Last October, Stanislas Guerini, Minister of Transformation and Public Service, launched the first experiment in the use of generative artificial intelligence in public services. For two months, around a thousand volunteer agents have used AI to write responses to users’ online reviews and comments.
The first results are proving conclusive for both agents and citizens. In public services that volunteered, one response in two is facilitated by AI and the average response time fell from 7 days to 3. 74% of users say they are satisfied with the response provided and 70% of agents have a positive feeling about the tool.
However, there is no question of automating everything. The responses generated by artificial intelligence are completed and personalized by the agent. This frees up time that he can devote to “ to responses that require specific attention. » The experiment is supervised by the Interministerial Directorate for Public Transformation (DITP) and the Interministerial Digital Directorate (Dinum).
Building on this initial assessment, the ministry intends to accelerate. “ Artificial intelligence holds the promise of major transformations, he believes in a press release. The civil service, France’s largest employer, is using this technology to make it available to its agents in order to strengthen the quality of support for the French. »
Albert, an AI “created by and for public officials”
If the current experiment concerns three generative AIs on the market – ChatGPT from OpenAI, LLaMA, the AI model developed by Meta and Bloom, an open source project led by the Franco-American startup Hugging Face – Dinum is developing, in parallel, a generative AI tool, “sovereign, free and open, created by and for public agents”.
Called Albert, this tool has the advantage of offering “ transparency of data sources and ease of access for all administrations “. In the coming months, Albert will be deployed in the France services network, alongside voluntary advisors.
With the same concern for sovereignty, another experiment concerns the creation of a French-speaking data hub to train an AI that better reflects our French culture than tools of Anglo-Saxon origin. “ On average, less than 0.2% of training data for artificial intelligence models is French,” recalls the ministry. A first beta version of this project, called Villers-Cotterêt, is planned in six months.
Furthermore, the use of AI will be extended to other ministries, particularly as part of the work to simplify relations between the administration and businesses announced by Bercy. Launched by Elisabeth Borne in September, the committee of experts for generative artificial intelligence should decide in early 2024 on “ the opportunity and modalities of this extension”.
A public-private “Alliance”
Dinum is intended to steer the government’s strategy on the subject. To do this, this summer it launched an incubator called Alliance. It brings together administrations – Cned, Cnil, INA, Arcom, Pôle emploi, etc. – research organizations and major schools – Inria, CentraleSupélec, CNRS, etc. – and companies, whether they are French flagships – Illuin Technology, OVHcloud, LightOn, Mistral AI – or GAFAM (Meta, Google).
Alliance objective: experiment and scale up sovereign and free products. To date, seven initiatives have been supported. Among them, we find CamemBERT 2.0. Led by the ALMAnaCH (Automatic Language Modeling and Analysis & Computational Humanities) team at Inria, this project consists of creating an updated corpus of texts in French for this French version of the famous BERT model (Bidirectional Encoder Representations from Transformers) .
CamemBERT is the most used natural language processing model for French companies with 22 million downloads since its publication at the end of 2019. Enedis uses it to categorize its thousands of customer requests per day, a task previously carried out manually. The return on investment for the electricity network manager is around 3 million euros per year.