Research & Policy Note on Generative

Artificial Intelligence for Customs

November 2023

The Secretariat of the World Customs Organization (WCO) anticipates that many Customs
authorities, in conjunction with both public and private stakeholders in international trade, will
progressively adopt GenAI as part of their systems in the foreseeable future. This assumption
is underpinned by several factors: (i) the diversity of GenAI applications that will supplement
already well-established AI systems within the public sector; (ii) the intuitive interface made
possible by natural language interaction, broadening the scope of end-users; (iii) the extremely
rapid spread of this technology throughout society; and (iv) the reaction of governments and
organizations in working on trust in AI to be achieved, for example, the G7 Hiroshima process1,
the European Union AI Act2, the United States Executive Order on AI3 and the last AI Safety
Summit in United Kingdom4.
Owing to its capacity to generate text, images, video, and audio in response to natural
language prompts, GenAI holds potential implications for professions rooted in intellectual
activities. This technology may lead to the transformation of existing jobs and the creation of
new jobs, which is a trajectory historically observed for all emerging technologies5.
It is therefore important for the Customs community to engage in the discussion on the
future impact of GenAI in public administrations. The objective of this note is to furnish a fair
perspective on GenAI as an emerging trend in technology, balancing its potential benefits,
current specific risks and limits. This note therefore aims to foster informed discussions on
GenAI within Customs administrations, the broader WCO community, and among the technical
and financial partners of Customs authorities worldwide.

1
In 2023, the G7 countries launched the Hiroshima AI Process aimed at establishing global
recommendations for trusted AI, but applied to so-called "advanced" AI (typically GenAI)
(https://www.politico.eu/wp-content/uploads/2023/09/07/3e39b82d-464d-403a-b6cb-dc0e1bdec642-
230906_Ministerial-clean-Draft-Hiroshima-Ministers-Statement68.pdf).
2
The European Union (EU) recently took into account GenAI-specific considerations in the AI Act
(https://www.europarl.europa.eu/news/en/press-room/20230505IPR84904/ai-act-a-step-closer-to-the-first-rules-
on-artificial-intelligence). In April 2023, Italy blocked access to ChatGPT for allegedly failing to comply with
data privacy regulations, before reinstating it.
3
In October 2023, the U.S. issued a comprehensive document embracing many topics related to AI,
including GenAI, defining policy guidelines and announcing the coming works of federal agencies in AI
(https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-
secure-and-trustworthy-development-and-use-of-artificial-intelligence/). In May 2023, the US government had
launched a working group on the limits and risks associated with GAI
(https://www.whitehouse.gov/pcast/briefing-room/2023/05/13/pcast-working-group-on-generative-ai-invites-
public-input/).
4
https://www.gov.uk/government/topical-events/ai-safety-summit-2023
5
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). Gpts are gpts: An early look at the labor
market impact potential of large language models. arXiv preprint arXiv:2303.10130.

1
 It is important to consider the limits of this note that is issued within a rapidly changing
technological environment: this note is based on a July 2023 research publication and the
outcomes of a seminal seminar on “GenAI for Tax and Customs Administrations” held in
October 2023 in partnership with academia6. This note is therefore one resource in the needed
research for any administration considering using GenAI.

The note is segmented into three main sections:

- A succinct overview of the technological principles of GenAI and its limitations that
affect its uses in Customs administrations.
- An exploration of the potential applications of GenAI within Customs, identifying
both existing uses that can be borrowed from other sectors and potential near-
future applications, subject to technological progression.
- A discussion on three strategic aspects for Customs related to GenAI integration:
evolving dynamics of human-machine interaction owing to natural language usage,
essential training modules for officials to maximize GenAI benefits while mitigating
inherent risks, and the critical task of establishing a sovereign training corpus fitting
with Customs needs.

1. GenAI at a glance

GenAI is recognized for its proficient interaction with humans through natural language
(i.e., without using computer language). As a core principle, it is essential to comprehend that
GenAI neither "thinks" nor “reads”, “writes” or “draws”. GenAI has no notion of the intrinsic
"meaning" of what it produces, it operates on language through computational and statistical
methods. GenAI is ultimately based on mathematical prediction of the word that follows to
form texts. Nevertheless, it acquires de facto writing, synthesis and summarizing capabilities.

1.1. Core technical principles of GenAI

GenAI applied to text generation can better help to understand some principles. Large
Language Models (LLM) actually form the core of the most publicized applications of GenAI,
notably conversational agents such as ChatGPT. These new LLM applications are at the
intersection of artificial intelligence, machine learning and natural language processing
technologies.

