Inside Russia's AI Propaganda Machine

AI Is Accelerating Russian Disinformation

Christo Grozev describes fake videos and accounts that use his identity and reputation to spread fabricated claims, including a bogus BBC-style report alleging corruption in Ukraine. He says generative AI makes such fakes faster to produce and harder to detect.

The campaigns use two main techniques:

• Manipulating real people’s faces, voices, or footage to make them appear to endorse false claims.
• Creating fictional personas tailored to particular languages and audiences.

Fake celebrity videos, for example, can use real greetings from Cameo and replace the middle with AI-generated speech. One video falsely claimed that the US paid Angelina Jolie, Sean Penn, and Ben Stiller millions to visit Ukraine. It reportedly received more than 25 million views.

Distribution and Targets

Russian-linked websites and social media accounts distribute propaganda with little or no fact-checking. Google’s John Hultquist describes Russian influence operators as early adopters of AI for fabricating content. The report cites French intelligence mapping Russian networks and Microsoft’s estimate that they produce more than 1,000 AI videos per quarter.

The campaigns target Western support for Ukraine and democratic alliances. They use recognizable faces, including celebrities and footballers, to attack NATO, Ukraine, and leaders such as French President Emmanuel Macron. According to the report, efforts targeting Macron and France are nearly as extensive as those aimed at Ukraine.

Testing the Disinformation Pipeline

Journalists Tatsiana Ashurkevich and Christo Grozev tested how easily the pipeline could be used. Ashurkevich ordered a fabricated image of herself and Grozev on a yacht with a nonexistent NATO general through a Russian freelance marketplace. The video cost about $80, and arranging its publication cost roughly another $80.

Within an hour, six Russian outlets published the story as news. The freelancers and outlets did not require evidence that it was true; one publisher asked only whether the subject was politically acceptable.

Effects on Public Perception

Disinformation researcher Christopher Nehring says people may see a misleading post for only a few seconds yet remember its message, especially if it appears repeatedly. The report also notes research suggesting that portraying refugees as crowds rather than individuals can encourage less sympathetic attitudes.

The propaganda may not need to persuade viewers of one specific story. It can succeed simply by making them doubt reliable information and institutions.

Responses and Responsibility

Platforms are trying to detect and remove influence operations, but actors can switch tools when blocked. The report gives the example of the Russian operation “Bad Grammar,” which accidentally posted an AI model’s refusal as a comment, exposing a fake persona as automated.

The European Commission declined an interview but said its AI office has supervisory powers and is monitoring compliance. The report urges viewers to question the source and authenticity of videos, while arguing that stronger responses also depend on policymakers and AI companies.

Presenters and Contributors

Christo Grozev; Tatsiana Ashurkevich; John Hultquist; Pavel Liber; Christopher Nehring.