French Treasury halts use of Alibaba's Qwen AI over biased outputs
France's Treasury Department halted a trial of Alibaba's Qwen AI model in June 2026 after high-ranking officials reported biased outputs on topics related to China. The tool had been tested by about 100 agents at the General Treasury Directorate for several weeks and was replaced with a French alternative from Mistral AI.
What happened
From early June 2026, the Direction générale du Trésor tested Alibaba's Qwen AI model within an internal system called HéphAIstos, which provided a conversational agent and multi-language transcription tool to about 100 agents out of 1,300 for handling daily tasks including confidential and sensitive data.
High-ranking officials alerted leadership to biased outputs on China-related topics, with users reporting oriented or biased responses. On 23 June 2026, the Treasury Department halted the trial after only a few days and deployed a French model from Mistral AI to replace it.
Reported impact
An anonymous high-ranking official expressed concern that the choice of model was "very serious for the way we advise the French government on commercial policy toward China" because "there is necessarily a pro-Chinese bias with software developed under the careful watch of authorities."
The directorate, which advises the government on international commercial policy, fiscal matters, economic conditions, and employment forecasts, had been relying on a system with outputs potentially influenced by the Chinese regime.
Index analysis
Multiple Chinese chatbots have previously been reported for laconic or misleading responses on sensitive topics including Uyghur repression and Taiwan's status.
Governance controls
The Treasury Department decided to immediately interrupt the experimentation after identifying observed biases in certain responses relating to Chinese subjects.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.