6
The publication and seminar have been co-organized with the Fondation pour les Etudes et la
Recherche sur le Développement International and its Institut des Hautes Etudes du Développement Durable.
The working paper is available in French and English at https://ferdi.fr/publications/comment-pensera-l-etat-
avec-chatgpt-les-douanes-comme-illustration-de-l-intelligence-artificielle-generative-dans-les-administrations-
publiques. See also, for the seminar, https://ferdi.fr/en/events/generative-artificial-intelligence-genai-in-public-
administrations. A finalized academic paper is under publication (Cantens, 2023 forthcoming, “How Will the
State Think with ChatGPT? The Challenges of Generative Artificial Intelligence for Public Administrations »,
AI & Society)

2
 Like all AI models, GenAI models need to be "trained". Unlike narrow AIs that are
trained on specific data, for example, on Customs data for risk analysis, GenAI is trained on a
corpus of texts large enough to embrace language in its greatest extension. Text corpora exist
and are available in open or commercial sources7.LLMs therefore “learn” how the language
“works”, according to two core principles that are important to take into account to understand
the capacities but also the limitations of GenAI.
Word embedding is a form of text representation that is now commonly used. Every
word is ascribed “weights” - numerical values - allocated to myriad “parameters”. For instance,
the term “dog” is vectorized based on its assigned values linked to parameters such as “living
being”, "animal", "canids", “humans”, “verb”, “plural”, and “noun”. The weights allocated to
parameters define each word's unique characteristics. Each word is therefore transformed into
a vector of numbers (the weights).
Distance and relationships between words: once words are digitally vectorized,
GenAI computes the relationships between words and probabilistically anticipates subsequent
words, including in complex sentences8. This capacity underpins its text generation
capabilities, which are vital for functions such as multilingual translation, summaries or
synthesis of texts.
Then, models are progressively refined, sometimes manually, by analyzing the
responses to improve their accuracy.

Beyond this basic and broad training (pretraining), GenAI can be trained on an
additional and domain-specific corpus of information (fine-tuning). For instance, an additional
training could be on Customs administration texts, texts of laws and regulations, texts defining
rules of origin, value or tariff classification, or descriptions of products or manufacturing
techniques for both legal and illegal products. The field of specific training is as diverse as the
uses of GenAI. While pretraining takes a significant amount of time and computer resources,
fine-tuning can be conducted in a simpler and more cost-effective way at the scale of a
Customs administration. This point will be addressed more specifically in the third section of
this note.

GenAI rests on neural networks that have been present for at least two decades. While
the foundational technologies of GenAI are therefore not entirely new, their contemporary
relevance arises from enhanced processing speeds and reduced data storage costs9. Two
major advances make the current GenAI more successful.

7
ChatGPT's training corpus is still confidential, but it is estimated at several hundred gigabytes of
documents (approximately 600 GB), supplied by WebText (a standard corpus of web pages) augmented by
Wikipdeia pages, academic articles and blogs. This corpus probably represents more than 500 billion words
(https://scoms.hypotheses.org/1059).
8
Providing a technical background of the functioning of a LLM is out of the scope of this paper, but
this process is (more or less) the « T » in ChatGPT, the Transformer technique that ensures the consistency of
the meaning of a sentence ; by associating words that can be distant in a complex sentence. This is called the
« attention mechanism » and refers to a seminal paper in the science of LLMs (Vaswani, A., Shazeer, N.,
Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need.
Advances in neural information processing systems, 30. (https://arxiv.org/abs/1706.03762v6).
9
https://www.zdnet.com/article/chatgpt-is-not-particularly-innovative-and-nothing-revolutionary-says-
metas-chief-ai-scientist/

3
 First, LLMs are "large" in the sense that they are trained on large text corpora and
because they have a very large number of parameters. As an example, ChatGPT 4 is
supposed to have approximately 1.7 trillion parameters; Llama 2 has been pretrained on 2
trillion “tokens” (single language units) and 1 million human annotations. These new scales,
compared with existing models in language analysis, mark the real evolution of LLMs.
Second, LLMs are now often multimodal, able to process, through language, multiple
types of information such as images or audio recordings. They can recognize objects in an
image/photo and analyze their relationships.

GenAI marks a crucial phase in AI development. Its models, termed "general-purpose"

or "foundation models", have no fixed objective. This diverges from the narrow AI models
currently utilized in Customs, which are tailored for specific functions, such as risk analysis
algorithms employed against Customs fraud. Moreover, GenAI interacts in natural language:
we write to it and read its responses. Unlike narrow AIs that return scores and probabilities,
even if it is based on statistical computation of language, GenAI is textual – it interacts through
language – and not numerical, which facilitates, broadens and diversifies human
engagements with the AI but also the risks of misuses (see section 3 on challenges for
Customs).

1.2. Current trends

Current GenAI trends reflect the rise of a GenAI-specific ecosystem. Two primary
trends have emerged, common to any new applications in the modern IT landscape:

1. Commercial vs. open-source solutions. The first trend is the partitioning of the
GenAI sector between a commercial offering, which is high-performance but whose
technical characteristics remain opaque, and an open-source offering, which is
initially probably inferior in terms of performance but benefits from global
communities of developers and should, therefore, perform equally or better than
the commercial sector. ChatGPT, for example, is maintained by OpenAI and offers
a subscription package for an optimal version, as well as dedicated services for
major operators. On the open-source side, for instance, Llama 2 is made freely
available to the public by Meta10.

2. Specialized applications. The second trend is the development of small

applications (plug-ins, or others via APIs) for major models, which are dedicated to
facilitating certain tasks, such as reading and summarizing large PDF documents,
or interfacing with other tools, such as combining GenAI code writing and GitHub
facilities, or GenAI and academic paper research. These plug-ins may be freely
accessible or fee-based. A range of IT services using GenAI is also being
developed.

