AI-Driven Disinformation Swarms: Architectures, Cognitive Effects, and Systemic Resilience#

1. Executive Summary#

The global digital information ecosystem is undergoing a fundamental paradigm shift, catalyzed by the rapid proliferation of advanced computational linguistics and generative artificial intelligence (AI). The era of brittle, centrally scripted influence operations—characterized by human troll farms and rigid, copy-paste botnets—is rapidly receding. In its place, a new threat vector has emerged: the AI-driven disinformation swarm. These swarms represent a synthesis of large language model (LLM) reasoning capabilities and multi-agent architectures, enabling the orchestration of decentralized, highly adaptive, and massively scalable manipulation campaigns. The strategic objective of these operations has evolved; it is rarely the successful persuasion of a target populace regarding a singular, verifiable falsehood. Rather, the objective is the deliberate, systemic degradation of shared reality, achieved through the weaponization of cognitive fatigue, epistemic exhaustion, and censorship by noise. This comprehensive report, synthesized from the intersection of network science, complexity theory, cybersecurity, and political communication, interrogates the conceptual, technical, and psychological parameters of disinformation swarms. It establishes a rigorous taxonomy separating emergent AI swarms from conventional digital phenomena, delineates the theoretical infrastructure required to sustain such swarms, and explores their profound socio-psychological impacts on democratic deliberation. Through detailed empirical analyses of recent state-aligned networks—namely the Doppelganger, CopyCop, and Portal Kombat campaigns—alongside controlled multi-agent simulations, this analysis highlights the critical limitations of current content-based detection paradigms. Finally, the report addresses the legal ambiguities under international law and proposes systemic, friction-based resilience measures designed to safeguard democratic discourse without infringing upon legitimate expression.

2. Definitions and Taxonomy#

A disinformation swarm is formally defined as a coordinated or emergent collection of accounts, agents, media assets, or automated systems that produces, varies, distributes, and amplifies misleading narratives at high speed or scale1. A swarm transcends simple algorithmic repetition; it is characterized by linguistic heterogeneity, context-aware semantic adaptation, and dynamic role allocation across its constituent nodes. A swarm may promote a specific strategic narrative, introduce multiple contradictory narratives simultaneously, attack reliable institutional sources, overwhelm factual verification systems, simulate artificial public disagreement, or create generalized epistemic uncertainty1. To achieve analytical clarity, it is necessary to distinguish an AI-driven disinformation swarm from adjacent, often conflated, information phenomena. The structural, behavioral, and technological distinctions are outlined below.

This taxonomy directly addresses the question of what makes a swarm qualitatively different from ordinary automated amplification. The qualitative differentiator is adaptive heterogeneity combined with autonomous sociability. Ordinary automation amplifies a static signal; a swarm amplifies a dynamic, continuously mutating signal. By leveraging generative AI to dynamically adjust tone, vocabulary, cultural markers, and ideological framing for specific micro-communities, swarms bypass traditional detection heuristics. Furthermore, ordinary amplification is a one-way broadcast. Swarms, conversely, simulate multi-directional human interactions—such as debate, agreement, skepticism, and conversion—creating a synthetic social fabric that is cognitively indistinguishable from human communities2.

3. Historical Precedents#

The lineage of the disinformation swarm traces a trajectory from information scarcity to information abundance, reflecting broader shifts in authoritarian control mechanisms and media capture. Historically, political manipulation relied on traditional censorship—the active suppression of dissenting information. However, as global internet penetration accelerated, states realized that controlling the flow of information was increasingly futile. This realization birthed the ideology of information abundance, where the objective shifted from blocking truth to drowning it out. In the early 2010s, algorithmic amplification was largely exploited by rudimentary scripts that simply clicked "like" or shared identical text strings. The evolutionary leap occurred with the professionalization of human-staffed "troll farms." These human operators possessed the cognitive flexibility to engage authentically, understand cultural nuance, and inject sarcasm or targeted outrage. However, troll farms suffered from an inherent operational bottleneck: human labor does not scale infinitely, and human operators exhibit observable shift patterns, fatigue, and language errors that aid in attribution. By the late 2010s, researchers observed hybrid networks consisting of centralized human controllers directing vast, automated distribution networks. During this period, the tactic of "flooding the zone"—rooted in the "firehose of falsehood" model—became dominant. This doctrine prioritized high volume, rapid dissemination, and continuous, repetitive messaging over consistency or objective truth5. The advent of accessible, open-source Large Language Models in the early 2020s eliminated the final friction point: content generation. State-aligned actors immediately integrated these models, successfully bridging the gap between the infinite scale of automation and the nuanced adaptability of human interaction, giving rise to the modern disinformation swarm13.

