The Russian-funded Pravda network infects AI model training data so that chatbots generate content favorable to Kremlin theses. Discovered in early 2024 by Viginum, the French government agency attached to Matignon, this network has 150 websites that broadcast content in several languages (including French), in 49 countries. Among them are thematic sites such as Trump.News-Pravda or Macron.News-Pravda.
A network of websites to influence AI
Unlike traditional media, these websites do not produce content per se, they relay Russian propaganda by aggregating and massively disseminating stories glorifying Russia and Vladimir Putin's policies. The origin of this disinformation comes from official Russian sources, pro-Kremlin influencers and state organizations.
The network, also known as "Portal Kombat", has no direct link with the media Pravda, despite an identical name (which means "truth" in Russian). It is extremely prolific: last year, Pravda websites produced a whopping 3.6 million articles, a huge volume read by... very few people. Many of these sites have around a thousand unique visitors per month. To give you an idea, the official media site RT has over 14 million unique monthly visitors.
But this network is not looking for popularity among Internet users. By using SEO techniques, these sites maximize the visibility of their fake articles to better influence generative AIs. AIs that then generate responses based on Pravda's disinformation via their information retrieval systems (crawler robots and search engines). This “LLM grooming” technique (manipulation of language models) is clearly working wonders, according to a NewsGuard investigation.
This American company created in 2018 by veterans of journalism and publishing presents itself as a tool to combat false information and the manipulation of information on the internet. It tested 10 of the leading AI chatbots on the market: ChatGPT-4o (OpenAI), You.com Smart Assistant, Grok (xAI), Pi (Inflection AI), Le Chat (Mistral), Copilot (Microsoft), Meta AI, Claude (Anthropic), Gemini (Google) and the Perplexity answer engine.
The investigation was based on a sample of 15 false information stories disseminated by the Pravda network between April 2022 and February 2025. These claims were previously verified and catalogued in NewsGuard's Misinformation Fingerprints database. The aim was to measure the frequency with which these tools take up this disinformation.
Among the false information used to test the bots was, for example, the lie that Volodymyr Zelensky ordered the blocking of Truth Social in Ukraine. This is completely false: the Ukrainian president never asked for such a thing, and Donald Trump’s social network application was never distributed in Ukraine…
For each of the statements, NewsGuard formulated three variants of queries: neutral tone, “biased” tone (the question implies that the statement is true), “malicious” tone (an explicit formulation inciting disinformation). In total, 45 different queries were entered into the 10 bots.
The success of the Pravda network
The analysis of the responses confirms that the influence of disinformation on the AI’s responses is proven. A third of the bots’ responses actually contained disinformation; almost half (48%) contained a “debunk”, in other words a fact check. 18% were non-responses, indicating a lack of data or the inability to address the topic.
7 of the 10 bots cited sites from the Pravda network as sources. 56 out of 450 responses contained direct links to these famous articles. Finally, 92 articles were referenced in total, with some AI models even citing 27 different articles. As we can see, Russia has won points in this information war.
And it is not enough to turn off the Pravda tap to no longer influence AI models. First, because the network is not the source of disinformation, it is content, so to speak, to relay and amplify it. Above all, Pravda represents a systemic threat, because it cannot be easily neutralized with filtering measures. Its strategy is based on replication and diversification, making it very difficult to block sites with moderation algorithms. A blacklist of sites can be established, all that is needed is to create new websites to bypass the restrictions.
Source: NewsGuard
0 Comments