1.3. Current limits of GenAI

10
These are examples only. The WCO Secretariat does not recommend any particular model.

4
 In essence, GenAI translates natural languages into concepts and probabilities,
displaying both generalist and domain-specific capabilities. One can easily "dialogue" with a
GenAI agent by conveying requests (prompts) and refining its responses. However, this ease
of use may obscure the limitations and risks that have to be taken into account before any
deployment. By design, GenAI has three specific limitations.

Explainability: the ability to understand the algorithm's path to its result. The
explainability of GenAI is currently elusive. Since GenAI is based on neural networks11, it is
impossible, even for its designers, to explain how the algorithm produces a specific result.
Explainability is a decisive feature for the choice of algorithms in public administrations12.
Explainability is also a necessity in the event of AI failing to respect ethical values: when AI
outputs cannot be explained, it is all the more difficult to understand where its shortcomings
come from. In Customs, explainability is often one of the criteria for choosing algorithms when
the administration is legally obliged to explain its decisions, which is why algorithms based on
decision trees, for example, have been so popular for selectivity.
Research into explainability is progressing, and attempts are being made to dissect
GenAI’s neural networks to explain them "in parts"13. However, at this stage, the lack of
explainability restricts GenAI's role largely to advisory and assistance functions rather than
being a centerpiece in decision-making processes for which accountability is a legal or user-
driven necessity.

Hallucination vs. bias. GenAI, akin to narrow AI, is susceptible to biases originating
from two human sources. The first source is the training data bias. AI training data might mirror
human prejudices, which, when processed by AI, can be amplified and expedited. For
instance, if Customs officials usually overlook certain stakeholders based on nonexplicit,
subjective and common criteria, AI may perpetuate this oversight. The machine cannot be
reproached for being "wrong" but for doing as badly as humans, faster and more consistently.
The issue of bias is well known in Customs, particularly in risk analysis algorithms. The second
source of bias is design-induced bias. Bias can be inadvertently incorporated by AI designers,
either through racially biased parameters or by disregarding the implementation context.
However, in the case of bias, it is paramount to remember that all AIs are subject to similar
biases and that these biases emanate from human inclinations and are not innate machine
flaws.
Compared to narrow AIs, GenAI has a unique limitation termed "hallucination", where
it produces a mistake in that it references nonexistent information. For example, there is a
“famous” case in which a lawyer used ChatGPT to prepare his closing argument, and
ChatGPT returned very precise references to case law that simply did not exist14. Any user of
GenAI will experience hallucinations that manifest as references to nonexistent information or

11
Jovanovic, M., & Campbell, M. (2022). Generative Artificial Intelligence: Trends and Prospects.
Computer, 55(10), 107-112.
12
Lasmar Almada, M. A., Gorski, Ł., Kuzniacki, B., Tylinski, K., Winogradska, B., & Zeldenrust, R.
(2022). Toward eXplainable Artificial Intelligence (XAI) in tax law: the need for a minimum legal standard.
World tax journal, 14.
13
https://www.anthropic.com/index/decomposing-language-models-into-understandable-components
14
https://www.nytimes.com/2023/06/08/nyregion/lawyer-chatgpt-sanctions.html

5
false attributions. These errors stem from the GenAI’s design and natural language generation
models15, and they seem to be unavoidable, even if it is expected that their frequency should
decrease over time. The unpredictable nature of these errors and their potential impact
therefore necessitates rigorous verification by human officials.
GenAI’s hallucinations pose three challenges. First, they are not systematic. The
stochastic nature of GenAI makes hallucinations fully random. Second, hallucinations are not
quantitatively detectable through any ceiling of certainty included in GenAI’s responses: GenAI
produces answers with the same degree of certainty, unlike narrow AI, which is numerical and
therefore provides scores and confidence intervals for its answers. Third, the impact of
hallucinations could be important due to the inherent human tendency to follow AI suggestions
when embedded into a decision-making process16. This is an additional reason why GenAI
should be envisaged only as an assistant and not be embedded into decision-making at the
moment, without strict and systematic human intervention. Any use of GenAI that would make
it autonomous should be excluded from the scope of possible uses by Customs. All its
productions, whether analytical or synthetic, must be verified by an official.

Reproducibility: GenAI may offer varied responses to the same question due to its
stochastic nature. If the same question is repeated, the meaning of the answers does not
change, but the expression can. The lack of reproducibility may put the administration at risk
if GenAI is entitled to directly respond to the general public. Even if the substance of the
answers does not differ, their clarity and accuracy could be affected, resulting in a breach in
the equality of users before the administration. For officers using GenAI as an assistant only,
the problem may not be legal but requires much vigilance. Indeed, GenAI systems variously
give both correct and incorrect answers when the same question is asked multiple times.

2. Uses of GenAI for Customs

As of the publication of this note, the Secretariat is not informed of Customs

administration that has integrated GenAI, with some exceptions in Customs and Tax
administrations or private companies providing services to Customs, currently testing GenAI17.

