Russian disinformation network flooded training data to manipulate Western AI chatbots, study finds
A Russian disinformation network called Pravda (“Truth”) has influenced leading AI chatbots’ output by publishing numerous articles that made their way into the bots’ training data, a new report from the analysis group NewsGuard reveals. According to researchers, this wasn’t just a side effect of Moscow flooding the web with false narratives — it was the initiative’s main goal. Here’s how the scheme worked.
A “well-funded” online Russian disinformation network called Pravda put out 3.6 million articles last year, many of which were processed by popular chatbots such as OpenAI’s ChatGPT-4o, Claude (Anthropic), Meta AI, Gemini (Google), and Copilot (Microsoft), according to a new report from the watchdog group NewsGuard. These chatbots reportedly reproduced narratives spread by Pravda in 33 percent of their responses.
The researchers’ findings align with a February report from the American nonprofit organization American Sunlight Project (ASP), which found that manipulating AI technologies, rather than reaching human users, was likely the Russian network’s primary goal. This method of influencing large language models (LLMs) has been termed “LLM grooming.”
According to NewsGuard, Pravda “is pursuing an ambitious strategy by deliberately infiltrating the retrieved data of artificial intelligence chatbots, publishing false claims and propaganda for the purpose of affecting the responses of AI models on topics in the news rather than by targeting human readers.”
The bitter truth is that events in Russia affect your life, too. Help Meduza continue to bring news from Russia to readers around the world by setting up a monthly donation.
Pravda does not produce original content but amplifies material from pro-Kremlin sources, including Russian state media, through a network of various websites. The NewsGuard study identified 150 sites in this network, around 40 of which publish content targeting Ukraine, 70 focus on Europe, and 30 target audiences in Africa, the Pacific region, the Middle East, North America, the Caucasus, and Asia. The rest are divided by topic.
The sites in the Pravda network publish content in a range of languages. Many of their domain names include names of cities and regions in Ukraine, such as News-Kiev.ru, Kherson-News.ru, and Donetsk-News.ru.
Over the three years of Russia’s invasion, Pravda has spread at least 207 disinformation narratives, including claims about “secret U.S. biolabs in Ukraine” and accusations of “Zelensky’s misuse of U.S. military aid,” according to NewsGuard.
ASP experts have warned of serious long-term risks associated with attempts to manipulate artificial intelligence. The more false narratives circulate in online media, the greater the likelihood that language models will begin to treat them as credible and incorporate them into their responses.
Axios noted that NewsGuard’s study comes just as Washington has reportedly halted the U.S. Cyber Command’s activities against Russia.
