At the beginning of June, around a hundred agents of the General Directorate of the Treasury (DGT) are offered a new in-house tool, called HéphAIstos. Under the hood of this conversational assistant, designed to help with writing, summarizing or transcribing, lies a powerful engine: Qwen AIa family of models developed by Alibaba.
Its aim was to assist civil servants, including in the processing of confidential data. But the experiment was stopped abruptly on June 23, barely a few weeks after its launch, following the discovery of obvious ideological biases in its responses.
This radical decision, taken at the end of June, follows internal alerts from senior officials concerning responses deemed “oriented” and pro-Chinese, calling into question the neutrality of the tool.
Why did Bercy suddenly disconnect the Chinese AI Qwen?
Bercy disconnected the Qwen AI model due to “ biases observed » in his responses on topics related to China. Senior officials quickly sounded the alarm, noting statements “ oriented » which were similar to a pro-Beijing political line.
This situation was judged “ very serious » by an agent, considering the advisory role of the DGT to the government on international trade policy.
The problem is not so much technical as doctrinal. While all AI models have biases from their training data, those of Chinese models are often suspected of being deliberately aligned with the censorship and propaganda of the Chinese Communist Party.
Independent tests have shown that Qwen describes Taiwan as “ inalienable element of China » and refuses to comment on the events in Tiananmen Square, proof of a very precise doctrinal compass.
Using such a tool at the heart of such a sensitive administration amounted to installing a potential vector of influence within the State itself.
What were the concrete risks of this experiment?
The risks were of two kinds: the data security et integrity of economic analysis. Although Bercy assures that the Qwen model, owned by the giant Alibabaoperated in isolation, “ without internet access or backdoor possibility “, zero risk does not exist.
The mere presence of software designed under the supervision of a foreign state posed a latent threat, especially when it manipulated confidential notes on tax policy or international economic relations.
Beyond the data leak, the most immediate danger was that of information pollution. By generating summaries or arguments tinged with a form of Chinese bias inherent, the tool could, subtly, influence the agents’ analyzes and, ultimately, the policy recommendations made to the government.
The real fortress is located in the training datawell before deployment, and in this area, the Chinese model was already compromised. The choice of this model, justified by its open source accessibilityappears in hindsight to be an incredible negligence.
What signal does this incident send about sovereign AI in France?
This incident is a resounding alarm signal about the fragility of French digital sovereignty. It comes at the worst time for the government, just a few days after the announcement with great fanfare of a plan to deploy theArtificial intelligence within the state.
This plan, supported by Minister David Amiel, aims to generalize a sovereign conversational agent, L’Assistantbased on French technologies and hosted on a secure infrastructure SecNumCloud (the state’s highest safety certification).
The Qwen debacle highlights a blatant contradiction between the discourse on sovereignty and certain practices, even if different solutions must be tested and the consequences drawn from them.
In response, Bercy immediately replaced the Chinese model with that of a French startup. The solution was to turn to the French nugget Mistral AIinstalled the next day.
This episode, although embarrassing, could serve as a shock to accelerate the adoption of national or European solutions reliable and auditable, even if the challenge of competing in performance with the American and Chinese giants remains.
