GAI is a branch of artificial intelligence that uses machine-learning algorithms trained on data to generate content (images, video, text, etc.) autonomously. The ChatGPT[2] conversational tool is the best-known example (GPT stands for generative pre-trained transformer). This chatbot, developed by OpenAI and first made available to the general public for testing in November 2022, provides answers to questions by analysing large quantities of previously viewed textual data, known as large language models (LLMs).

The “language-centric” nature of GAI distinguishes it from other exclusively “number-centric” forms of AI that have, to date, been deployed in the public administration sector. The paper therefore raises questions about the implementation of GAI-based solutions in public administration: what are the specific benefits, risks and limitations of GAI, and what effects will it have on the relationship between officials and machines?

The author begins by analysing the potential specific benefits of GAI for Customs administrations: machine translation, creation of computer code for data analysis, uniformity and correctness of administrative language, content creation, assistance in drafting policy documents and assistance in fraud detection.

If such benefits were to be realized, it would first be necessary to develop a training corpus specific to Customs in order to ensure that the GAI responses are as accurate as possible. A fine-tuning process is furthermore offered by a number of GAI companies, making it possible to customize the LLM according to the user’s needs.

If this corpus were to be shared, it could lead to an increase in the quantity and quality of the analyses performed by Customs officers, not least because they would be able to draft reports in their native language. The dissemination of this information in other languages would help eliminate any disparities in intellectual production between states and provide all civil servants with access to the same corpus of knowledge, regardless of the language in which it is originally produced.

The author then attempts to deconstruct the underlying reasons for the widespread mistrust of GAI. He explains, for example, that, while the use of AI in decision-making commits the administration in the form of a decision for which the latter must be able to provide an explanation, GAI is limited to an advisory or assistance role in thinking and writing. Another question is: should people be informed that they are conversing with a GAI agent in the context of their relations with a public administration? After all, civil servants are expected to act in an impartial, objective and rational manner, not unlike a machine, and people expect an answer to be binding on the administration, whether it comes from a human or a machine. Accordingly, it is the responsibility of the administration itself to determine the legal value of the machine’s answers.

The author then proceeds to review the technical limitations of this conversational tool: the GAI-driven conversational agent does not quote its sources, it draws its responses from a corpus of information that is subject to time limitations, it does not always respond in the same way to the same queries, and it makes mistakes. Although ways of overcoming these limitations have been found, one important constraint on the use of GAI by Customs is perhaps more difficult to resolve: the need to preserve the confidentiality of the information and workflows of the administration. Some of the corpus and many of the questions asked by officials in sensitive areas are confidential. The solution for administrations lies in internalizing GAI and fine-tuning LLMs as described above. Nevertheless, the author does warn that, while sharing data with other departments can have its advantages, it also raises specific concerns – especially about protecting the confidentiality and security of data – and, in some cases, may even be prohibited.

The final part of the paper explores what thinking with the assistance of a language-based GAI tool would mean in the Customs context, as well as the question of whether GAI will replace Customs officials. In response to this last question, the author reassuringly maintains that, strictly speaking, GAI cannot be used to replace civil servants. It is more a question of officers having to incorporate into their day-to-day work tasks that would previously have been outsourced. This will require civil servants to work a little more autonomously, to do their own translations and proofreading with the assistance of GAI and to focus their efforts on specific parts of the thought process. Officers will need to improve their writing skills in order to meet the specific objectives of their administration, cultivate their critical vigilance and develop their critical thinking, creativity and ability to think outside the box, particularly with regard to issues where the development of a strategic vision is a highly complex and detailed process.

Despite the many concerns raised by the use of GAI, along with its technical limitations and the uncertainty over the structure of its economic model, GAI is set to become, in the not too distant future, a routine work tool, integrated into professional practices in the same way as the internet and search engines. To maintain control of the machine and ensure that it helps Customs officers to become better experts, Customs administrations will need to increase their appetite for knowledge and their officers’ ability to conduct a critical examination of the written material produced.

More information
https://ferdi.fr/en/publications/how-will-the-state-think-with-the-assistance-of-chatgpt-the-case-of-customs-as-an-example-of-generative-artificial-intelligence-in-public-administrations

[1] https://ferdi.fr/en/publications/how-will-the-state-think-with-the-assistance-of-chatgpt-the-case-of-customs-as-an-example-of-generative-artificial-intelligence-in-public-administrations

[2] https://openai.com/blog/chatgpt