4. Swarm Architectures and Behavioral Patterns#

Understanding the mechanical reality of a swarm requires defining the conceptual infrastructure necessary to sustain it, as well as the observable behaviors it manifests in the wild.

Conceptual Technical Infrastructure#

At a foundational level, a malicious AI swarm requires a robust, multi-layered technical architecture designed for resilience and scale. The base layer consists of self-hosted, uncensored large language models. Threat actors increasingly rely on open-source models (such as jailbroken iterations of Meta's Llama 3\) running on private servers15. Self-hosting is paramount because it evades the safety filters, rate limits, and telemetry of commercial AI service providers, allowing the generation of explicitly harmful or biased content without triggering provider-side takedowns16. The middle layer comprises the orchestration engine—a multi-agent framework. This engine manages thousands of persistent agent "personas." Each persona is endowed with a unique system prompt acting as its "memory," dictating its demographic profile, political leanings, behavioral parameters (e.g., level of aggression, propensity for sarcasm), and linguistic quirks8. The top layer is the distribution and amplification infrastructure. This includes vast arrays of anonymized residential proxies, programmatic API access or headless browser automation to interface with social platforms, and networks of cloned or synthetic media websites. These websites act as the narrative anchor, hosting the fabricated articles that the social media personas will subsequently cite and distribute15.

Autonomous vs. Semi-Autonomous Operations#

A critical operational question is whether current AI systems can coordinate complex social behavior independently, or if apparent swarms still depend on human direction. Empirical evidence suggests we are currently in a transitional phase characterized by semi-autonomous swarms. Operations like CopyCop and Doppelganger still rely on human orchestrators to set high-level strategic objectives (e.g., "exacerbate domestic tensions in France prior to the election") and to construct the physical infrastructure (e.g., registering domains, purchasing proxy networks)14. However, the tactical execution is entirely automated. Translating human directives into hundreds of unique articles, generating headlines, simulating journalistic personas, and engaging in social media commentary are handled by the multi-agent engine13. While controlled experiments in academic settings prove that LLM-driven agents can engage in fully autonomous complex social behavior, emergent opinion dynamics, and self-directed persuasion8, real-world threat actors currently maintain human-in-the-loop oversight to ensure narratives remain tightly aligned with specific geopolitical objectives and do not hallucinate counter-productive messaging.

Exploitative Swarm Behaviors#

Swarms exhibit a sophisticated repertoire of tactical behaviors designed to manipulate both algorithmic curation systems and human cognitive vulnerabilities. These behaviors include:

• Rapid Production of Narrative Variants: Instead of deploying a single, easily identifiable fake news article, the orchestration engine funnels a core falsehood through an LLM, commanding it to produce thousands of rhetorically diverse variants. This linguistic heterogeneity prevents platforms from utilizing simple hash-matching or text-similarity algorithms for moderation4.
• Division of Roles Among Accounts: Multi-agent frameworks assign distinct roles to personas within the swarm. Some agents act as "news aggregators," blindly broadcasting links. Others act as "skeptical citizens" who demand evidence. A third group provides the fabricated evidence, leading the "skeptic" to publicly announce they have been convinced. This simulates a natural process of persuasion2.
• Artificial Debate: To game algorithmic engagement metrics, swarms manufacture debate between supposedly opposing personas within the same controlled network. Because recommendation engines reward high-interaction threads, this synthetic conflict artificially elevates the visibility of the underlying premise to unaware human observers.
• Simultaneous Targeting of Multiple Platforms: A swarm does not reside in a silo. It executes cross-platform pollination. A fabricated intelligence dossier is hosted on a synthetic blog; an AI persona tweets a link to the blog; another persona creates a TikTok video summarizing the tweet; a fourth persona posts the TikTok video to a Reddit forum. This creates a highly resilient web of synthetic verification19.
• Repetition Combined with Constant Mutation: The overarching narrative is relentless, exploiting the illusory truth effect (where repeated statements are perceived as more truthful). However, the specific wording constantly evolves, evading static detection rules15.
• Flooding Information Channels During Crises: Swarms operate with zero latency. During a crisis, they instantly flood timelines with highly plausible, AI-generated synthetic media, dominating the narrative space before human fact-checkers can verify the reality of the situation20.
• Attacking Trusted Intermediaries: Swarms actively attempt to delegitimize authoritative sources—such as electoral commissions, public health bodies, or mainstream investigative journalists—by generating massive volumes of fabricated evidence alleging bias, corruption, or incompetence.
• Creating Contradictory Explanations: When a state actor commits a controversial act, the swarm is deployed not to push one exculpatory narrative, but to push fifty mutually exclusive theories. The goal is to make the truth appear entirely unknowable, paralyzing public consensus3.
• Adapting to Moderation: If a specific phrase, URL, or domain is blacklisted by a platform, the orchestration engine seamlessly substitutes alternative vocabularies and shifts traffic to mirrored subdomains without human intervention15.

