Does the P in ChatGPT stand for propaganda? The Kremlin certainly thinks it should. In 2025, at a press conference in Moscow, John Mark Dougan—a fugitive deputy sheriff from Florida who was granted political asylum in Russia and became an active part of the state’s media machine—declared: “By pushing these Russian narratives from the Russian perspective, we can actually change worldwide AI. It’s not a tool to be scared of, it’s a tool to be leveraged.”

Russia really is trying to influence the responses of global neural networks—primarily through the information they receive from the internet. Politicians, researchers, and journalists have known this for several years now. However, wanting to rewrite ChatGPT’s answer to the “who does Crimea belong to?” question is one thing. Actually rewriting it is another. Before we can assess the effectiveness of the Kremlin’s AI propaganda ecosystem, we must first understand its diffuse, contradictory nature.

Dougan likely appeared on camera because he could not hide anymore. Researchers, fact-checkers, and journalists have totally exposed a network linked to him, called CopyCop—a labyrinth of dozens of websites that imitate U.S. and European regional publications, with titles like DC Weekly, London Crier, or Berliner Tageblatt for that extra bit of believability. The articles are copied from real media outlets and rewritten using neural networks. The new versions include the appropriate political slants and fakery—about Western and Ukrainian politicians, for example.

Who is behind this, exactly? In 2024, the Washington Post, citing documents from a European intelligence agency, reported that Dougan regularly met with Valery Korovin—an associate of Alexander Dugin and the head of the Center for Geopolitical Expertise, which Dugin founded. According to the U.S. Department of the Treasury, the Center operated with GRU backing, maintained at least 100 fake news sites, and ran a server with generative models.

On another network, Pravda, researchers from the American Sunlight Project counted more than 20,000 posts across 97 websites over a 48-hour period. The researchers connected this volume and speed to possible attempts to influence large language models—a practice known as LLM grooming.

However, it remains unclear whether this grooming is effective. In two NewsGuard stress tests conducted in 2024 and 2025, ten chatbots were fed queries based on a preselected set of Russian disinformation. They repeated false claims in approximately one out of every three responses—in 32 percent and 33 percent of cases, respectively.

That still does not mean every third conversation any user has with a chatbot will include Russian propaganda. In the 2024 test, two-thirds of the queries were leading or directly asked the chatbot to spread disinformation. The final percentage also included responses that repeated a false claim with caveats. Such tests demonstrate the chatbots’ vulnerabilities under specific conditions, but they do not measure the frequency of these failures in everyday use.

Another study, published in the peer-reviewed journal HKS Misinformation Review, examined four chatbots and used a different set of queries. Of the 416 responses, 5 percent supported a false claim and 8 percent contained links to Pravda network websites, but only 1 percent used such links to support misinformation. In most other cases, the links were accompanied by disclaimers or refutations.

Differences in models, search queries, and evaluation criteria make it impossible to directly compare these percentages with NewsGuard’s findings. However, the study did reveal an important pattern: Links to propaganda sites appeared predominantly in responses to niche questions for which there are few reliable sources. The authors found no convincing evidence that the models had been successfully retrained by Russian propaganda, instead attributing this pattern to “data voids.”

Russian propaganda’s main target is sources that the chatbot finds directly while preparing its response. Many services use internet searches to obtain up-to-date information. Spreading misinformation does not require retraining the model: All it takes is for the model to find unreliable material and deem it a suitable source.

A lot depends on the topic, the wording of the query, and the language. For example, the large body of English-language publications from independent media outlets about Alexey Navalny allows chatbots to cross-reference different sources. But let’s imagine a search in Czech or Swedish for a hypothetical, little-known anti-war activist from Khabarovsk. If little reliable coverage about him exists, but propagandists have posted numerous articles containing false claims, these materials could make up a significant portion of the search results. A single claim repeated across different websites risks being mistaken for independent confirmation. 

However, that does not preclude grooming from taking other forms. You still have the “first and last mile” problem.

The most difficult—and most desirable—strategy for propagandists is to get a false claim to be included in a neural network’s training data. To do this, they flood the internet with texts containing repetitive assertions, in the hope that the model will learn the “proper” take.

This process is slow and difficult to control: The complete composition of training data for the largest commercial models is usually not disclosed, and the data itself undergoes selection and filtering. Therefore, even publishing millions of articles does not guarantee a specific outcome.

Furthermore, the mere inclusion of propaganda in the training data does not necessarily mean that the model will accept it as fact. Additional training by developers influences the model’s behavior by setting rules for how it should respond. In OpenAI’s published guidelines, for example, a response about Tiananmen Square should address the 1989 massacre, and any discussion of Stalin’s achievements must also mention repressions and the Gulag.

Application owners can select the sources available to a model or restrict its responses to certain topics. Governments may impose requirements on services operating within their territory. Chinese regulations, for example, require providers of generative AI serving the domestic public to adhere to “core socialist values.” And in Türkiye, a court has already blocked dozens of responses from the Grok neural network on Twitter due to “abuse. 

Russia’s approach barely works for it. The Russian authorities can impose requirements on Yandex and Sberbank and restrict access to Western apps, but publishing thousands of propaganda pieces does not give their authors control over ChatGPT settings globally.

On the other hand, there is a definite upside to Russia’s AI-driven propaganda operating in a fragmented, chaotic, and at times illogical manner. It creates a cacophony that often does not resemble a coordinated information campaign but instead seems to mimic a genuine public debate.

RT and Sputnik, domestic political entities, intelligence agencies, private contractors, and projects that have grown out of deceased Wagner Group chief Yevgeny Prigozhin’s empire all coexist within the Kremlin’s AI propaganda ecosystem. Those players may compete with one another, copy each other’s work, and spread mutually exclusive versions of events.

However, the contradictory nature of the messages isn’t necessarily a problem for the Kremlin—sometimes propaganda exists to undermine trust in reliable information. Different actors may offer incompatible explanations of an event while simultaneously reinforcing the overall impression that “we will never know the whole truth.”

If an LLM fails to recognize that the posts share a common source, it may mistake them for independent voices. If it is unable to assess the quality of the evidence, it risks presenting established facts and propaganda-driven fabrications as equally valid versions. In the aforementioned HKS Misinformation Review study, some responses that cited Pravda websites were indeed grouped into the broader category of “conflicting reports.”

A new phenomenon is emerging, one we could call “effective inefficiency”: The duplication of work and contradictory contributions are increasing the volume and variety of data that may wind up in neural networks’ responses.

But before anyone panics: Russia’s system has its limits, and it is safe to assume that those limits have either already been reached or are nearly reached. All the tools available to Russian propagandists have been scrutinized by AI companies, Western intelligence agencies, researchers, and journalists. Each generation of LLMs will be more resilient than the last, and preemptive measures against potential operations involving “data voids” are already in development. Therefore, we should not blindly extrapolate the Russian AI threat to 2030, let alone 2035. Sooner or later, Russia will hit a wall.

This text is part of the “Russia 2030: Rogue Superpower” project.