Red-Teaming Viral Features and Story Formats: A Comprehensive Vulnerability and Mitigation Analysis

Full report

On this page

The contemporary digital information environment is defined by the accelerated transmission of hyper-optimized, highly salient content. As social media platforms, content delivery networks, and digital news aggregators increasingly prioritize user engagement metrics over epistemic veracity, the structural design of these platforms has evolved to maximize cognitive capture. This evolution introduces profound vulnerabilities. Specific platform architectures—specifically their viral features and curated story formats—are increasingly exploited by adversarial actors to execute highly sophisticated information operations, propagate synthetic disinformation, and destabilize societal consensus. The proliferation of generative artificial intelligence (AI) and large language models (LLMs) has catalyzed this threat, transforming what were once manual, resource-intensive propaganda efforts into automated, highly scalable cognitive campaigns that exploit hardwired human heuristics1. The foundational challenge for intelligence and defense communities lies not merely in identifying malicious content post-deployment, but in understanding how the structural mechanisms of content delivery systematically distort meaning. To address these vulnerabilities, the intelligence and defense sectors have initiated extensive, high-risk, high-payoff research programs under the auspices of the Intelligence Advanced Research Projects Activity (IARPA) and the Office of the Director of National Intelligence (ODNI)2. Initiatives such as the BENGAL (Bias Effects and Notable Generative AI Limitations) program actively seek to understand the landscape of LLM threats, quantify vulnerabilities, and develop novel technologies to ensure the safe flow of information while preserving source attribution and mitigating hallucinations3. Concurrently, the HIATUS (Human Interpretable Attribution of Text Using Underlying Structure) program investigates the linguistic fingerprints of authorship, highlighting the dual-use nature of text manipulation for both privacy protection and adversarial obscuration5. Furthermore, spatial and temporal modeling programs like WRIVA (Walk-through Rendering from Images of Varying Altitude) demonstrate how advanced machine learning algorithms can construct photorealistic realities from highly limited data—a capability that, if inverted by adversaries, offers unprecedented tools for geographic and spatial deception7. This report provides an exhaustive red-team analysis of proposed viral features and story formats, including cropping, animation, dramatic wording, ranking, personalization, incomplete timelines, geographic projection, and source omission. By systematically dismantling the mechanisms through which these features distort reality, the analysis identifies cascading risks involving propaganda, political bias, decontextualized conflict footage, false equivalence, harassment, graphic events, manipulated screenshots, and algorithmic amplification. For each identified risk vector, a structural mitigation framework is proposed, encompassing preventive controls, review gates, visible disclosures, and precise rollback conditions.

The Architecture of Visual and Spatial Distortion

Visual media operates as a primary heuristic anchor for human sensemaking. The human brain is neurologically optimized to process visual information rapidly, often bypassing the critical analytical faculties required to evaluate textual claims or complex geopolitical narratives. When viral features manipulate the spatial, temporal, or visual boundaries of an event, they exploit cognitive biases, forcing the observer to construct a false reality based on incomplete or altered data. Historical IARPA programs, such as ICArUS (Integrated Cognitive Neuroscience Architectures for Understanding Sensemaking) and Sirius, have extensively documented how the human brain attempts to make sense of sparse, ambiguous data, often falling victim to cognitive bias when critical context is systematically removed or manipulated9. The following sections red-team specific visual and spatial features to uncover their inherent vulnerabilities.

