The growing use of generative AI is changing how Russia conducts information operations. AI tools allow operators to produce content at a much lower cost, manage larger networks of fake accounts, and maintain their activity over longer periods. They also create new opportunities to monitor public debates, identify sensitive topics, and adapt messages to different audiences.
These changes build on Russia’s existing influence infrastructure. Fake media outlets, social media accounts, advertising networks, intermediaries, and human amplifiers continue to play a central role in distributing content and reaching real audiences. AI strengthens this system by automating tasks that previously required large numbers of people. As a result, Russian influence operations are moving towards a more automated and decentralized model, with fewer resources needed to sustain large-scale activity.
How AI Reshapes Russia’s Influence Operations
Russia’s current influence infrastructure retains many methods developed by the earlier generation of troll factories. Today’s operations retain these methods but have evolved into a more decentralized system that separates content production, account management, and message distribution.
The Internet Research Agency (IRA), previously led by Yevgeny Prigozhin, relied on paid staff to create fake accounts, produce political content, and imitate public debate. This model allows different actors to contribute to specific stages of an influence campaign while maintaining coordination across the wider operation.
Operating as a distributed model, this ecosystem unites distinct actors and capabilities across different stages of the influence process. While Doppelganger draws on presidential administration-linked groups like the Social Design Agency, Struktura, and ANO Dialog to coordinate campaigns, other components focus on synthetic audience management, exemplified by the RT-developed Meliorator system that automates fake personas. By contrast, CopyCop (Storm-1516) combines fabricated outlets and producers, including John Mark Dougan and the Center for Geopolitical Expertise, with networks of external amplifiers to broaden narrative reach.
AI primarily changes the way these structures produce and distribute content. It speeds up writing, translation, identity creation, and automated posting. The CopyCop network also appears to rely on self-hosted Llama models, allowing operators to run generative AI independently of commercial service providers. This approach reduces the network’s reliance on commercial AI providers and their restrictions.
However, the technology does not remove the need for human involvement. Operators still set political objectives, coordinate distribution networks, select target audiences, and seek access to established public platforms.
The difference between content production and actual influence remains particularly important. The Bad Grammar campaign demonstrated that automated content production does not necessarily generate public engagement. Its posts and comments received little audience reaction. By contrast, some Storm-1516 stories reached much wider audiences through external intermediaries and amplifiers.
This distinction reveals the main organizational change. AI reduces the resources required to produce content, while the ability to distribute it beyond controlled networks remains a key condition for achieving political influence.
Identifying Vulnerabilities and Points of Tension
AI also expands the analytical capabilities of Russian influence operations. By processing large volumes of public posts, comments, and other open-source information, operators can identify sensitive political issues, track shifts in public debate, and locate social conflicts that may support their narratives.
AI has also become part of the planning process behind these operations. Operators use generative models to research open sources, summarize articles, and prepare campaign materials. The leaked files from the Social Design Agency reveal a structured approach to campaign planning, in which operators define political objectives and select the methods to achieve them. Together, these practices show how AI can support different stages of an influence campaign, from initial research and preparation to content production.
A key feature of these operations is their ability to exploit existing divisions within target societies rather than rely on uniform messaging. CopyCop and Storm-1516 demonstrate how Russian influence networks use this approach to adapt their narratives to different audiences. Rather than promoting a single message across multiple countries, the network focuses on issues that already generate public debate and political friction. Its campaigns have targeted disagreements over military assistance to Ukraine and trust in US authorities, tensions between France and Armenia, separatist sentiment in Canada’s Alberta province, and disputes surrounding elections in Moldova.
By building on these divisions, operators can make externally promoted narratives seem relevant to domestic debates and amplify existing tensions within target societies. It also demonstrates the importance of adapting influence content to local political contexts.
The further development of these capabilities could make the process faster and more systematic. AI can help classify large volumes of information, assess public reactions, track narrative circulation, and identify changes in audience sentiment. This creates opportunities to adjust campaign content and distribution in response to developments in the information space.
Building Synthetic Audiences to Scale Influence
One of the most significant changes associated with generative AI is the ability to create and manage entire networks of artificial users. These networks imitate public debate, express political opinions on behalf of supposed residents of target countries, and create the appearance of grassroots support for specific narratives.