15
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., ... & Fung, P. (2023). Survey of hallucination in
natural language generation. ACM Computing Surveys, 55(12), 1-38.
16
Alon-Barkat, S., & Busuioc, M. (2023). Human-AI interactions in public sector decision making:
"automation bias" and "selective adherence" to algorithmic advice. Journal of Public Administration Research
and Theory, 33(1), 153-169.
17
India Customs is experimenting LLMs for real-time Harmonized System classification inconsistency
identification (WCO Asia/Pacific Customs News September 2023, pp. 38-40). A private company providing risk
and trade analysis platform for Customs showcased the use of GenAI during the October 2023 session of the
WCO Data Innovation Hub in Brussels (see also https://altana.ai/blog/altana-and-dataminr-form-partnership-to-
bring-ai-powered). French government launched a public experimentation on GenAI to support officials
responding to users (https://www.modernisation.gouv.fr/presse/stanislas-guerini-experimente-lintelligence-
artificielle-generative-dans-les-services). The French Tax administration also announced it is launching internal
experiments of GenAI for investigation purposes (https://acteurspublics.fr/articles/comment-la-dgfip-negocie-le-
virage-de-lintelligence-artificielle-generative).

6
 At the national level, several governments have established dedicated entities to
consider GenAI implications and applications. Of particular note, the Singapore government,
in collaboration with a private company, intends to provide an assistant to all its civil servants
that centralizes numerous applications across its various administrations18.

This section provides an overview of the potential uses of GenAI by Customs. It takes
into account the limitations mentioned in the previous section: GenAI is not integrated into any
decision-support process for which the administration should be accountable and is only
envisaged in assistance functions.

2.1. Uses derived from other domains

Table 1 presents an inventory of potential applications derived from existing uses in

other sectors or those explored by researchers and the Secretariat. Detailed references can
be found in the Appendix.

Customs – a new generation Current customs chatbots are limited by expert

public of chatbots systems and dictionaries. GenAI makes interaction
relations more intuitive and more easily multilingual.
One risk often mentioned is that the user may not be
able to know whether he or she is interacting with a
human agent or a GenAI agent.
This risk can be addressed both legally and
technically. Some national or regional legislation, such
as the European Union's AI Act, requires GenAI to
include forms of marking that enable anyone to verify
the human or machine source of a discourse (letter,
text, etc.). Technical solutions known as watermarking
are being explored.
The administration must also assess the importance
of identifying the interlocutor according to whether the
response provided to the user is legally binding or not,
the potential consequences of the use of any
information given, the likelihood of errors, and the
expectation of authority and reliability. For example, it
may be tolerable legally for the user of a chatbot on
tariff classification to be informed that he is
communicating exclusively and systematically with an
artificial agent whose responses have no legal force.
However, as misclassification can cause significant
financial harm to the user, a significant error rate can
be expected for products not specifically covered in

18
https://www.smartnation.gov.sg/media-hub/press-releases/31052023/

7
 the training material, and that users of an official
Customs site expect a very high level of authority and
reliability, the above disclaimer may well be
insufficient to protect the administration from claims or
reputational risk.
communication Automatic writing of articles and social network
assistance messages.
reading Classification of texts/messages/articles according to
assistance their positive or negative character, applicable for
example to the reading of responses to
questionnaires or the analysis of user opinions.
Conception writing Writing, summarizing, correcting spelling and syntax,
assistance improving style and bringing it into line with
professional usage, taking notes during conferences
and meetings.
assistance with Production of summaries on a legal field or issues,
text research and based on a multilingual body of knowledge.
analysis
assistance with Automatic writing of computer code, implementation
digital data of machine learning models, visualization of statistical
analysis parameters.
project Choice of project monitoring methods, answers to
management problems during the project.
assistance
negotiation Adopt different points of view, discuss from critical
assistance perspectives.
Training profile selection Generation of job descriptions, trainer profiles,
assistance automatic reading of CVs.
training Proposal of training plans, training content, adaptation
assistance of training to the individual level of participants.
Investigation assistance in Automated collection and representation of evidence
and collecting digital collected on digital storage
intelligence evidence
analysis Intelligence fusion (open sources, minutes, notes,
assistance etc.), representation of large bodies of information in
the form of "knowledge graphs", interrogation of
bodies of information, synthesis, drafting of minutes
and investigation reports.

Table 1. Uses tested and applicable in a Customs context

2.2. Exploratory uses

Within the realm of exploratory applications, we identify two kinds of potentialities,

albeit speculative but based on existing and tested GenAI capacities.

8
 GenAI to be an interface between civil servants and narrow AI or paired with expert
systems to enhance user-machine interaction through natural language processing. Within
Customs, possible applications might include GenAI support for tariff classification search
engines, enabling the facilitation of synonym use, multilingual capabilities, and interaction on
the relationships between classification options and associated regulatory texts, all of which
would contribute to progressively refining the classification of a product during the
conversation with the AI. For risk analysis, GenAI might serve as an interface between civil
servants and algorithms, helping understand risk types, correlating with past cases, or
answering any queries a civil servant might encounter when processing machine-identified
risks. In all such scenarios, the amalgamation of Customs documents, such as investigative
reports with regulatory texts, and the possibility of interacting with an AI putting this knowledge
at disposal through natural language would streamline the flow of information for Customs
officers, assisting them in drawing connections between suspected fraud and previously
identified instances.

