same technology carries a structural flaw, hallucination, that has already produced a legally significant, adjudicated case of commercial harm in Moffatt v. Air Canada, and that institutional actors from the IMF to the WCO to major logistics firms are actively building safeguards against. This paper has argued that when hallucination's costs — inflated compliance burdens, information asymmetry, liability exposure, and eroded trust in digital trade tools — are examined through the lens of established non-tariff-barrier theory, they meet the functional definition of an NTB, even though no government intends or imposes them. Recognising AI hallucination explicitly as an emerging category of trade friction, rather than treating it purely as a technology-quality issue, is a necessary first step toward the kind of coordinated policy response, capacity-building, and empirical measurement that traditional NTBs have received over the past three decades.
Download
Download XML