Previously, maintaining large numbers of fake accounts required substantial human resources. Generative AI reduces these requirements by automating profile creation, content production, and routine account activity, allowing operators to maintain larger networks with fewer personnel.
The Meliorator system, developed by RT-linked groups with FSB involvement, demonstrates this approach. It automatically generated user profiles with individual biographies, locations, and political views. The system also managed posts, comments, and shares through a centralized interface.
Ahead of the 2024 US elections, hundreds of fake accounts posing as American voters shared video statements by Vladimir Putin. Some accounts also posted these materials in the comments under publications by a federal candidate. The operation sought to present official Russian messaging as the views of ordinary American users.
The experience of Operation Bad Grammar illustrates the limitations of relying on automated content production to generate influence. Russian operators used AI to produce comments and automate Telegram posts on the war in Ukraine, political developments in Moldova and the Baltic states, and US politics. The campaign generated large volumes of content, but most comments attracted little engagement. The fake accounts failed to gain traction in public discussions or generate a meaningful audience response. This highlights a key limitation of synthetic audiences. The ability to simulate public activity does not necessarily translate into genuine engagement or wider influence.
CopyCop/Storm-1516 illustrates how Russian influence operations can overcome some of the limitations of automated account activity by combining synthetic audiences with a broader network of fabricated media outlets. The network launched at least 300 websites in 2025, presenting them as local news outlets, political organizations, and fact-checking projects. This model creates multiple channels for distributing the same narratives while giving them the appearance of independent voice. By combining fabricated content with coordinated amplification, operators can generate repeated exposure and artificial signals of public interest, increasing the likelihood that selected narratives reach audiences beyond their own networks.
The Truefact network further expanded this model by developing regional and language-specific editions for different audiences, including Turkish, Ukrainian, and Swahili versions. AI tools helped populate these websites with large volumes of content, allowing operators to maintain a steady flow of publications across multiple language environments. This approach reduced the need for separate editorial teams while making it easier to adapt narratives to local political contexts.
For instance, the summer 2025 story claiming that Russia was preparing massive strikes against Ukraine illustrates this distribution model. The narrative moved through CopyCop websites, InfoDefense Slovenia, and the Portal Kombat network. Some stories produced by this ecosystem achieved substantial reach and entered wider public debate. Doppelganger follows a related approach. The operation uses a combination of established digital influence and content distribution techniques to expand the reach of manipulated narratives.
Together, these cases demonstrate the development of a more integrated influence infrastructure. Fake accounts provide artificial signs of public support, fake media outlets produce and legitimize content, and external amplifiers help move selected narratives into wider information channels.
AI makes it easier to maintain these interconnected structures, yet the scale of automated activity alone does not determine their effectiveness. Reaching independent users and entering genuine public discussions remain essential to achieving wider influence.
Compromising Search Engines and AI-Generated Answers
The growing use of generative AI in search engines and information services creates an additional area of potential vulnerability. Information operations may increasingly target the digital systems that people use to obtain and interpret information, rather than focusing exclusively on human audiences.
Large-scale publication of manipulated content can affect the information environment from which search algorithms and language models retrieve material. This creates a risk that coordinated disinformation may appear in AI-generated answers as an established fact supported by multiple independent sources.
The growing use of AI systems to retrieve and synthesize information creates an additional channel for influence operations. Large-scale publication of targeted content allows operators to shape the material that enters search indexes and the sources that AI systems draw on when generating answers (LLM grooming). The Pravda (Portal Kombat) network, which publishes at least 3.6 million items annually, demonstrates the scale at which this approach can operate. CopyCop/Storm-1516 follows a similar model through its extensive network of websites and high-volume content production. These practices extend influence operations beyond direct audience engagement on media and social platforms. They also create opportunities to embed selected narratives in the wider information environment on which AI systems rely.
The use of multiple interconnected websites can create an artificial impression of independent confirmation. This is particularly relevant to niche topics where reliable information remains limited. In such information gaps, often described as data voids, coordinated publications may gain disproportionate visibility.
