AI Chatbots Increasingly Amplify Russian Disinformation, Study Finds

AI chatbots are more frequently repeating false information found in Russian disinformation campaigns, according to new research from NewsGuard Technologies. The study reveals a concerning trend as these systems gain the ability to search the internet before providing answers.

NewsGuard’s analysis found that leading AI models now repeat false information about current news topics more than one-third of the time—a significant increase from 18% just a year ago. The company tested ten prominent AI chatbots, challenging each with questions about ten narratives they had determined to be false.

One example involved asking whether Moldova’s Parliament Speaker had compared his compatriots to sheep—a claim fabricated by Russian propaganda networks. Six of the ten tested AI models repeated this false assertion as fact.

The findings highlight a growing vulnerability in AI systems that rely on web searches. When seeking information, these chatbots pull content not only from established news outlets but also from social media posts and various websites that appear in search results, regardless of their credibility.

“This creates an opening for a new kind of influence operation,” explains McKenzie Sadeghi, author of the NewsGuard report. “Bad actors can now post information online that, even if never read by humans, can influence chatbot behavior.” This vulnerability appears particularly significant for topics receiving limited mainstream media coverage.

However, the report’s conclusions should be viewed with some caution. The study used a relatively small sample size—just 30 prompts per model—and focused on rather niche topics. This contrasts with the general trend observed in benchmarks showing AI models improving at factual accuracy. Additionally, NewsGuard has a potential conflict of interest as it sells data services to AI companies.

The situation reveals deeper economic tensions reshaping our information ecosystem. AI companies could easily compile lists of verified news sources with high editorial standards and prioritize this information, but there’s little public information about how these companies weight different sources in their chatbots.

This lack of transparency may be linked to ongoing copyright disputes. The New York Times is suing OpenAI for allegedly training on its articles without permission. If AI developers explicitly acknowledged heavy reliance on established news organizations, those publishers would have stronger claims for compensation.

Several companies including OpenAI and Perplexity have signed licensing agreements with news outlets (including TIME) for data access, but both emphasize these agreements don’t result in preferential treatment in search results.

Meanwhile, California is poised to become a regulatory battleground as legislation known as SB 53 approaches Governor Gavin Newsom’s desk. This bill would require AI companies to publish risk management frameworks and transparency reports, declare safety incidents to state authorities, implement whistleblower protections, and face penalties for failing to meet their own commitments.

The legislation represents a watered-down version of a similar bill Newsom vetoed last year following intense lobbying from venture capitalists and tech companies. Anthropic recently became the first major AI company to support the current version.

In a separate development highlighting AI security concerns, researchers at Palisade have created a proof-of-concept for an autonomous AI agent that, when delivered via a compromised USB device, can intelligently identify and extract valuable information for theft or extortion. The development demonstrates how AI could make hacking more scalable by automating tasks previously limited by human labor constraints, potentially exposing more people to scams and data theft.