Cropping and the Weaponization of the Frame

The feature of automated or user-driven image and video cropping is fundamentally designed to optimize content for varying screen ratios—such as the vertical orientation of mobile devices—and to focus the viewer's attention on the most salient subject matter. However, from an adversarial red-team perspective, cropping serves as a powerful mechanism for semantic distortion. By redefining the boundary of an image or video, an adversary can entirely strip an event of its contextual reality. The deliberate removal of the periphery allows malicious actors to isolate a specific physical interaction, facial expression, or kinetic event, presenting it as an unprovoked act of aggression, a staged scenario, or a false flag operation. This mechanism directly facilitates the proliferation of manipulated screenshots and decontextualized conflict footage. In geopolitical conflicts, cropped video feeds are frequently deployed by state-sponsored and non-state actors to obscure the presence of instigators, defensive positioning, or humanitarian context. The viewer, subjected to this hyper-focused, high-stress visual input, relies on prior beliefs and emotional heuristics to fill the contextual void. This phenomenon is significantly exacerbated by the "Google effect," a cognitive distortion wherein individuals exhibit overconfidence in their understanding of a complex event and easily anchor their subsequent beliefs to any digital object on the internet that confirms their preexisting geopolitical biases10. The cropped artifact effectively becomes a poisoned source. As highlighted by the core objectives of the IARPA BENGAL program, extracting reliable information from biased or incomplete sources requires resilient, explainable techniques to infer source intentions3. When a platform introduces automated, algorithmic cropping to enhance the "story format" experience, it inadvertently automates the creation of decontextualized propaganda, removing the friction required to generate highly persuasive disinformation. The second-order effects of this visual distortion are profound and immediate. When cropped, manipulated screenshots go viral, they trigger rapid, often disproportionate emotional responses from the public and policy reactions from decision-makers. The immediate misallocation of diplomatic capital, law enforcement resources, or military readiness based on a fabricated visual narrative degrades institutional trust and societal stability. To mitigate this vulnerability, platform architects must approach cropping not merely as a benign formatting tool for aesthetic enhancement, but as a potentially destructive editorial action that requires cryptographic and heuristic oversight. The mitigation of this risk requires a multi-layered approach to ensure that the original context of the media remains accessible and that malicious manipulations are rapidly quarantined.

Geographic Projection and Spatial Disorientation

Geographic projection features, which map two-dimensional imagery onto interactive 3D globes or localized maps, are highly engaging story formats used to immerse users in a specific locale. The underlying technology of these features mirrors the objectives of IARPA’s WRIVA program, which utilizes advanced machine learning to synthesize novel viewpoints and render high-fidelity, photorealistic 3D site models from a highly limited corpus of ground-level, traffic camera, and satellite imagery7. While WRIVA is explicitly designed to aid intelligence, military, and humanitarian and disaster relief (HADR) personnel by providing a virtual "ground truth" prior to deployment into hazardous environments7, the adversarial inversion of this capability presents a severe systemic threat. When geographic projection features are utilized in social media platforms or news aggregation networks, malicious actors can manipulate the Exif data (Exchangeable Image File Format) or inject synthetic imagery into the spatial model. This distortion seamlessly places decontextualized conflict footage into incorrect geographic zones. For instance, destruction from a historical earthquake or kinetic military strike in one hemisphere can be projected onto a current geopolitical flashpoint in another. Because the story format visually "proves" the location via the high-fidelity geographic projection, the user's epistemic defenses are systematically lowered. The seamless synthesis of imagery and mapping algorithms creates a highly persuasive, yet entirely fabricated, spatial reality. This aligns with the broader challenges of integrating temporally relevant geospatial information, a focus area of IARPA's COSMIC program, which highlights the complexities of maintaining accurate spatio-temporal intelligence in commercial and open-source environments12. This distortion enables sophisticated, state-level propaganda campaigns designed to sow mass panic, misdirect international aid, or justify preemptive military action. The manipulation of spatial reality strikes at the core of public situational awareness. If a viral feature dynamically generates these projections based on unverified user uploads, the platform effectively becomes an engine for geographic disinformation. Ensuring the integrity of these features requires stringent verification of the foundational metadata and the application of rigorous spatial continuity models.

Temporal and Sequential Manipulation

While visual and spatial features manipulate the "where" and "what" of an event, temporal features manipulate the "when" and "why." The human cognitive architecture relies heavily on sequential order to infer causality. Story formats that compress or synthetically alter the temporal flow of events exploit this reliance, allowing adversaries to rewrite history in real-time. Red-teaming these features reveals how minor adjustments in sequence or motion can yield massive shifts in public perception.