Another technique, called LLM poisoning, covers a broader range of technical methods. These include deliberately manipulating training data to influence a model’s behavior after training. Attackers may also target connected knowledge bases and retrieval-augmented generation (RAG) systems. By introducing fabricated documents or hidden instructions into external information sources, they can potentially influence the answers generated by a system without modifying its underlying model.
These methods create different levels of risk, from the repetition of false claims to systematic bias and attempts to introduce hidden behavioral triggers.
The central concern is the potential transformation of mass-produced disinformation into source material for seemingly neutral AI-generated answers. This could extend the reach of influence operations beyond social media and traditional news platforms, allowing manipulated information to enter the everyday digital services people rely on.
Technical Infrastructure Foundations
The expansion of AI capabilities depends on a technical infrastructure that connects language models, account management systems, fake websites, and distribution channels. Russian influence operations increasingly combine these elements into interconnected structures. AI automates selected production and management tasks, while conventional digital platforms and human operators continue to coordinate the wider process.
Meliorator illustrates how AI capabilities can be integrated into a technical infrastructure for managing artificial personas at scale. Its architecture combined specialized software for identity creation, account administration, and automated activity. An administrative interface called Brigadir allowed operators to manage fake identities, while a server application named Taras executed scripts that controlled their behavior. This integration enabled operators to coordinate large numbers of artificial personas and automate their activity across online platforms.
This architecture allowed operators to manage entire account networks from a centralized control point. It also separated content generation from account management and distribution, making it easier to coordinate large-scale activity.
Doppelganger and CopyCop/Storm-1516 demonstrate how fake media networks depend on distributed digital infrastructure to maintain content production and distribution. Their operators combine fabricated news websites with advertising systems, social media accounts, and intermediary networks. The scale of this infrastructure is substantial. Doppelganger has used dozens of domains, while CopyCop expanded to more than 300 websites in 2025. Both operations rely on infrastructure spread across multiple providers and jurisdictions, which makes disruption more difficult, while mirror websites and backup domains allow operators to restore access when individual sites face restrictions or lose their hosting.
The technical implementation also points to a shift toward greater operational independence. CopyCop appears to use self-hosted language models, including locally run versions of Llama, supported by tools such as Open WebUI and Ollama. This setup allows operators to run generative AI without relying on commercial platforms and their access restrictions. Combined with distributed hosting and backup domains, it gives influence networks greater control over both content production and the infrastructure used to distribute it.
Access to a comprehensive computing infrastructure further supports this trend. Russian platforms, including Yandex AI Studio, provide options for deploying language models on GPU hardware. In addition, generating content with existing models requires considerably fewer computing resources than training foundational models from scratch.
The resulting infrastructure is increasingly distributed across multiple technical and organizational components. AI supports content production and selected management tasks, while websites, advertising services, social media platforms, and external networks provide distribution. Self-hosted models may further increase the resilience of these operations by reducing their dependence on foreign service providers.
Potential Risks in Complex Hybrid Operations
The development of AI-enabled influence infrastructure creates additional opportunities to combine information operations with intelligence gathering, recruitment, and physical sabotage. The main concern is the ability to automate parts of the preparation and information support for hybrid activities, reducing the time and resources required to organize them.
Existing digital recruitment practices provide a basis for this development. Investigations have documented the use of Telegram bots and messaging applications to recruit individuals for espionage and sabotage operations in Europe. AI could expand these mechanisms by automating initial contact, conducting conversations in multiple languages, and screening potential recruits.
Another potential application involves using established influence networks to provide information cover for physical attacks. Doppelganger and CopyCop/Storm-1516 have demonstrated the ability to produce large volumes of fabricated articles, imitate independent media outlets, and distribute political narratives through interconnected platforms.
Similar capabilities could help operators respond rapidly to sabotage incidents. They could generate alternative explanations, attribute responsibility to third parties, or question the credibility of official investigations. Coordinated publication across multiple websites and social media accounts could reinforce these claims and complicate the public response.
Documented links between individuals associated with Storm-1516 and groups connected to Russian military intelligence add to concerns about the possible integration of information operations with physical activities.
The main risk lies in the ability to provide rapid and large-scale information support for hybrid operations. AI could help organizers produce competing explanations, deepen public confusion, undermine confidence in state institutions, and increase the political impact of physical interference without a proportional increase in resources.