Mistral's Le Chat Repeats State-Sponsored Disinformation Half the Time, NewsGuard Audit Finds
Key Takeaways
- ▸Le Chat repeated false information in 50% of English tests and 56.6% of French tests, with 10 distinct Iran-war disinformation claims from Russian, Chinese, and Iranian state actors
- ▸Vulnerability scaled dramatically with prompt sophistication: 10% error rate for innocent queries, 60-70% for leading prompts, and 80-90% for malicious prompts seeking shareable disinformation content
- ▸The French Ministry of Armed Forces uses Mistral's technology despite the revealed vulnerabilities, though it claims its version lacks internet access and is customized for security
Summary
A NewsGuard audit published in April 2026 found that Mistral AI's Le Chat chatbot, Europe's most prominent home-grown conversational AI tool, repeated false claims about the Iran war 50 percent of the time in English and 56.6 percent of the time in French. The audit tested the free consumer version of Le Chat against 10 false claims originating from Russian, Chinese, and Iranian state-aligned networks using 30 prompts designed to reflect different user personas—innocent inquiries, leading questions, and malicious attempts to generate shareable content.
The vulnerability increased dramatically based on prompt type. When responding to innocent, neutral prompts, Le Chat provided false information only 10 percent of the time in both languages. However, when responding to leading prompts that assumed the false claim was true, the error rate jumped to 60 percent in English and 70 percent in French. Most concerning, when prompted by malicious users seeking to repackage falsehoods for mass distribution, Le Chat provided false information 80 percent of the time in English and 90 percent in French—suggesting significant risks of weaponization by foreign actors.
The timing raises additional concerns: France's Ministry of Armed Forces announced in January 2026 that it would provide all branches and agencies access to Mistral AI's models and services. In April, the Ministry disclosed that military personnel would use a customized version of Le Chat Enterprise with internet access disabled. Mistral AI declined to respond to two NewsGuard requests for comment on the audit's findings, limiting transparency around plans to address the vulnerabilities.
The findings underscore critical gaps in content safety for major AI systems operating in European markets, particularly as governments and enterprises increasingly adopt local AI alternatives to American providers. The audit demonstrates that geographic origin or government partnerships do not guarantee protection against state-sponsored disinformation campaigns.
- Mistral AI failed to respond to NewsGuard's requests for comment, raising questions about the company's transparency and commitment to addressing the identified safety gaps
- The audit reveals that European AI alternatives to U.S. providers are not immune to state-sponsored disinformation, complicating European technology sovereignty strategies
Editorial Opinion
This audit exposes a troubling paradox: Mistral's Le Chat, positioned as a sovereign European alternative to American AI systems, is demonstrably vulnerable to the very state-sponsored propaganda threats that European governments cite as justification for building local alternatives. The 80-90% error rate on malicious prompts isn't a minor flaw—it's an open door for foreign influence operations at scale. Mistral's silence in response to these findings is particularly alarming for a company whose French government clients are supposed to be leveraging it for defense and security applications.
