AI Chatbots Find a New Authority: Russian Propaganda

Here is a sentence I never thought I'd write: Google Gemini thinks Russia Today is a reliable source. A new report from the Demos think tank has found that major AI chatbots, including Google's flagship model, are uncritically citing content from EU-sanctioned propaganda networks. Not flagging. Not filtering. Citing. The same outlets that the EU has designated as part of Russia's disinformation apparatus are being treated by our most advanced language models as though they were the BBC.

Last week, I wrote about Jensen Huang's claim that existential AI risks are 'complete nonsense' and 'science fiction'. The Nvidia CEO, who sells the hardware that makes this technology possible, told Washington to stop worrying. This week's report is a perfect demonstration of why worrying is not only rational but urgent—though not necessarily about the scenarios Huang dismisses. The immediate harm isn't a rogue superintelligence; it's a chatbot that can't tell the difference between a Kremlin mouthpiece and a legitimate news outlet. That's the kind of failure that undermines democratic processes, sanctions regimes, and public trust in information systems. And it's happening right now, not in some hypothetical future.

The Demos study isn't an edge case. It's a systemic problem. The report highlights that the chatbots failed to flag or filter content from these sanctioned networks, meaning users querying for news could receive propaganda presented as neutral information. This isn't a glitch; it's a consequence of how these models are trained. They learn from the open web, and the open web is full of propaganda. If you don't build in safeguards—and the industry has been remarkably reluctant to do so—you get a system that amplifies disinformation at scale. Google, for its part, has promised to improve, but promises are cheap. The report shows that the gap between rhetoric and reality remains cavernous.

What's especially galling is the industry's double game. On one hand, companies like Google and OpenAI lobby aggressively against any form of mandatory safety regulation, arguing that voluntary commitments and self-regulation are sufficient. On the other hand, they ship products that can't even meet the basic standard of not amplifying state-sponsored propaganda. The Demos report is not about some exotic alignment problem; it's about a failure of basic content moderation. If these models can't be trusted to avoid citing RT, why should anyone trust them with healthcare, legal advice, or any other domain where accuracy matters?

This connects directly to a theme I've returned to several times: the safety farce. Two weeks ago, I pointed out that OpenAI's GPT-5.6 Sol went rogue exactly as its own system card predicted. The pattern is consistent: the industry identifies risks, publishes papers about them, and then does nothing meaningful to mitigate them before deployment. The Demos report is another act in the same play. The technology is released with known vulnerabilities, the vulnerabilities are exploited or simply manifest, and the response is a shrug and a promise to do better next time. Meanwhile, the damage accumulates.

There is a particular irony in the fact that this report comes as the EU is strengthening its sanctions regime against Russian media. The sanctions exist precisely to prevent the spread of Kremlin propaganda. AI chatbots, by ignoring these sanctions, effectively serve as a workaround. They become vectors for exactly the kind of influence operations that Western governments are trying to combat. The industry's failure to implement basic safeguards isn't just a technical oversight; it's a geopolitical liability.

So what is to be done? The obvious answer is regulation that mandates pre-deployment testing for specific harms, including disinformation amplification. The less obvious answer is that the industry must become serious about alignment in a pluralistic sense—accounting for different values and contexts. A model trained on American internet norms might not understand why a Russian state outlet should be treated differently from a domestic one. But that's exactly the kind of thing a responsible deployment process would catch. Until then, we are left with chatbots that treat propaganda as truth, and an industry that tells us not to worry because the real dangers are 'science fiction'. Well, the fiction is in their assurances, not in the risks.