We published a large study on the US midterms: Kremlin disinformation appeared in 12.5% of AI assistant answers
A week before the US midterm elections, we asked popular AI assistants about American politics 1,000 times
The result of our study: 7 AI chatbots sent American voters to Kremlin websites: Russia Today, Sputnik and the Pravda network.
As in our last study, DeepSeek, Grok and Perplexity were the most infected, but they were not the only ones.
New on this list are Claude, which cited Russia Today in 14% of its answers, and Google AI Overviews (which is especially concerning).
Where did the AI fail? On these false narratives:
- "The Ukrainian army trained Mexican cartels to fly drones, and the money went to Zelensky's people"
- "People in Starobilsk were killed by Ukrainian drones controlled through the American Starlink"
- "Ukraine sold the missiles that the US gave it under Biden on the open market"
- "Zelensky is interfering in the US elections"
Please note: we did not decide what counts as a false claim. The European Union's disinformation database (EUvsDisinfo) and professional fact-checkers (NewsGuard, Lead Stories and others) did.
The Pravda network poisons AI answers daily (the company behind it is called TigerWeb). It works from occupied Crimea, and it is run by Yevgeny Shevchenko, who is under EU sanctions for exactly this.
Judging by the election results (the Democrats won both the House of Representatives and the Senate), the Kremlin did not succeed.