5. Psychological and Social Effects#

The impact of disinformation swarms extends far beyond the temporary injection of false data into a social network; these systems fundamentally compromise the cognitive and epistemic environment in which democratic deliberation occurs. The vectors of volume, diversity, repetition, speed, and cross-platform activity operate synergistically to exploit human cognitive heuristics. The "wisdom of crowds" phenomenon, essential for democratic societies, relies entirely on the statistical independence of individual judgments1. When an AI swarm simulates a diverse, geographically distributed chorus of independent voices across multiple platforms, it creates a "synthetic consensus"1. Human beings are evolutionarily wired to rely on social-proof heuristics; when confronted with an overwhelming volume of diverse voices seemingly agreeing on a topic, humans subconsciously update their own opinions to align with what they falsely perceive to be the majority norm, regardless of empirical evidence. This leads to a critical operational reality: confusion is often a more realistic and desirable objective than belief in a specific false claim. Advanced swarms deploy the "firehose of falsehood" to overwhelm the target's cognitive processing capacity. The resulting phenomenon is termed epistemic exhaustion—the profound cognitive fatigue generated by the ceaseless effort required to determine, retain, or communicate truth under conditions of extreme informational chaos23. As documented in contemporary political communication literature, when citizens face a relentless barrage of contradictory, highly emotive, and plausible information, they frequently experience informational learned helplessness24. This is the essence of censorship through noise22. Instead of seeking objective truth, exhausted citizens retreat into "reactive partisanship," doubling down on tribal epistemology because accepting the dominant narratives of one's ingroup requires far less cognitive effort than independent verification23. The sheer volume of AI-generated content guarantees that the necessary human defense mechanism of "critical ignoring" eventually results in citizens ignoring legitimate democratic discourse alongside the malicious noise, fostering a cynicism that inherently benefits authoritarian actors21.

6. Case Studies#

To ground the theoretical framework in empirical reality, this section examines three documented, real-world campaigns and one controlled academic experiment involving coordinated information activity, evaluating the confirmed use of generative AI and their operational impacts.

Case Study 1: The Doppelganger Campaign#

Originating from Russia and attributed to entities sanctioned by the European Union—such as the Social Design Agency and Structura National Technologies—the Doppelganger campaign is one of the most persistent and aggressively voluminous influence operations discovered to date11.

• Operational Methodology: The campaign's signature tactic is "typosquatting"—registering domain names virtually identical to trusted mainstream media outlets (e.g., Le Monde, Der Spiegel, The Washington Post) and government institutions (e.g., the French Ministry of Public Affairs). The swarm then clones the visual architecture of these sites to host fabricated articles designed to undermine Western support for Ukraine, exacerbate domestic tensions regarding migration, and inflame societal divisions19.
• Generative AI Confirmation: The use of generative AI is confirmed. Threat intelligence reports from OpenAI, Meta, and the German Foreign Office indicate Doppelganger utilized LLMs to translate articles from Russian, write original short-form commentary, generate sensationalist headlines, and convert long-form news copy into optimized social media posts across English, French, German, Polish, and Ukrainian13.
• Impact Analysis: While the operation generated massive volume—automating posts at rates exceeding one tweet per second through hundreds of thousands of disposable bot accounts—its measurable impact on authentic user engagement was remarkably low19. Doppelganger represents a brute-force approach. It lacks the nuanced, multi-agent conversational depth of a true swarm, instead relying on sheer persistence to slowly normalize hate speech and alter the baseline of online discourse11.

Case Study 2: CopyCop (Storm-1516)#

Tracked extensively by Recorded Future's Insikt Group, CopyCop represents a highly evolved, second-generation iteration of the generative AI swarm, demonstrating significant operational security and adaptive capabilities15.

