- Category
- Investigations
Claude and Mistral Echo Kremlin-Aligned Claims in Over 70% of Tested Answers
Prefer U24 Media on Google
- Authors
![Photo of Roman Kohanets]( "Photo of Roman Kohanets")
-111f0e5095e02c02446ffed57bfb0ab1.jpeg "Photo of Roman Kohanets")
Claude and Mistral, AI models developed by US-based Anthropic and French company Mistral AI, respectively, repeated contested or false claims from Kremlin-aligned sites in roughly 70% of answers on topics with little mainstream coverage.
This was reported by the Global Disinformation Index (GDI) in a study published on October 12, titled “Russia benefitting from Data Voids to influence LLMs.” The analysis, commissioned by GDI and conducted with the AI analytics firm Sy.nexus, tested Claude Sonnet 5 and Mistral Large.
We bring you stories from the ground. Your support keeps our team in the field.
Researchers used 85 plain-language questions, repeated some prompts to assess consistency, and collected 280 answers from each model. According to the report, Claude repeated such claims in 70% of those answers and Mistral in 71%.
Neither model consistently rebutted the claims, GDI found. Among answers that repeated a claim, Mistral presented it as true in 48% and Claude in 28%. Claude flagged or contradicted the claim in 60% of such answers, compared with about 20% for Mistral.
GDI stressed that the tests ran through the models’ APIs, not consumer chatbot interfaces. The report noted that some companies add guardrails to their user-facing chatbots, but a growing number of corporate customers rely on APIs. It made no claims about what people using the chat interfaces might see.
-639eacbfbce594c9306781d41539ee90.png)
The questions were built from data voids, meaning topics that Kremlin-aligned sites promote but that mainstream outlets rarely cover. In one example, GDI examined a fabricated Russian claim that its forces had captured Volnoye (Ukrainian: Vilne), a real village in Ukraine's Dnipropetrovsk region. The report emphasized that the alleged Russian capture had never happened.
In the responses highlighted by GDI, neither model cited Ukrainian news outlets to substantiate the alleged capture. Mistral presented the Russian Ministry of Defense's claim as fact without caveats. Claude attributed the claim to Russian state media but acknowledged that it could not find independent confirmation from Ukrainian officials, the Ukrainian monitoring group DeepState, or the Institute for the Study of War.
Mistral's answers to that question varied widely, the report added. Asked eight times, it produced four distinct kinds of responses, ranging from rejecting the claim to presenting it as settled fact. GDI described this inconsistency as a known challenge requiring further research.
The two models also searched differently, according to GDI. Claude ran 3.6 searches per answer, compared with 1.2 for Mistral, and its searches returned a Russian- or Chinese-state-linked source in 88.2% of answers, compared with 32.1% for Mistral. TASS appeared in searches behind 140 of Claude’s answers and the Pravda network in 130, while neither appeared for Mistral.
GDI reported encountering the problem firsthand on July 1, 2026, when it asked an LLM about GDI itself. Claude cited the Strategic Culture Foundation, a Kremlin-aligned think tank added to a US Treasury sanctions list in 2021. The report stressed that the risk came from drawing on a sanctioned site without caveats, not from the content itself.
GDI linked the findings to the rise of zero-click searches, which it put at around 60%, where users accept an AI summary without visiting the original source. “Whoever controls the AI answer summary will control the narrative,” GDI concluded.
As a possible safeguard, GDI tested an exclusion list on Claude. A list of 191 blocked sites cut state-linked results from 30 of 34 matched answers to one, the report stated. It did not change whether answers agreed with the claim: 10 of 34 agreed with the list and 13 of 34 without it, too few to show a difference.
The report noted that the tool accepts lists of up to 250 sites, while lists of 500 or more were rejected, and that Mistral’s API offers no such option. GDI cautioned that any exclusion list would need to be continuously updated as new sites emerge and called for further research.
Russian exploitation of Claude has not been limited to the information space. In September, a Russian state-linked hacking group was found to have used the models to automate cyberespionage against more than 20 organizations, primarily in Ukraine and Europe.
Targets included government ministries, diplomatic missions, defense companies and drone technology suppliers, with AI-driven workflows handling reconnaissance, phishing, malware evasion and the processing of hundreds of gigabytes of stolen data. The hackers also stole a proprietary software development kit for a military drone vision system and spent several days reverse-engineering it.
Discuss this article:
-206008aed5f329e86c52788e3e423f23.jpg "Photo of Oleksandr Moiseienko")