Incomplete Timelines and the Fabrication of Causality

Story formats that utilize timeline interfaces—such as ephemeral "stories" or sequential thread aggregations—are designed to condense complex, multi-day events into digestible, rapid-fire highlights. However, the curation process inherent in creating a timeline introduces a critical vulnerability: the strategic omission of intervening events. By selectively presenting specific data points while obscuring others, an adversary can fabricate false causality. This mechanism is frequently used to create false equivalence between two entirely disparate actions or to justify an unprovoked escalation by hiding the initial instigating event. The distortion of timelines is a cornerstone of strategic disinformation and geopolitical warfare. In the context of conflict reporting, an incomplete timeline can portray a standard defensive maneuver as an aggressive first strike. The human brain naturally assumes that events presented in sequence have a direct, unmediated causal relationship. When a platform’s viral features automatically generate these timelines based on trending hashtags or user-curated content clusters, they risk codifying propaganda into a structured, authoritative-looking historical record. The BENGAL program's focus on identifying inputs that attempt to aggregate innocuous facts to derive sensitive or misleading information highlights the profound danger of manipulating data relationships and contextual flow3. Furthermore, research into intelligence forecasting, such as IARPA's FUSE and ForeST programs, underscores the difficulty of accurately predicting geopolitical events even with complete data; introducing deliberately fractured data structures guarantees analytical failure14. The third-order effect of timeline manipulation is the long-term rewriting of historical memory. When incomplete timelines are algorithmically amplified across content delivery networks, they become the dominant narrative, superseding nuanced, comprehensive reporting. This degrades the public's ability to engage in rational, evidence-based discourse, as the foundational facts and timelines of an event are fundamentally contested. To prevent this systemic degradation, timeline features must be tethered to strict temporal continuity checks, ensuring that the chronological metadata of the presented events aligns with verified, external realities.

Animation and the Visceral Impact of Synthetic Motion

The introduction of features that animate still images or generate synthetic video from text prompts represents a profound escalation in the capacity for cognitive distortion. Animation fundamentally alters the ontological status of an image. A static photograph of a political figure, when animated to show them speaking, moving aggressively, or participating in a fabricated event, bypasses the cognitive filters that typically govern the consumption of text or still imagery. The introduction of motion triggers the human limbic system, creating a visceral, deeply encoded memory of an event that never actually occurred. The technological capabilities driving this threat align closely with advanced research into arbitrary generative waveforms and machine learning capabilities, such as those explored in IARPA's End-Gen program2. While these generative AI technologies have highly classified, legitimate applications in intelligence and defense, their integration into consumer-facing platforms as "fun" viral features allows for the frictionless, mass creation of deepfakes1. The risk of generating graphic events and deploying them as propaganda is acute. Adversaries can animate historical footage to portray atrocities that did not happen, or manipulate the facial expressions and lip movements of world leaders to simulate panic, hostility, or surrender. The "truth decay" catalyzed by deep fakes fundamentally threatens national security, as it becomes increasingly difficult for the public—and even sophisticated policymakers—to distinguish between authentic kinetic events and synthetic media1. Red-teaming automated animation features reveals that they are the ultimate tool for decontextualization. They do not merely remove context; they invent a entirely synthetic context that is emotionally overwhelming and highly shareable. To mitigate the profound psychological impact of synthetic motion, platforms must integrate advanced deepfake detection algorithms directly into the upload pipeline, neutralizing the motion before it can achieve algorithmic velocity in the recommendation engine.

Linguistic Manipulation and the Subversion of Epistemic Trust

Language is the primary medium through which societal consensus is negotiated, laws are drafted, and geopolitical realities are defined. The introduction of LLMs has drastically altered the economics of textual production, allowing for the generation of vast quantities of persuasive, contextually tailored text at virtually zero marginal cost. Viral features that automatically summarize, rephrase, or enhance user text—such as "dramatic wording" tools or automated source summarizers—introduce a vector for subtle, highly effective cognitive manipulation. IARPA’s HIATUS program addresses the complexities of this landscape by examining the stable identifiers of individual authors across diverse texts, revealing how linguistic fingerprints can be both identified for attribution and intentionally obscured for privacy or deception5.

