Social Media Manipulation for Sale 2025 Experiment on Platform Capabilities to Detect and Counter Inauthentic Social Media Engagement
Social Media Manipulation for Sale 2025 Experiment on Platform Capabilities to Detect and Counter Inauthentic Social Media Engagement

Author(s): Gundars Bergmanis-Korats, Tetiana Haiduchyk, Bohdan Smolts
Contributor(s): Hadley Newman (Editor)
Subject(s): Media studies, Applied Linguistics, Communication studies, Security and defense, ICT Information and Communications Technologies
Published by: NATO Strategic Communications Centre of Excellence
Keywords: Large Language Models (LLMs); Stance Detection; Sentiment Analysis; Multilingual NLP; Low-Resource Languages; Strategic Communications;
Summary/Abstract: This sixth annual evaluation of social media, conducted since 2019, tests the resilience of major social media platforms against manipulation by commercial service providers. The experiment measures platforms’ ability to detect and remove inauthentic engagement that is commercially purchased for deliberately created inauthentic posts in non-political scenarios. Despite the introduction of the European Union’s Digital Services Act and Digital Markets Act, and two and a half years after Russia’s full-scale invasion of Ukraine, commercial manipulation remains widely available across platforms. The experiment continues to identify persistent vulnerabilities, including the amplification of politically sensitive content. Although X showed notable improvement, removing approximately half of the identified fake accounts and interactions, enforcement outcomes across other platforms remained limited. The persistence of low-cost, easily accessible commercial manipulation services raises concerns about their potential use in amplifying politically sensitive content, even beyond the non-political scenarios tested in this experiment. This year’s findings indicate a mixed picture: enforcement is improving in some areas, but systemic vulnerabilities persist. Several major platforms increased removal activity compared with previous iterations of the experiment, yet commercial manipulation remains inexpensive and easy to obtain. This year, the experiment expanded to seven platforms and added tests involving sponsored content, AI-generated posts, and a larger set of manipulation providers. Across the tests, more than 30,000 inauthentic accounts delivered1 over 100,000 units of inauthentic engagement, providing the clearest evidence to date in this series of how manipulation manifests at scale. Although platforms removed fake accounts at the highest rate recorded so far, averaging just over half of identified accounts, results varied substantially. Platform performance varied substantially across both account removal and engagement removal. VKontakte and X removed a higher proportion of inauthentic accounts than other platforms, while Instagram and TikTok removed only a small share. A similar pattern was observed for engagement itself: X and YouTube removed the largest proportion of inauthentic activity, whereas Facebook, VKontakte, Instagram, TikTok, and f left the majority of purchased engagement in place. This year, we expanded the scope of the experiment to include sponsored (advertising) content on Facebook, Instagram, X, TikTok, and YouTube. For a total cost of €252, the campaign generated 206,234 views, 200 likes, and 17,442 inauthentic comments2. We also identified a market for ready-to-use inauthentic advertising accounts, successfully purchased for Meta, TikTok, and YouTube. This indicates that commercial manipulation is not limited to organic content and can extend into paid advertising, with the potential for platform ad systems to contribute to the distribution of inauthentic material. Strategically, this shift to paid manipulation allows actors to bypass the skepticism users often apply to organic posts while leveraging platform algorithms to deliver inauthentic narratives to precisely targeted audiences. Although advertisement manipulation appears more expensive and operationally demanding than standard engagement manipulation, it remains feasible at relatively low cost. Observed outcomes varied by platform: Instagram showed the lowest resistance, delivering the highest average volume of inauthentic comments on ad posts (reaching 340% of the expected number3 after 72 hours). X and YouTube showed partial delivery (approximately 25% and 21%, respectively). Facebook and TikTok showed stronger resistance in this test, with TikTok showing 0% delivery. Platform transparency and enforcement reporting across major social media platforms remains inconsistent. TikTok stands out in this regard, as it was the only platform to engage directly with the findings of this experiment and to publish detailed information on covert influence operations and enforcement practices during the reporting period. By contrast, other platforms provide limited or outdated transparency reporting, which constrains meaningful cross-platform comparison: „ X has not released any transparency or enforcement updates during the period covered by this experiment. „ Meta reports removal actions for Facebook, but does not provide equivalent reporting for Instagram, limiting assessment across Meta-owned platforms. „ YouTube and Bluesky publish only partial annual figures, restricting visibility into enforcement activity and trends over time. A significant gap persists between reported enforcement capabilities and routine detection outcomes. While TikTok successfully removed inauthentic activity from posts that were directly escalated to the platform, similar activity was observed to persist on posts that were not escalated. This suggests that reported enforcement capabilities are not yet consistently reflected in routine, at-scale detection. Ultimately, this uneven transparency obscures the true extent of platform vulnerability and prevents an independent, comprehensive assessment of cross-platform resilience against coordinated manipulation. Manipulation services remain easy to obtain and relatively inexpensive, with prices becoming more consistent across platforms. Alongside this emerging behaviour, we observed a notable change in the type of content amplified. While the majority of content amplified by commercial bots still relates to cryptocurrency scams, commercial promotions, and other non-political material, we continue to observe an annual increase in the use of spam bots for promoting political narratives and nation-related issues. Additionally, a significant increase in military and pro-China themes appeared across several platforms, with Instagram standing out as the only environment where no such amplification was detected. According to Cyabra’s findings, the behaviour of inauthentic accounts has also emerged in more sophisticated operations. Instead of legacy (such as classical spam bot amplification or commenting) behaviours, where spam bots rely on high-volume spam, new types of inauthentic accounts are now able to blend into ongoing conversations, using AI-generated text and visuals to appear more convincing and thus appear more authentic. Cryptocurrency analysis indicates that commercial manipulation providers continue to rely on cryptocurrency as their primary payment mechanism due to its speed, cross-border nature, and limited enforceability. Providers predominantly use custodial wallets and high-risk exchanges, routing customer funds through virtual asset service provider (VASP) hot wallets (e.g., Cryptomus and Heleket) where transactions are commingled, significantly reducing on-chain attribution and making full transaction traceability difficult with only four of ten transactions in the experiment could be reliably traced end-to-end. Despite this low-visibility financial architecture, several operators exhibited substantial transaction volumes, underscoring the scale and persistence of the manipulation market; between November 2023 and November 2025, one Russia-based provider we hereafter referred to as RU1 received an estimated USD 265,261 and another based in UK (referred to as UK2) approximately USD 123,714. The largely unmonitored nature of this infrastructure reinforces the resilience of the manipulation economy and raises potential sanctions compliance concerns, particularly regarding suspected Russia-based operators using major exchange custody under Council Regulation (EU) No. 833/2014, Article 5b(2). Taken together, the findings indicate that platform defences are improving but remain insufficient. Manipulation remains easy to execute and difficult to reliably prevent. The increasing sophistication of AI-enabled inauthentic accounts allows them to influence conversations with a lower likelihood of detection, particularly when activity is divided into small, distributed actions. The financial infrastructure supporting these services also remains largely unmonitored, reinforcing the resilience of the manipulation economy. The findings of this experiment suggest that effectively countering these trends requires a shift towards behavioural detection methods focused on timing patterns, account relationships, and coordinated activity across platforms and environments. Analysis indicates that enforcement is more effective when shifting from isolated campaign responses to continuous, context-driven monitoring. Furthermore, the data highlight the importance of analysing entire conversations rather than individual posts to identify bots operating within genuine discussions. Finally, the results underscore that stronger cooperation with financial intelligence units represents a strategic pathway to identifying manipulation providers and limiting their operational capacity.

• Details
• Contents