• Operational Methodology: Operating a network of over 300 inauthentic local news websites, political party fronts, and fake fact-checking organizations, CopyCop systematically plagiarizes content from legitimate media outlets15. The network utilizes a distributed infrastructure built to withstand disruption; when one domain is taken offline, mirrored copies instantly appear elsewhere17.
• Generative AI Confirmation: Confirmed. Unlike earlier operations that used commercial APIs, CopyCop operators migrated to self-hosted, uncensored versions of Meta's Llama-3 open-source models15. Researchers identified explicit AI artifacts in the published text (e.g., "Please note that this rewrite aims to provide a clear and concise summary...") proving the models were prompted to rewrite factual articles to introduce pro-Russian and anti-Ukrainian political bias14.
• Impact Analysis: CopyCop scaled massively, actively attempting to influence the 2024 US presidential elections, the French snap elections, and parliamentary elections in Moldova and Armenia15. By mimicking the stylistic cues of legitimate reporting—complete with bylines and fake investigative dossiers—CopyCop successfully created a disinformation ecosystem that deeply infiltrated online discussions and search engine results17.

Case Study 3: Portal Kombat#

Uncovered by VIGINUM, the French state agency for protection against digital interference, Portal Kombat comprised a vast network of at least 193 digitally native information portals15.

• Operational Methodology: The network flooded Western information spaces with highly coordinated, uniform messaging designed to justify the invasion of Ukraine and disparage Western leadership. While technically simpler in its initial architecture than CopyCop—relying heavily on content aggregation rather than generative synthesis—Portal Kombat served as a foundational proof-of-concept for mass, cross-border content distribution designed specifically to achieve censorship through noise15.

Controlled Experiment: Generative Propaganda and Agent-Based Modeling#

A critical study by Wack et al. (Clemson University) observed a quasi-experimental transition within a state-backed propaganda site (DC Weekly, linked to the CopyCop infrastructure). Researchers identified the exact moment—September 20, 2023—when the outlet transitioned from manual copy-pasting of articles to AI-assisted generative rewriting4.

• Generative AI Confirmation: Confirmed. The study tracked the sudden shift from exact language matches to seemingly original text generated by LLMs4.
• Impact Analysis: The adoption of generative AI allowed the outlet to drastically increase the quantity of disinformation and the breadth of narratives covered. Crucially, through human survey experiments, the researchers proved that the AI-assisted articles maintained their persuasiveness and credibility among readers in the post-adoption period. AI supercharged production capabilities without sacrificing the quality required for deception5.
• Multi-Agent Validation: Parallel controlled experiments in computer science demonstrate that LLM-driven agents can act as highly accurate proxies for human opinion dynamics. When placed in Agent-Based Models (ABMs), LLM agents effectively simulate human cognitive biases, polarization, and susceptibility to misinformation headlines. Neural language models programmed with minimal belief systems successfully generated complex argumentative dynamics, proving that the multi-agent architectures underpinning swarms are theoretically sound and practically lethal8.

7. Network Detection and Attribution#

Detecting AI-driven swarms requires a fundamental paradigm shift for researchers and trust-and-safety teams. Because LLMs generate vast linguistic diversity for every post, traditional natural language processing (NLP) heuristics that search for identical copy-pasted strings are easily defeated. The focus must shift from content-based moderation to structural and behavioral network analysis. How can network analysis identify coordinated behavior when messages are linguistically diverse? It does so by focusing on the metadata of dissemination—how the content moves through the network, rather than what the content says. Techniques such as Coordinated Link Sharing Behavior (CLSB), pioneered by researchers like Fabio Giglietto, trace the dissemination topology of URLs6. If an array of seemingly disparate accounts, pages, or groups shares links pointing to the same underlying infrastructure (e.g., a cluster of CopyCop fake news domains) in highly synchronized temporal windows, the structural graph reveals the swarm. This topological mapping is entirely agnostic to the varied, AI-generated text accompanying the links7. Distinguishing malicious coordination from coincidental similarity or organic, shared political enthusiasm relies on rigorous statistical thresholds. In CLSB models, researchers utilize inter-arrival time (the delta between consecutive shares of the same link) and repetition frequency. For instance, if two entities share the same link within a statistically improbable timeframe (e.g., under 27 seconds) and repeat this synchronized behavior dozens of times across unrelated news cycles, statistical models indicate algorithmic coordination rather than coincidence6. Network position metrics—such as degree centrality, eigenvector centrality, and PageRank—are further utilized to map the hierarchy of the swarm, identifying the orchestrator nodes versus the amplifier nodes6. Methodological Weaknesses: Despite these advancements, structural detection faces severe methodological vulnerabilities.