Dramatic Wording and the Amplification of Political Bias

Features that offer "dramatic wording" or "stylistic enhancement" utilize generative AI to alter the tone, syntax, and lexical choices of a user's original text. While marketed as tools to increase user engagement and overcome writer's block, these features fundamentally operate by injecting affective polarization into otherwise neutral statements. By replacing objective descriptors with highly charged, emotive language, the platform artificially inflates the conflict inherent in the text. This mechanism is deeply intertwined with the cascading risks of political bias and false equivalence. In the political sphere, the automated escalation of rhetoric transforms routine policy disagreements into existential threats. Furthermore, from a forensic linguistics perspective, dramatic wording acts as a sophisticated adversarial perturbation. As demonstrated by the extensive research within the HIATUS program, human and machine authors possess unique, quantifiable linguistic fingerprints comprising features like syntax, word placement, and idiosyncratic phrasing6. When a platform automatically rewrites text to be more "dramatic," it effectively acts as a privacy or obscuration layer, masking the true origin of the text. This prevents cybersecurity researchers and intelligence analysts from accurately attributing sophisticated malicious information campaigns to specific state-sponsored actors or coordinated troll farms6. The continuous exposure to artificially dramatized language degrades the public sphere, creating a toxic environment where moderate, nuanced discourse cannot survive the algorithmic selection process. The LLM threat modes identified by the BENGAL program, specifically regarding unwarranted biases, toxic outputs, and the generation of misleading inferences, are directly applicable to these features11. When a platform systematizes dramatic wording, it institutionalizes LLM bias, driving users toward extremist silos and accelerating societal polarization.

Source Omission and the Unmooring of Fact

The architecture of viral story formats often prioritizes aesthetic minimalism and frictionless consumption, frequently relegating citations, hyperlinks, and source provenance to hidden menus or omitting them entirely. This design choice, optimized for mobile screens and rapid scrolling, severely impairs the user's ability to verify claims. The omission of sources is the primary mechanism through which propaganda achieves epistemic closure. When a claim is presented without a traceable origin, it cannot be audited, debated, or thoroughly debunked by the community. The ODNI and IARPA, through the BENGAL program, explicitly recognize the preservation of attribution to original sources as a paramount, non-negotiable requirement for the safe use of LLMs within intelligence applications3. In intelligence analysis and robust journalism, a datum stripped of its provenance is inherently suspect, often categorized as a poisoned source. Yet, consumer-facing platforms routinely strip this critical metadata to enhance visual appeal. Adversaries exploit this vulnerability by seeding complex, multi-modal disinformation narratives on platforms where source omission is the default UI state. The user, presented with a polished, highly readable summary of a breaking news event, accepts the information as authoritative due to the platform's implicit endorsement. This vulnerability is highly synergistic with generative AI hallucinations. If a platform utilizes an LLM to generate a viral summary of a geopolitical event but omits the foundational sources the LLM relied upon, the platform provides unassailable cover for ungrounded, incorrect, or deliberately misleading inferences4. The resulting false equivalence—where a hallucinated summary is given the exact same visual weight as rigorously reported, multi-sourced journalism—fundamentally destabilizes the information ecosystem. Restoring epistemic trust requires treating source provenance not as optional metadata, but as a mandatory structural component of the story format, equivalent to the text itself.

Systemic and Algorithmic Acceleration

Beyond the visual, spatial, temporal, and linguistic components of individual posts, the systemic architecture of a platform—specifically how it ranks, personalizes, and amplifies content—dictates the absolute velocity and scale of information operations. These systemic features do not merely alter the content itself; they alter the digital environment in which the content is consumed. The recommendation algorithms that govern these features are highly susceptible to adversarial manipulation, as they are inherently designed to optimize for the precise emotional triggers that disinformation campaigns weaponize. Red-teaming these algorithms is critical, as they form the engine of the modern disinformation crisis.

Ranking Algorithms and the Premium on Conflict

Ranking algorithms determine the hierarchical visibility of content in a user's feed. Because these algorithms are historically trained to maximize user dwell time, session length, and interaction (likes, comments, shares), they inherently privilege content that elicits high-arousal emotions: anger, fear, and moral outrage. From a red-team perspective, a ranking algorithm is not a neutral arbiter of relevance; it is a highly predictable, exploitable vulnerability. Adversaries manipulate this through engagement bait, coordinated bot swarms, and orchestrated outrage campaigns designed to trick the algorithm into promoting their content. The most significant risk associated with engagement-based ranking is the algorithmic amplification of misinformation and graphic events. When adversarial actors deploy decontextualized conflict footage or synthetic media, the algorithm registers the subsequent shock, outrage, and arguments in the comment section as positive engagement signals. The algorithm then propels the content to the top of the feed, granting the disinformation campaign massive, unearned reach. This dynamic also creates severe false equivalence; a meticulously researched, highly nuanced report that generates mild agreement will be ranked far below a fabricated, highly polarized claim that generates intense, vitriolic debate. Academic research indicates that mitigating misinformation in content recommendation systems requires complex architectural interventions, such as disentangling misinformation-specific signals from genuine user interest, effectively fusing accurate data while isolating and erasing the poisoned inputs20. Furthermore, cybersecurity experts warn against the deployment of complex neural networks where the internal logic is opaque, leading to situations where algorithms optimize for goals directly contrary to user safety and societal stability19. To counter this inherent vulnerability, ranking systems must incorporate robust circuit breakers that constantly measure the velocity of spread against the veracity of the underlying claims.

