Recently, around a hundred agents from the General Directorate of the Treasury tested Alibaba’s Qwen AI model. However, the AI was ultimately disconnected due to a glaring lack of neutrality on certain geopolitically sensitive topics important to China.
Bercy and the Qwen AI: It’s Over!
As a reminder, Qwen (Tongyi Qianwen) is a family of artificial intelligence models developed by Alibaba Cloud, the computing arm of the Chinese tech giant. Since its launch in early 2025, the AI has become one of the world leaders with exceptional performance and a strategy focused on open source, enabling it to function entirely offline. Yet, while this trajectory appears rapid, the Qwen AI is currently quite controversial.
As Le Figaro revealed in a June 26, 2026 article, the Ministry of Economy and Finance (Bercy) ended a trial period for the Qwen AI. For several weeks, about a hundred agents from the General Directory of the Treasury (DGT) had been testing the model within the HéphAIstos project. It involved an internal generative AI experiment to validate the agents’ assistance in handling complex macroeconomic data and in synthesizing financial reports and strategic analysis notes. It also touched on the handling of sometimes confidential or sensitive economic information.
The fact is that Bercy detected a glaring lack of neutrality in Qwen regarding topics dear to China. This issue is quite recurrent at the level of Chinese AIs. More precisely, the problem stemmed from the post-training political orientation of the algorithm itself. When entering queries related to the international economy and Franco-Chinese relations, the agents observed biased, oriented, or even censored responses. In short, the AI was reproducing exactly the official ideological line of the Chinese government.
A Controversy at Several Levels
From a data security standpoint, the infrastructure was indeed sealed, operating in a closed loop with no web access. However, the AI’s responses, incompatible with the neutrality of the French administration, pushed Bercy to halt the collaboration. Now, the French startup Mistral AI has taken up the mantle and is deploying “L’Assistant,” a sovereign model that, in the long run, should equip nearly a million public agents.
Since its launch, the Qwen AI has already sparked controversy several times. At the end of 2025, the Australian Strategic Policy Institute (ASPI) demonstrated in a report that the model applied blocking and selective censorship filters depending on languages. Moreover, although Qwen has the knowledge required, there is a restrictive layer of algorithmic safety. Yet, that layer represented precisely the major problem during the DGT tests in France.
Finally, it should be noted that recently, the American startup Anthropic accused Alibaba of having carried out the largest distillation hack to date. According to the company, operators linked to the Qwen lab used nearly 25,000 fake accounts to conduct 28.8 million exchanges with the Claude model over a six-week period. The objective of this maneuver was to illegally and cheaply extract Claude’s reasoning and coding capabilities in order to train and improve the Qwen models.
