AI chatbots repeat pro-Russian disinformation as facts, audit finds

Major commercial AI chatbots, including OpenAI’s ChatGPT and Mistral, are frequently repeating pro-Russian disinformation as verified fact, according to a new investigation by the disinformation watchdog NewsGuard. The audit highlights a growing vulnerability in generative AI systems, where probabilistic algorithms prioritise high-volume content over reliability, allowing narratives from the sprawling Pravda network to infiltrate responses.

In March 2025, NewsGuard found that 33 per cent of tested chatbots repeated false narratives from the Pravda network. By January 2026, this figure had risen to 50 per cent, with models citing Pravda-affiliated websites as sources. Chine Labbé, senior vice president at NewsGuard, noted that the Pravda network comprises 370 sites that published approximately six million articles in 2025. This staggering volume ensures that statistically, content aligned with Kremlin interests is more likely to be selected by the models.

The disparity in accuracy is starkly evident across different languages. While English and French queries often trigger correct flagging of false claims, responses in less widely spoken languages such as Finnish, Danish, and Slovenian frequently validate disinformation. A test conducted by FRANCE 24 and replicated from Nordic fact-checking network Nordis showed Microsoft’s Copilot correctly identified a false claim about a Danish student killed in Ukraine as fake in English and French, but incorrectly affirmed the rumour when asked in Finnish or Slovenian.

Experts suggest this vulnerability may stem from intentional data poisoning or "data voids" where reliable information is scarce. Pipsa Havula of Nordis pointed to low-quality, machine-generated text in languages like Finnish, which appears designed to target bots rather than human readers. She noted that some Kremlin informants have theorised such strategies, although she acknowledged that gaps in reliable information could also explain the prevalence of these narratives in specific linguistic regions.

The issue extends beyond chatbots to search features and image verification tools. Google’s AI Overviews and Lens have also been compromised, with tests showing that reverse image searches often return false information for debunked AI-generated images. Marc Faddoul of AI Forensics warned that trust and safety standards vary significantly between tech giants, with some implementing fewer safeguards than others.

With 20 per cent of French people already using AI for news, according to a survey by France’s media watchdog Arcom, the stakes are rising. Researchers are calling for stricter measures, including blacklisting known propaganda sites and adopting whitelisting approaches for sensitive topics such as elections and public health. As Generative Engine Optimisation (GEO) emerges, both malicious and well-intentioned actors are working to ensure their narratives are adopted by AI tools, threatening to crush facts under alternative realities.