GenAI to cohesively link text and images. GenAI is already capable of describing
an image. In Customs, narrow AI is used to detect only certain kinds of illegal goods in X-ray
images but cannot “read” and “compare” the textual description of goods provided by the
importer and the image provided by the scanning machine. GenAI now makes it conceivable
to compare images and texts related to Customs operations: between images taken from
photographs of goods or container scans, on the one hand, and on the other the textual
descriptions of the same goods provided in the manifests and tariff classifications of the
Customs declarations. It may therefore be possible to detect anomalies, or to assess the
vagueness of the descriptions of the goods in the cargo manifests as a risk criterion, possibly
linked to the assessment of the homogeneity of the cargo as assessed by the AI on the
scanner image.

GenAI is suited for assistance rather than decision-making. GenAI cannot assess the
robustness of its results in terms of probabilities, as narrow AI does. Setting predefined
acceptability thresholds for its outputs, as done for narrow AI, is therefore much more difficult.
However, its uses as an assistant present immense potential for administrative tasks and may
renew civil servants' interest in their work.

2.3. Benefits

Three main benefits are expected from GenAI.

Cost efficiency. Using GenAI could lead to significant reductions in governmental

operational expenses by internalizing tasks traditionally outsourced, such as editorial
assistance, communication material creation, and translation. Regarding data analysis, which
generates many difficulties for Customs that need to hire specialists, the operational costs of
GPT-4 are between 0.45% and 0.70% of that of a data analyst, depending on their expertise19.

19
Cheng, L., Li, X., & Bing, L. (2023). Is GPT-4 a Good Data Analyst? arXiv preprint arXiv:2305.15038.
https://arxiv.org/pdf/2305.15038.pdf.

9
However, mobilizing GenAI for data analysis should not replace Customs staff in the short or
mid-term; GenAI would only “democratize” basic data analytics for Customs officials.

Enhanced analysis. GenAI can increase the quality of analyses conducted by civil
servants. It provides access to a vast, multilingual corpus of knowledge for public policy,
enables comprehensive document analysis for intelligence specialists, and facilitates
multilingual dissemination of administrative stances at both regional and international levels.
Access to a multilingual corpus of knowledge and capacity to write in his or her mother tongue
will be an important benefit for any Customs officer, particularly in non-English speaking
countries. However, as previously discussed in section 1, skilled officers will still be needed to
make profit from interacting with GenAI and monitoring its responses. Moreover, the
representation of languages in the training corpus of publicly available GenAI agents varies
among languages, and the risks of errors and mistranslations increase for languages poorly
represented in the training corpus20.

Consistency. GenAI could contribute to ensuring both semantic and stylistic

uniformity in administrative documents.

While efficiency gains are often highlighted, they should be assessed judiciously.
Although document quality improvements might expedite hierarchical validations, civil
servants using GenAI might not consistently experience time savings. They might be tasked
with additional responsibilities and higher quality expectations. Furthermore, the influx of
accessible information and improved analytical capacities might legitimate more time invested
in research or preliminary investigations. Finally, as discussed in the subsequent section
addressing GenAI limitations and risks, civil servants might need to allocate specific time to
verify AI-generated content.

2.4. Costs

Evaluating the operational costs of GenAI’s implementation is complex due to the

variety of its applications and deployment techniques, ranging from centralized national
approaches, as seen in Singapore, to deployments on individual laptops using open-source
models.
However, GenAI's relative autonomy from prevailing administrative IT systems might
reduce its installation expenses compared to narrow AI. GenAI could be installed, for relatively
basic assistance uses, as a more or less autonomous system within the Customs IT system,
while narrow AI usually needs to be integrated into the existing IT system, which often raises
many difficulties. WCO experts noted a recurring challenge faced by Customs administrations
in scaling up AI, transforming the experimentation in labs into full deployment in the
administration. This problem is common to narrow AI in all organizations21. Given its relative

20
Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., ... & Fung, P. (2023). A multitask,
multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv
preprint arXiv:2302.04023.
21
Davenport, T. H., & Ronanki, R. (2018). Artificial intelligence for the real world. Harvard business
review, 96(1), 108-116.

10
autonomy, GenAI should, at least initially, benefit from more straightforward administrative
integration. Notably, numerous private companies have rapidly integrated their GenAI
systems. This consideration is crucial for Customs data and IT strategies that should
distinguish the strategy for GenAI from that for narrow AI.

In essence, GenAI offers the potential for diminished operational expenses and
improved analytical quality on both strategic and operational fronts. Its time-saving potential
for civil servants, however, warrants further exploration. The integration strategy for GenAI,
especially within the realm of Customs data and IT strategies, necessitates distinct
consideration from that for narrow AI.

3. Challenges for Customs

3.1. Impact of GenAI on officials

GenAI is likely to lead to a redefinition of tasks and introduce new ways of working for
many civil servants. Administrations will have to promptly evolve their perspective on the
interaction between AI and civil servants, taking into account the large number of civil servants
impacted by GenAI.

In terms of organization, GenAI will undertake specific tasks, easing the burden on
certain positions (communication, analysis, translation, etc.). Some civil servants will need to
integrate GenAI as an assistant, a supplementary tool handling tasks previously performed by
their peers. For instance, officers could be tasked to draft communication articles by
themselves with the assistance of their “personal” GenAI or create training materials from the
beginning to the end. GenAI could also democratize data analysis for a larger number of non-
data-scientist Customs officers, giving more capacities to officials and amplifying their
analytical capacities. Taking advantage of GenAI should therefore require an adaptation of
data access policies in Customs administrations.

