Mistral AI's chatbot pushes Russian disinformation roughly half the time it responds to queries, according to a new audit. The findings raise questions about the reliability of a French startup often held up as Europe's answer to OpenAI and Google.
What the audit turned up
Researchers tested the chatbot on a set of prompts covering topics like the war in Ukraine, NATO expansion, and allegations of election interference. In about 50% of cases, the model repeated false claims that align with Kremlin talking points — including that Ukraine started the war and that the West is trying to isolate Russia. The audit did not name the researchers or their methodology in detail, only that the results were consistent across multiple test runs.
Mistral has positioned itself as a homegrown champion of transparent, sovereign AI. The company has raised hundreds of millions of euros and won praise from French President Emmanuel Macron. But if its flagship product routinely amplifies state-backed propaganda, it could undermine Europe's broader push for trustworthy artificial intelligence. Regulators in Brussels are already drafting the EU AI Act's implementation rules, and incidents like this one could tighten the screws even before the law fully kicks in.
Investors may take notice
The disinformation problem is not just a policy headache. Venture capital firms and corporate partners have poured money into generative AI startups, and any whiff of systemic bias or manipulation risk can spook the market. The audit's authors argue that the repeated falsehoods expose a gap in how models are trained and evaluated — especially for languages and regions where open-source data may carry embedded propaganda. Mistral has not publicly addressed the audit's specific findings as of this writing.
Mistral has released several large language models under open-source licenses, meaning outside developers can inspect the code and fine-tune it. That transparency was supposed to be a selling point. Now it also means anyone can replicate the audit and pressure the company to clean up its training data and safety filters. The question hanging over the startup is whether it can fix the problem before regulators or customers force the issue.