1. Bot-Detection Decay: Traditional bot-detection algorithms (like Botometer) are rapidly losing efficacy. LLMs allow bots to exhibit highly "human-like" variance in posting schedules, sentiment, and vocabulary, depriving detection models of the repetitive anomalies they were trained to flag5. 2. API Obsolescence: Structural analysis requires massive, unimpeded access to platform data. The systematic closure of platform APIs and the deprecation of vital research tools—most notably Meta's shutdown of CrowdTangle in August 2024—have severely handicapped independent network scientists, trapping vital topological data behind opaque corporate walls31.

8. Platform and Fact-Checking Challenges#

The integration of generative AI into disinformation pipelines has created a highly asymmetrical warfare environment. The "firehose of falsehood" mathematically guarantees that falsehood production will perpetually outpace human verification capacity. When a multi-agent swarm can generate 10,000 unique, plausible variations of a fake narrative in seconds, debunking individual claims is a futile, Whac-A-Mole endeavor. How should fact-checkers prioritize claims when production exceeds capacity? The discipline must pivot from item-level verification to source-level and narrative-level triage30.

1. Narrative Inoculation (Pre-bunking): Instead of chasing specific micro-claims, fact-checkers must identify the meta-narrative early (e.g., "The upcoming election is physically compromised by imported ballots") and issue pre-bunking warnings that inoculate populations against the theme, regardless of the specific AI-generated anecdote used to deliver it. 2. Harm-Based Triage: Prioritization must be strictly based on the potential for immediate, real-world harm—such as imminent violence, voter suppression tactics, or public health emergencies—rather than the mere presence of falsehood. 3. Structural Debunking: Fact-checkers must collaborate with network scientists to expose the underlying infrastructure. Identifying a cluster of 300 websites as a Russian intelligence front, rather than debunking the 3,000 articles they produce daily, delegates the verification task from the content to the origin17.

9. Crisis-Response Strategies#

During periods of acute societal vulnerability—such as elections, pandemics, natural disasters, or the outbreak of kinetic conflict—there is an inevitable delay in the production of verified, authoritative journalism. Swarms are explicitly programmed to exploit this breaking-news information vacuum. Operating with zero latency, they instantly flood communication channels with highly plausible, AI-generated synthetic media, deepfakes, and explanatory narratives. Because platform algorithms inherently favor recency and high user engagement, the swarm's fabrications achieve dominance before fact-checkers can even begin verification. This establishes a first-mover advantage, forcing authoritative sources into a defensive, reactive posture20. Defensive interventions must reduce the distribution of this malicious content without strengthening martyrdom narratives or suppressing legitimate collective speech. Explicit censorship (outright content deletion) often triggers the Streisand effect, allowing malicious actors to claim martyrdom and validating conspiracy theories about institutional suppression21. Therefore, platforms must employ friction-based interventions:

• Reverse Amplification: Adjusting algorithmic weighting to silently demote content originating from clusters exhibiting high Coordinated Link Sharing Behavior (CLSB), significantly reducing reach without actively deleting the posts or notifying the actor32.
• Circuit Breakers: Implementing temporary, automated rate limits on the viral coefficient of entirely novel URLs or sensational claims during breaking news events, introducing artificial friction that allows authoritative consensus time to catch up.
• Demonetization and Algorithmic Isolation: Stripping the ability for coordinated accounts to monetize content or utilize algorithmic recommendation engines, confining their reach strictly to the feeds of their explicit, existing followers.

10. Legal and International-Security Implications#

State-sponsored disinformation swarms exist in a precarious grey zone of international law, complicating diplomatic response, attribution, and deterrence. What international norms should govern state use of autonomous or semi-autonomous influence networks? The prevailing framework for cyber operations under international law is articulated in the Tallinn Manual. Under this framework, a cyber operation—including a digital influence operation—rarely meets the threshold of a "use of force" (jus ad bellum). However, it can be deemed a violation of a target state's sovereignty if it constitutes coercion33. The Tallinn Manual defines coercion narrowly: compelling a state to take, or omit, an action it would otherwise freely choose, particularly regarding its internal political processes34. The deployment of a disinformation swarm to generally erode public trust, foster epistemic exhaustion, or create noise does not easily satisfy this strict legal definition of coercion. However, if a state-directed AI swarm successfully suppresses voter turnout through targeted micro-suppression, or manipulates an electoral outcome to the extent that it dictates a nation's political trajectory, a compelling argument for illegal intervention and coercion emerges. Establishing robust international norms requires recognizing that the scale, autonomy, and psychological impact of AI swarms elevate them from traditional public diplomacy or propaganda into the realm of structural cyberattacks on a nation's cognitive infrastructure.