At an individual level, there are risks of “anthropization”: civil servants might either
overvalue the machine’s outputs due to its systematic and “mathematical” nature or, on the
contrary, harbor excessive distrust22. Utilizing GenAI therefore accentuates the importance of
cultivating new skills, especially critical thinking. Beyond their critical vigilance to detect GenAI
hallucinations, civil servants will have to demonstrate their added value in the intellectual
production process. They will have to learn to think both with and against GenAI and consider
it a sparring partner in their analyses. If this condition of critical vigilance and thinking could
be achieved, GenAI could foster a rejuvenated interest in the analytical responsibilities of civil
servants, be it for policy, strategic or operational purposes. This could also help to position
Customs as advisors to their governments, as advocated by the WCO's data strategy.

22
Researchers have demonstrated significant perception biases when productions are provided by AIs,
both positive and negative biases. See Longoni, C., Fradkin, A., Cian, L., & Pennycook, G. (2022, June). News
from generative artificial intelligence is believed less. In 2022 ACM Conference on Fairness, Accountability, and
Transparency (pp. 97-106) and Longoni, C., Cian, L., & Kyung, E. J. (2023). Algorithmic Transference: People
Overgeneralize Failures of AI in the Government. Journal of Marketing Research, 60(1), 170-188.

11
However, it should be noted that high-level critical thinking skills are often a highly sought-after
attribute in employees, so the prevalence of the necessary skill levels to achieve this high-
functioning use described above in a particular workforce may be a question to consider.

Using GenAI, complemented with data science, fortifies the analytical capabilities of
Customs, at all levels. However, this opportunity will only be fully profitable with the
development of specific qualities for civil servants: better statistical abilities to formulate
precise questions to GenAI and vigilance and critical thinking to check and bring added value
to GenAI outputs. This development of critical faculties relies on specific training and a
conducive and supportive managerial environment.

3.2. An urgent training imperative

The Secretariat observes that prohibiting Customs officers from accessing online
GenAI agents, which are already proficient in many outlined tasks, might be difficult, if not
counterproductive. Pending the implementation of sovereign GenAI by states, the Secretariat
alerts Members to two primary risks associated with online available GenAI agents.

Confidentiality breaches. Private companies operating GenAI agents on the Internet

leverage user-IA interactions to refine their models. Through direct negative assessments
provided by users, these companies aim to prevent the risk of disseminating illegal content
(glorification of criminal acts, dissemination of hateful or racist content, advice on acquiring,
consuming or manufacturing illegal products, etc.). Furthermore, the users’ evaluations are
used to improve the quality of the training corpus by enriching it with the best semantic,
grammatical and stylistic interactions between the user and machine. Users’ conversations
with GenAI agents therefore represent a capital of knowledge necessary to GenAI companies.
Although users can opt to delete their interactions, it is impossible to guarantee that this is
actually the case or that these companies are not sharing information with the states that host
them. For Customs, this poses dual threats: divulging sensitive data during interactions and
inadvertently exposing administrative strategies and core interests.
Infringement of the intellectual property of documents. Sharing documents with
GenAI, directly or through plug-ins, for various tasks (reading assistance, synthesis,
summary…) might compromise their ownership, even if they are nonconfidential. The
Secretariat therefore invites Members to scrutinize the terms of sharing documents with public
GenAI agents. In some cases, any shared document is likely to be used by the company
managing the GenAI agent and therefore becomes, in a way, its property.
These risks, linked to the use of online GenAI agents, are not particularly new. They
are merely a transposition into a new context of the risks usually encountered on the Internet
in general. Customs officials are already bound by confidentiality and discretion duties
appropriate for these risks. The administrations’ focus should therefore be on enhancing
awareness and training rather than introducing new rules.
.
When governments roll out sovereign GenAI for their administrations, these
confidentiality and intellectual property risks will be resolved. However, other challenges linked
to AI limitations, as developed above (mainly, lack of explainability, hallucinations and

12
reproducibility, as discussed in section 1.2), will persist. Misunderstanding these limitations
could have various repercussions, including errors in public policy analyses, poor technical
choices, damage to the administration's reputation, unequal treatment of users, or
investigations invalidated by the courts.
Given the widespread potential application of GenAI, the large volume of civil servants
impacted by GenAI should be a primary concern for administrations. The deployment policy
of GenAI dramatically differs from the deployment of narrow AI intended for a limited group of
civil servants.

The Secretariat therefore stresses the importance of establishing national guidelines

for GenAI usage, reiterating online confidentiality and discretion rules. The Secretariat also
recommends that Members educate civil servants about the opportune and efficient use of
GenAI, especially fostering their creative and critical faculties.

Table 2 below lists all the risks identified to date, both in relation to the use of generative
AI on the Internet and to a misunderstanding of the current limits of GenAI.

