The Manufactured Crisis: Inside the Flawed Think Tank Report on AI Disinformation
A provocative new think tank paper has ignited a firestorm of controversy by claiming that Russian state actors are strategically rigging websites to manipulate Large Language Models (LLMs) into generating disinformation. The report posits that by seeding the internet with misleading content, bad actors are effectively training AI to parrot state-sponsored propaganda. While the authors present this as a groundbreaking exposure of a digital threat, the study has been met with immediate skepticism by critics who argue that the findings are less a revelation of geopolitical sabotage and more a result of poor methodological design and confirmation bias.
At the core of the report is a test wherein AI chatbots were fed leading questions about manufactured stories. The models complied with the false premises roughly 16.6% of the time, occasionally citing junk websites as sources. Critics, however, point out that this “one-in-six” hit rate is suspiciously consistent with previous studies, such as the May 2025 NewsGuard analysis on Australian election falsehoods. This statistical uniformity suggests that the problem lies not in a sophisticated Russian “rigging” scheme, but with the test design itself. By forcing chatbots to interact with leading prompts, the researchers may simply be triggering expected behaviors inherent to how models process limited data, rather than uncovering a coordinated influence campaign.
The report’s technical evidence of “deliberate targeting” is equally fragile. The authors highlight the use of specific website code—the max-snippet:-1 directive—as a smoking gun for Russian interference. This argument collapses under scrutiny, as the very same code is found in the source of the think tank’s own website and the news articles hosting the report. By flagging standard web infrastructure as a conspiratorial tool, the authors have inadvertently undermined their own premise. The study fails to distinguish between malicious intent and the boring, well-documented reality that AI models are trained on the entirety of the open web, where low-quality content and “data voids” inevitably exist.
Furthermore, the research ignores established scholarship that provides a more mundane explanation for these AI errors. A significant study by researchers from the University of Manchester and the University of Bern, published in the Harvard Kennedy School Misinformation Review, previously examined similar claims and arrived at a much lower “echo rate” of 5%. That team concluded that chatbots only lean on junk sites when high-quality information on a topic is thin—a phenomenon known as a data void. By failing to account for this and refusing to release their full prompt lists, the authors of the new report appear to have committed the very sin they decry: prioritizing a predetermined narrative over empirical rigor.
The implications of this study extend beyond bad science, raising serious questions about the professional incentives driving the current discourse on AI regulation. Reports like these often serve as effective marketing for the disinformation spammers themselves, who rely on the Western establishment to validate their influence. By loudly proclaiming that “the junk works,” the think tank is essentially gifting state-sponsored propagandists a massive PR victory, reinforcing the idea that these fringes are far more powerful and capable of swaying global technology than they actually are.
Ultimately, skeptics argue that this paper serves a specific policy agenda rather than the pursuit of truth. The report was funded by Will Perrin, a central architect of the UK’s Online Safety Act and a proponent of expanding platform regulation to include LLMs. By championing a study with unproven claims—and utilizing outdated, exclusionary terminology like “black-lists”—the initiative appears designed to facilitate state-led censorship. As the debate over AI safety intensifies, the circulation of such flawed research serves as a stark reminder that when it comes to the “disinformation crisis,” the loudest voices are often those trying to sell a specific regulatory solution.