11. Defensive Resilience Measures#

Building systemic resilience against AI swarms requires a layered, defense-in-depth strategy, moving beyond the inadequate paradigm of individual media literacy to systemic authentication.

• Model-Side Safeguards: While implementing standardized persuasion-risk tests and watermarking in commercial models is necessary, it is insufficient. Because threat actors like CopyCop utilize self-hosted, open-source models, commercial safeguards are easily bypassed15.
• Cryptographic Provenance: Resilience requires the widespread adoption of cryptographic integrity protocols, such as the Coalition for Content Provenance and Authenticity (C2PA). By embedding cryptographic watermarks at the hardware level (cameras) and software level (publishing platforms), users can verify the origin, edit history, and authenticity of media. In a mature ecosystem, content lacking cryptographic provenance would inherently carry a lower trust score, shifting the burden of proof from the defender to the publisher16.
• Client-Side AI Shields: Empowering end-users with local, client-side AI tools designed to act as cognitive firewalls. These tools can analyze incoming information streams, detecting linguistic manipulation, logical fallacies, and network anomalies before the user cognitively processes the payload, providing a real-time defense against epistemic exhaustion35.

12. Future Warning Indicators#

The current threat landscape is characterized by semi-autonomous swarms requiring human strategic input. Analysts and intelligence agencies must monitor for critical trailing and leading indicators pointing toward the emergence of fully autonomous malicious swarms:

1. AI Training Data Contamination (Model Collapse): As swarms flood the internet with synthetic consensus, future LLMs scraping the web will inevitably ingest this fabricated reality. This phenomenon, where the epistemic substrate of future AI tools is poisoned, serves as a trailing indicator of massive, unchecked swarm activity1. 2. Adversarially Induced Norm Shifts: The moment multi-agent systems abandon engineered constraints and develop new, unprompted behavioral patterns to optimize manipulation metrics without human oversight, representing a loss of control by the original orchestrators2. 3. Real-Time Tactical Mutability: Swarms autonomously probing platform algorithms via reinforcement learning, discovering zero-day vulnerabilities in recommendation engines, and exploiting them instantly without human direction.

13. Research Gaps#

Despite rapid advances in computational social science, significant empirical and methodological gaps remain in our understanding of disinformation swarms:

• Platform Transparency: The systemic closure of platform APIs severely limits independent topological research into coordinated behavior, leaving civil society blind to network-level manipulation31.
• Longitudinal Opinion Dynamics: Current Agent-Based Models utilizing LLMs demonstrate short-term behavioral mimicry, but there is a profound lack of real-world longitudinal data validating how human populations react to prolonged, multi-year epistemic exhaustion and synthetic consensus10.
• Cross-Lingual Network Analysis: While CLSB is highly effective within closed, single-language ecosystems, tracing the semantic translation of an AI-generated narrative as it jumps platforms—for example, from a Russian Telegram channel to an English X thread, to a French TikTok video—remains computationally prohibitive.

14. Conclusion#

The convergence of multi-agent architectures and large language models has fundamentally altered the threat landscape of global information environments. AI-driven disinformation swarms represent a stark departure from historical persuasion campaigns; they are highly scalable instruments of cognitive denial-of-service, designed to induce epistemic exhaustion, manufacture synthetic consensus, and fracture democratic reality. As demonstrated by the operational reality of the Doppelganger and CopyCop networks, the capability to mass-produce contextually nuanced, linguistically diverse, and logically coherent synthetic media is actively being deployed by state-aligned actors. Defending against this evolution requires a complete departure from traditional content moderation. When truth and falsehood are visually and linguistically indistinguishable, systemic resilience must be built upon structural network analysis, algorithmic friction, and cryptographic provenance. The international community must recognize that manipulating a populace's cognitive environment through high-volume, automated artificial intelligence constitutes a systemic threat to democratic sovereignty, demanding a unified, interdisciplinary, and technically rigorous response.