13
 Risk area Risks Effects Answer
Privacy Conversations with generative AI agents on the Violation of confidentiality Do not share any individual, nominative
Internet can be stored by the companies of individual data data.
deploying them. (company names,
addresses, etc.). Estimate the sensitivity of the subject to
be dealt with before sharing it with
Disclosure of the generative AI (ask yourself whether you
administration's strategic would organize a public conference on
interests to third parties. this subject in your administration).
Errors Like narrow AI, GenAI is subject to biases Production of biased, Check all results and sources. Do not
contained in the training corpus. In addition, erroneous analyses and assume that the machine can be only
GenAI produces "hallucinations", such as facts, dissemination of unverified biased.
data and references that do not exist. facts.
Hallucinations are not systematic (the same
question asked several times may generate a
hallucination just once).
Document It is possible to share entire documents with Loss of exclusive Estimate the sensitivity of the document
ownership GenAI, for further processing or analysis. ownership of the before sharing it.
However, some GenAI companies include in document.
their legal provisions that they can use any
document shared by the user for their own
purposes.
Plagiarism GenAI can propose answers that paraphrase Damage to the reputation Use anti-plagiarism software.
authors or documents without quoting them. of the administration.
Cite sources in all texts produced by the
administration.

14
Temporal validity Given the high cost of training GenAI agents - Failure to take account of Complement the conversation with GenAI
of information for the most efficient of them - training cannot be the latest facts, contexts, by using specialized search engines
launched over short periods of time. Information scientific advances or (academic, press, etc.).
accessible mobilized by GenAI has a temporal legal frameworks in GenAI
limit. analysis. Use GenAI agents that complement their
analyses with Internet searches.
Sources Some GenAI agents do not provide sources for Errors in the information Check specialized search engines
the ideas or analyses they report in sources provided by
conversations, or they provide false ones GenAI during a Use plug-ins adapted to academic
(hallucinations). conversation. research that provide the actual sources
of ideas and papers.
Reproducibility Since GenAI is a stochastic process, it cannot Potential for unequal Evaluate the sensitivity of accuracy in the
produce exactly the same answers, textually or treatment of users (to be responses expected from GenAI when
visually, to an identical request. studied on a case-by-case GenAI may be in direct contact with the
basis) public.
Explainability GenAI provides answers that cannot be Inability to account for Do not use GenAI autonomously in
explained. It relies on artificial intelligence administrative decisions. decision-making processes.
technologies (neural networks, deep learning)
which do not allow to know precisely the Difficulty detecting the
sequence of computer operations leading to the causes of unequal
result. treatment by the
administration.

Table 2. Risks, limits and possible responses on which to train civil servants.

15
 3.3. Building a training corpus

In the current digital landscape shaped by GenAI, the strategic emphasis for
government agencies will shift from the choice of algorithms to the building of the training
corpus. It will become imperative for government entities to develop specialized training
corpora tailored to their unique requirements. Such a corpus should ensure that the GenAI
agents' responses are aligned with the administrations’ missions and core topical interests, as
well as consistent with administrative terminologies.

For instance, it has been suggested that in the realm of tariff classification, a dedicated
training corpus would encompass regulatory texts pertinent to the Harmonized System (HS),
the nomenclature, legal notes, explanatory notes and classification opinions, as well as
national internal documents issued on classification cases, records of misclassification fraud
cases, or even product specifications. However, it should be noted that such a corpus is highly
unlikely to be sufficient to achieve consistently good results in an automated way. The nature
of current GenAI means it does not “classify” in the sense of applying the HS’s legally binding
rules, but instead it relies on having a corpus associating goods with classification and the
above list would only contain a fraction of the possible descriptions of goods. Broadening it to
include all declaration data would mean a greater range but would be training the system on
data with a very high probability of significant error rates in classification, depending on the
performances of the Customs officers in classification, and would require significant
reinforcement learning from human feedback. In the same way, updates to the HS or the
national tariff would pose great difficulties as it would alter classifications from what the bulk
of the material learnt for affected products.

In areas where the possibilities are greater, a broad perspective to aid policy analysis
should be considered, and the training corpus should incorporate contemporary strategic or
innovative topics for the administration, including press articles, research papers, and
essential administrative texts (both public and internal).

Current advancements in GenAI allow for the fine-tuning23 of such corpora based on
specific needs. GenAI companies already offer customization to align the model with user
requirements. Tailored solutions have emerged in sectors such as the legal domain24 and

23
OpenAI already has this online offer to refine its own conversational agent, which of course comes at
a cost https://platform.openai.com/docs/guides/fine-tuning
24
See https://www.reuters.com/technology/bar-exam-score-shows-ai-can-keep-up-with-human-
lawyers-researchers-say-2023-03-15/
See also Harvey for lawyers https://www.sequoiacap.com/article/partnering-with-harvey-putting-llms-
to-work / and current projects https://www.pwc.com/gx/en/news-room/press-releases/2023/pwc-announces-
strategic-alliance-with-harvey-positioning-pwcs-legal-business-solutions-at-the-forefront-of-legal-generative-
ai.html

16
within private enterprises providing writing assistance tools to their employees25. Notably, such
adaptability is a standard feature across all GenAI models, even within open-source solutions.

An added advantage of fine-tuning the training corpus is the potential creation of

models with fewer parameters. Such models might be less resource intensive, facilitating local
implementations, shortening training periods, and enabling more frequent training and
updates.
At the national level, GenAI therefore causes a paradigm shift from algorithms to a
training corpus. With narrow AI, training data were cost-free, derived from existing databases
and the IT Customs clearance systems. The development of training corpora for GenAI now
represents a strategic investment for government agencies, both in terms of human and
financial capital.

