Le Chat fails on Iran war propaganda in 60 percent of leading prompts

WHERE LE CHAT FAILED

NewsGuard's audit, as reported by The Decoder, tested Mistral's Le Chat on ten false claims tied to Russian, Iranian, and Chinese narratives. The sharpest result comes from leading prompts about Iran: according to the report, the model repeated disinformation in 60 percent of those attempts.

The spread matters. On neutral prompts, the error rate was 10 percent. On leading prompts, it rose to 60 percent. On malicious prompts, it reached 80 percent. That does not look like a random slip; it looks like a system that becomes much weaker once the user supplies direction, confidence, or manipulative framing.

Mistral has positioned itself as a European alternative to U.S. AI labs, emphasizing openness, speed, and data sovereignty. That argument can still matter for buyers that do not want all AI infrastructure tied to the United States. But NewsGuard's finding shows the edge of that promise: a model's geopolitical origin says too little about whether it resists war propaganda.

The most sensitive part of the story is France's defense use. The source says France's Ministry of Defense uses a customized offline version of Le Chat. The audit, as described in the available material, does not test that exact isolated instance, so the detail should not be inflated into a claim about a specific military system. The risk is still clear: if the public model accepts adversarial framing this easily, defense users need to know what their version actually fixes.

SOVEREIGNTY IS NOT SAFETY

It is not enough to say that every large language model sometimes hallucinates. This is not about a creative error. It is about a model's willingness to repeat politically useful falsehoods when a prompt nudges it in that direction. In war narratives, that failure is not cosmetic; it can amplify someone else's information operation at almost no cost to the attacker.

The methodology still deserves caution. The available report summarizes the results but does not provide the full view of every query, sample choice, and attack design. That means the numbers should be read as a serious signal for verification, not as a complete forensic map of every Le Chat deployment.

Mistral's best response would be boring and concrete: publish a clearer account of the testing, explain where safeguards failed, and give a timeline for fixes. Defense and government users need more than the claim that a model is European or offline. They need evidence that it cannot be easily pushed into delivering an adversary's narrative.

The lesson is broader than one French model. AI sovereignty only means something if it includes sovereignty over quality, evaluation, and safety standards. Without that, local infrastructure can become a local way to distribute someone else's disinformation.