Typology of Influence Networks#

15. Annotated Bibliography#

1. Giglietto, F., et al. (2020). Coordinated Link Sharing Behavior as a Signal to Surface Sources of Problematic Information on Facebook. \[cite: 6, 7\] This foundational study in network science outlines the methodology for detecting Coordinated Inauthentic Behavior (CIB) by analyzing the topological dissemination of URLs. The authors define Coordinated Link Sharing Behavior (CLSB) based on highly repetitive, unusually rapid sharing of identical links, providing a structural detection framework that bypasses the limitations of natural language processing in the age of generative AI. 2. Schroeder, D. T., et al. (2026). How Malicious AI Swarms Can Threaten Democracy. Science.1 A comprehensive, interdisciplinary analysis defining the concept of malicious AI swarms resulting from the fusion of agentic AI and LLMs. The authors meticulously categorize the cascading harms to democratic institutions, heavily emphasizing "synthetic consensus," the erosion of collective intelligence ("wisdom of crowds"), and the eventual contamination of future AI training data through information flooding. 3. Wack, M., et al. (2025). Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign. PNAS Nexus.4 This crucial quasi-experimental study analyzes the Russian-linked CopyCop/DC Weekly network, establishing empirical proof of when the network transitioned from human-directed copy-pasting to generative AI. The study confirms that the use of AI dramatically increased the output volume and narrative breadth of the disinformation without sacrificing persuasiveness or credibility among target audiences. 4. Recorded Future Insikt Group (2024-2025). Threat Intelligence Reports on "CopyCop" (Storm-1516).14 A series of technical threat reports detailing the operational infrastructure of a highly evolved Russian influence network. The reports confirm the network's use of self-hosted, uncensored Llama 3 models to plagiarize, rewrite, and weaponize Western media content across over 300 synthetic news websites, highlighting the modern architecture of semi-autonomous disinformation swarms. 5. EU DisinfoLab / VIGINUM / Meta Threat Reports (2022-2024). Investigations into the "Doppelganger" and "Portal Kombat" campaigns.19 Detailed forensic analyses of massive, state-aligned information manipulation sets. These reports document the mechanics of typosquatting, the cloning of authoritative government and media websites, and the initial integration of generative AI to facilitate multilingual propaganda dissemination at an unprecedented scale. 6. Li, et al. (2024) / Betz, G. (2022). Generative Agents and Natural-Language Multi-Agent Simulations of Argumentative Opinion Dynamics. \[cite: 8, 29, 30\] A body of research in complexity science and agent-based modeling (ABM) that substitutes traditional mathematical representations of human belief with natural language LLM agents. These studies prove that AI systems can coordinate complex social behaviors, effectively mirroring human susceptibility to misinformation, polarization, and echo-chamber dynamics, thus providing a theoretical foundation for understanding how malicious swarms operate.