At the international level, training corpora should gain even more prominence, not only
economically but also politically and ideologically. Constructing these corpora necessitates
discerning content selection, which in turn defines an established "truth" in knowledge. From
a geopolitical lens, these corpora might emerge as new avenues for political influence, when
a more advanced country shares its training corpus with a less advanced one.

Given the ongoing discussions around the establishment of sovereign GenAI, the
Secretariat emphasizes the need for Members to proactively assess the organizational impact
of GenAI, so that Customs challenges and missions are taken into account in national
discussions. Initiating GenAI projects promptly will allow civil servants to better envision their
future and be prepared for their evolving roles.

4. Conclusion

While early adoption of GenAI by Customs can potentially provide major successes to
become more data-driven, it can just as easily be a failure, particularly in reputational terms.
Judging how and when to use GenAI requires a good understanding of the technology, its
capabilities, strengths and weaknesses, and the prerequisites to achieve the promised
outcomes. It is therefore important to explore the risks and consequences of GenAI for
Customs organizations, officials’ interaction with AI and relationships with the rest of society.
Research, experiments and knowledge sharing through continuous training of Customs
officials are crucial to make technology more performant and civil servants well prepared for
its appropriate uses.
Technology and innovation is one focus area in the WCO's strategic plan, and the
Secretariat will continue supporting its Members, by developing knowledge on GenAI, its
applications and its appropriation in Customs, based on research and in cooperation with
Members and academia.

25
https://asia.nikkei.com/Business/Companies/Panasonic-unit-deploys-ChatGPT-style-AI-to-improve-
productivity

17
 5. Appendix. Bibliographical references on uses.

replacing French initiative launched for civil servants responding to users to use
chatbots generative AI26 .
writing assistance Babl, F. E., & Babl, M. P. (2023). Generative artificial intelligence: Can
ChatGPT write a quality abstract?". Emergency Medicine Australasia.

Huang, J., & Tan, M. (2023). The role of ChatGPT in scientific

communication: writing better scientific review articles. American
Journal of Cancer Research, 13(4), 1148.

Herbold, S., Hautli-Janisz, A., Heuer, U., Kikteva, Z., & Trautsch, A.
(2023). AI, write an essay for me: A large-scale comparison of human-
written versus ChatGPT-generated essays. arXiv preprint
arXiv:2304.14276.

Jiao, W., Wang, W., Huang, J. T., Wang, X., & Tu, Z. P. (2023). Is
ChatGPT a good translator? Yes with GPT-4 as the engine. arXiv
preprint arXiv:2301.08745.

Pu, D., & Demberg, V. (2023). ChatGPT vs Human-authored Text:

Insights into Controllable Text Summarization and Sentence Style
Transfer. arXiv preprint arXiv:2306.07799

Xames, M. D., & Shefa, J. (2023). ChatGPT for research and

publication: Opportunities and challenges. Available at SSRN 4381803.

Yang, X., Li, Y., Zhang, X., Chen, H., & Cheng, W. (2023). Exploring the
limits of chatgpt for query or aspect-based text summarization. arXiv
preprint arXiv:2302.08081
assistance with Opdahl, A. L., Tessem, B., Dang-Nguyen, D. T., Motta, E., Setty, V.,
text research and Throndsen, E., ... & Trattner, C. (2023). Trustworthy journalism through
analysis AI. Data & Knowledge Engineering, 146, 102182

Pierce, N., & Goutos, S. (2023). Why Law Firms Must Responsibly
Embrace Generative AI. Available at SSRN 4477704.
assistance with Cheng, L., Li, X., & Bing, L. (2023). Is GPT-4 a Good Data Analyst?
digital data arXiv preprint arXiv:2305.15038. https://arxiv.org/pdf/2305.15038.pdf
analysis
https://ieg.worldbankgroup.org/blog/fulfilled-promises-using-gpt-
analytical-tasks

26
https://www.gouvernement.fr/upload/media/content/0001/05/0a63326de41e9a36d7878966030e3ce55
e98bddf.pdf

18
project Minelle, F., & Stolfi, F. (2023). AI to support PM: a ChatGPT quality
management assessment (ß test). URL: https://pmworldlibrary.net/wp-
assistance content/uploads/2023/05/pmwj129-May2023-Minelle-Stolfi-AI-to-
support-PM-a-ChatGPT-quality-assessment-3-1.pdf
reading https://ieg.worldbankgroup.org/blog/fulfilled-promises-using-gpt-
assistance analytical-tasks
assistance in Rodrigues, F. B., Giozza, W. F., de Oliveira Albuquerque, R., & Villalba,
collecting digital L. J. G. (2022). Natural language processing applied to forensics
evidence information extraction with transformers and graph visualization. IEEE
Transactions on Computational Social Systems.

Henseler, H., & van Beek, H. (2023). ChatGPT as a Copilot for

Investigating Digital Evidence. https://ceur-ws.org/Vol-3423/paper6.pdf
survey assistance Graham, S., Yates, D., & El-Roby, A. (2023). Investigating antiquities
trafficking with generative pretrained transformer (GPT)-3 enabled
knowledge graphs: A case study. Open Research Europe, 3, 100.

19