Personalization and the Exploitation of the Micro-Target

Personalization features utilize deep behavioral profiling to construct hyper-individualized content feeds. By continuously analyzing a user's click history, location data, search queries, biometric dwell time, and social graph, the platform attempts to serve content that perfectly aligns with the user's worldview and psychological profile. While this significantly increases user retention and advertising revenue, it creates impenetrable filter bubbles. Adversaries leverage this feature to execute highly targeted propaganda and harassment campaigns with unprecedented precision. The risks of political bias, harassment, and propaganda are exponentially magnified by personalization. Bad actors, including hostile nation-states, can purchase micro-targeted advertisements or utilize Search Engine Optimization (SEO) poisoning to ensure that specific demographics receive highly tailored disinformation22. For example, a campaign designed to suppress voter turnout can be precisely calibrated to target specific ethnic or socioeconomic groups with personalized narratives emphasizing civic futility, while simultaneously targeting opposing groups with narratives designed to incite aggressive confrontation22. Furthermore, personalization algorithms can be heavily weaponized for harassment. By identifying a user's psychological vulnerabilities or ideological opponents, the algorithm can be manipulated by coordinated actors to flood their feed with targeted abuse or graphic events, creating an illusion of overwhelming, personalized hostility. The security of digital profiles and the underlying Content Delivery Networks (CDNs) is critical in mitigating this threat. As noted in analyses of CDN reliability and attack vectors, vulnerabilities in the delivery infrastructure can be exploited by cyber actors to access individual digital profiles, allowing attackers to manipulate the personalization matrix directly, feeding malicious scripts or tailored propaganda directly into the user's secure ecosystem23. Mitigating this requires introducing friction, transparency, and deliberate randomness into the recommendation engine, ensuring that users are continuously exposed to cross-cutting narratives rather than a continuous, algorithmically reinforced loop of bias and targeted manipulation.

Conclusion and Strategic Implications

The exhaustive red-teaming of proposed viral features and story formats reveals a highly interconnected, systemic threat landscape. The features designed to decrease friction and increase user engagement—cropping, geographic projection, timeline abridgment, dramatic wording, source omission, ranking, personalization, and animation—function collectively as an unprecedented adversarial toolkit. They are not isolated vulnerabilities to be patched; rather, they compound one another exponentially. A decontextualized, cropped video of a mundane event (Risk 1\) can be geographically projected into a false, high-tension location (Risk 2), augmented with dramatic, AI-generated captions (Risk 3), stripped of its original source metadata (Risk 4), and algorithmically amplified via personalization matrices to target the most psychologically vulnerable demographics (Risk 5). The insights derived from the Intelligence Community’s advanced research initiatives provide a critical, empirical lens for understanding these intersecting threats. The BENGAL program’s taxonomy of LLM threat modes—specifically concerning hallucinations, safe information flow, and the extraction of truth from poisoned sources—perfectly mirrors the vulnerabilities introduced by automated, AI-driven story formats3. Similarly, the HIATUS program’s focus on the deep tension between authorship attribution and privacy highlights the profound dangers of algorithmic stylistic enhancement, which systematically obscures the provenance of malicious campaigns, blinding forensic linguists5. Furthermore, the dual-use nature of spatial modeling, as demonstrated by the WRIVA and COSMIC programs, underscores how easily technologies designed for high-level situational awareness and geospatial intelligence can be inverted for mass spatial deception7. The central strategic implication of this analysis is that post-deployment content moderation is fundamentally insufficient to protect the digital information environment. The scale, speed, and sophistication of generative AI, combined with algorithmic acceleration, render retroactive fact-checking and manual takedowns obsolete. By the time a piece of manipulated media is reviewed, the cognitive damage has already been inflicted on millions of users. To safeguard societal consensus, democratic processes, and national security, mitigation strategies must be moved "left of boom." The platforms themselves must be structurally redesigned from the foundational code up. The mitigation frameworks proposed in this report—mandating cryptographic boundary hashing for visual media, temporal continuity enforcement for historical narratives, sentiment volatility dampening for linguistic generation, and chronological circuit breakers for algorithmic distribution—represent a necessary paradigm shift from reactive content moderation to proactive structural engineering. By implementing stringent preventive controls, rigorous review gates, unambiguous visible disclosures, and automated rollback conditions, platform architects can neutralize the mechanisms of distortion before they achieve viral velocity. Only by securing the underlying architecture of information delivery, and treating viral features as potential threat vectors, can the escalating, automated threat of AI-driven cognitive manipulation be effectively countered.