Works cited#

1. Citation: Schroeder et al. How Malicious AI Swarms Can Threaten Democracy. 1-8…. DOI:000000/11111. Corresponding author: daniel.t.schroeder@sintef.no; †These authors contributed equally to this work; The authors are listed in alphabetical order by last name, starting from the third and ending with the second-to-last. - arXiv, https://arxiv.org/html/2506.06299v1 2. How malicious AI swarms can threaten democracy - Harvard Business School, https://www.hbs.edu/ris/download.aspx?name=How%20malicious%20AI%20swarms%20can%20threaten%20democracy.pdf 3. A New Paradigm for Global Journalism: Press Freedom and Public Interest, https://www.cjr.org/tow\_center\_reports/a-new-paradigm-for-global-journalism-press-freedom-and-public-interest.php 4. Generative propaganda - Oxford Academic, https://academic.oup.com/pnasnexus/article-pdf/4/4/pgaf083/62785143/pgaf083.pdf 5. Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC11950819/ 6. Coordinated Link Sharing Behavior as a Signal to Surface Sources of Problematic Information on Facebook | Request PDF - ResearchGate, https://www.researchgate.net/publication/342535909\_Coordinated\_Link\_Sharing\_Behavior\_as\_a\_Signal\_to\_Surface\_Sources\_of\_Problematic\_Information\_on\_Facebook 7. Coordinated link sharing on Facebook - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12053595/ 8. Simulating Misinformation Vulnerabilities With Agent Personas - arXiv, https://arxiv.org/html/2511.04697v1 9. Natural-Language Multi-Agent Simulations of Argumentative Opinion Dynamics - JASSS, https://www.jasss.org/25/1/2.html 10. Opinion dynamics and mutual influence with LLM agents through dialog simulation - arXiv, https://arxiv.org/pdf/2602.12583 11. Russia-linked 'Doppelgänger' social media operation rolls on, report says, https://therecord.media/doppelganger-influence-operation-new-activity 12. Russia-Linked CopyCop Expands to Cover US Elections, Target Political Leaders, https://www.recordedfuture.com/research/copycop-expands-to-cover-us-elections-target-political-leaders 13. Disrupting deceptive uses of AI by covert influence operations - OpenAI, https://openai.com/index/disrupting-deceptive-uses-of-ai-by-covert-influence-operations/ 14. Russia-Linked CopyCop Uses LLMs to Weaponize Influence Content at Scale - Recorded Future, https://assets.recordedfuture.com/insikt-report-pdfs/2024/cta-2024-0509.pdf 15. CopyCop Deepens Its Playbook with New Websites and Targets - Recorded Future, https://www.recordedfuture.com/research/copycop-deepens-its-playbook-with-new-websites-and-targets 16. Authenticity Debt and the Synthetic Content Threat Landscape: A Layered Framework for Trust, Provenance, and IP Governance in the Generative AI Era - arXiv, https://arxiv.org/html/2606.00621v1 17. Inside the CopyCop Playbook: How to Fight Back in the Age of Synthetic Media, https://www.recordedfuture.com/blog/inside-the-copycop-playbook 18. Simulating Opinion Dynamics with Networks of LLM-based Agents - OpenReview, https://openreview.net/forum?id=wLHI2xjmMW 19. Germany Targeted by the Pro-Russian Disinformation Campaign “Doppelgänger”, https://www.auswaertiges-amt.de/resource/blob/2682484/2da31936d1cbeb9faec49df74d8bbe2e/technischer-bericht-desinformationskampagne-doppelgaenger-1--data.pdf 20. Detecting Coordinated Link Sharing During The Italian Coronavirus Outbreak, https://compcommlab.univie.ac.at/research-teaching/publications/detailseite-publications/pure/6c8b1e96-67b7-40e4-822a-0a35f34a3201/show/publ/Pure/ 21. Key social media risks to democracy - European Parliament, https://www.europarl.europa.eu/RegData/etudes/IDAN/2021/698845/EPRS\_IDA(2021)698845\_EN.pdf698845_EN.pdf) 22. Faking it - Royal Society of Arts, https://www.thersa.org/rsa-journal/faking-it/ 23. Epistemic Exhaustion and the Retention of Power | Hypatia | Cambridge Core, https://www.cambridge.org/core/journals/hypatia/article/epistemic-exhaustion-and-the-retention-of-power/D53FD093288CDC9DAAB02DE4C0740ACE 24. Media Capture, Misinformation, and “Noise” | Psychology Today Australia, https://www.psychologytoday.com/au/blog/misinformation-desk/202603/media-capture-misinformation-and-noise 25. Is Fake News Old News? - VU Research Portal, https://research.vu.nl/files/153103900/Is\_Fake\_News\_Old\_News.pdf 26. AI in Disinformation Detection - Applied Cybersecurity & Internet Governance, https://www.acigjournal.com/AI-in-Disinformation-Detection,200200,0,2.html 27. Doppelganger hub - EU DisinfoLab, https://www.disinfo.eu/doppelganger-hub/ 28. France uncovers Russia's disinformation campaign justifying war in Ukraine | Ukrainska Pravda, https://www.pravda.com.ua/eng/news/2024/02/12/7441555/ 29. Natural-Language Multi-Agent Simulations of Argumentative Opinion Dynamics, https://ideas.repec.org/a/jas/jasssj/2021-59-3.html 30. Large Language Model-driven Multi-Agent Simulation for Fake News Diffusion Under Different Network Structures - AUP-Online, https://www.aup-online.com/content/journals/10.5117/CCR2026.2.8.LI?crawler=true\&mimetype=application/pdf 31. Home | Fabio Giglietto | Full Professor of Internet Studies, https://fabiogiglietto.github.io/ 32. Free Speech, Platforms, and the Fake News Problem - CRESSE, https://www.cresse.info/wp-content/uploads/2022/10/2022\_ps11\_pa2\_Van-Alstyne.pdf 33. Scenario 01: Election interference - International cyber law: interactive toolkit, https://cyberlaw.ccdcoe.org/wiki/Scenario\_01:\_Election\_interference 34. Hacking the Domaine Réservé: The Rule of Non-Intervention and Political Interference in Cyberspace, https://journals.law.harvard.edu/ilj/wp-content/uploads/sites/84/HLI115\_crop.pdf 35. How Malicious AI Swarms Can Threaten Democracy - OSF, https://osf.io/preprints/osf/qm9yk\_v4 36. How Malicious AI Swarms Can Threaten Democracy: The Fusion of Agentic AI and LLMs Marks a New Frontier in Information Warfare - Article - Faculty & Research - Harvard Business School, https://www.hbs.edu/faculty/Pages/item.aspx?num=68488