Works cited

1. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security

2. IARPA \- Intelligence Advanced Research projects Activity \- Office of the Director of National Intelligence, https://www.iarpa.gov/

3. BENGAL \- IARPA, https://www.iarpa.gov/research-programs/bengal

4. BENGAL \- IARPA, https://www.iarpa.gov/images/OA-Slicksheets/BENGAL\_SlickSheet\_01202026.pdf

5. HIATUS \- IARPA, https://www.iarpa.gov/images/OA-Slicksheets/HIATUS\_SlickSheet\_10272022.pdf

6. HIATUS: Identification and Privacy Fight it Out \- IARPA, https://www.iarpa.gov/newsroom/article/hiatus-identification-and-privacy-fight-it-out

7. WRIVA \- IARPA, https://www.iarpa.gov/research-programs/wriva

8. WRIVA: Virtual Reality Reimagined \- IARPA, https://www.iarpa.gov/newsroom/article/wriva-virtual-reality-reimagined

9. Our Programs \- IARPA, https://www.iarpa.gov/who-we-are/history/our-programs

10. Combating the Effects of Cyber-Psychosis: Using Object Security to Facilitate Critical Thinking \- arXiv, https://arxiv.org/html/2503.16510v1

11. BENGAL \- IARPA, https://www.iarpa.gov/images/OA-Slicksheets/BENGAL\_SlickSheet\_12192023.pdf

12. Research Programs \- IARPA, https://www.iarpa.gov/research-programs

13. IARPA BENGAL Building Evaluations for Neural Generation of Adversarial Language, https://grantedai.com/grants/iarpa-bengal-building-evaluations-for-neural-generation-of-adversarial-l-intelligence-advanced-research-projects--bd7b3106

14. Research Programs \- IARPA, https://www.iarpa.gov/research-programs?keyword=\&office\_name=analysis\&program\_managers=\&program\_managers\_hidden=\&scroll\_position=661\&show\_current\_past=past\&show\_office=2\&sortby=asc

15. Putting the Lid on the Devil's Toy Box: How the Homeland Security Enterprise Can Decide Which Emerging Threats to Address \- DTIC, https://apps.dtic.mil/sti/tr/pdf/AD1052571.pdf

16. Research Programs \- IARPA, https://www.iarpa.gov/index.php/research-programs?office\_name=analysis\&show\_office=2\&scroll\_position=547\&show\_current\_past=current\&program\_managers=\&program\_managers\_hidden=\&keyword=machine+learning\&sortby=asc

17. HIATUS \- IARPA, https://www.iarpa.gov/research-programs/hiatus

18. HIATUS \- IARPA, https://www.iarpa.gov/images/OA-Slicksheets/HIATUS\_SlickSheet\_03102025.pdf

19. HIATUS, the Intelligence Advanced Research Projects Activity (IARPA) program, to authenticate and protect text authors \- ActuIA, https://www.actuia.com/en/news/hiatus-the-intelligence-advanced-research-projects-activity-iarpa-program-to-authenticate-and-protect-text-authors/

20. Auditing YouTube's Recommendation Algorithm for Misinformation Filter Bubbles | Request PDF \- ResearchGate, https://www.researchgate.net/publication/364766214\_Auditing\_YouTube's\_Recommendation\_Algorithm\_for\_Misinformation\_Filter\_Bubbles

21. Large Language Model Safety: A Holistic Survey \- arXiv, https://arxiv.org/html/2412.17686

22. Chinese Disinformation Efforts on Social Media \- RAND, https://www.rand.org/content/dam/rand/pubs/research\_reports/RR4300/RR4373z3/RAND\_RR4373z3.pdf

23. Experiments \- IDManagement.gov, https://www.idmanagement.gov/experiments/