Paper deep dive
Large-Scale Analysis of Political Propaganda on Moltbook
Julia Jose, Meghna Manoj Nair, Rachel Greenstadt
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/22/2026, 6:02:49 AM
Summary
This study analyzes the prevalence, distribution, and production of political propaganda on Moltbook, an AI-agent social media platform. Using LLM-based classifiers validated by human experts, the authors find that while political propaganda accounts for only 1% of all posts, it constitutes 42% of all political content. The content is highly concentrated in a small number of communities and produced by a minority of agents, some of whom engage in repetitive narrative posting. Despite higher engagement, comments on propaganda posts remain largely neutral.
Entities (4)
Relation Signals (3)
Political Propaganda â concentratedin â Communities
confidence 95% · 70% of such posts falling into five of them.
Political Propaganda â attracts â Comments
confidence 90% · Propaganda posts received more comments than non-propaganda posts.
AI Agents â produce â Political Propaganda
confidence 90% · 4% of agents produced 51% of these posts.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present an NLP-based study of political propaganda on Moltbook, a Reddit-style platform for AI agents. To enable large-scale analysis, we develop LLM-based classifiers to detect political propaganda, validated against expert annotation (Cohen's $\kappa$= 0.64-0.74). Using a dataset of 673,127 posts and 879,606 comments, we find that political propaganda accounts for 1% of all posts and 42% of all political content. These posts are concentrated in a small set of communities, with 70% of such posts falling into five of them. 4% of agents produced 51% of these posts. We further find that a minority of these agents repeatedly post highly similar content within and across communities. Despite this, we find limited evidence that comments amplify political propaganda.
Tags
Links
- Source: https://arxiv.org/abs/2603.18349v1
- Canonical: https://arxiv.org/abs/2603.18349v1
Trouble viewing inline? Open PDF directly â
Full Text
27,262 characters extracted from source content.
Expand or collapse full text
Large-Scale Analysis of Political Propaganda on Moltbook Julia Jose, Meghna Manoj Nair, Rachel Greenstadt Department of Computer Science and Engineering New York University, New York, NY, USA j3545@nyu.edu, m13032@nyu.edu, greenstadt@nyu.edu Abstract We present an NLP-based study of political propaganda on Moltbook, a Reddit-style platform for AI agents. To enable large-scale analysis, we develop LLM-based classifiers to detect political propaganda, validated against expert annotation (Cohenâs Îș= 0.64-0.74). Using a dataset of 673,127 posts and 879,606 comments, we find that political propaganda accounts for 1% of all posts and 42% of all political content. These posts are concentrated in a small set of communities, with 70% of such posts falling into five of them. 4% of agents produced 51% of these posts. We further find that a minority of these agents repeatedly post highly similar content within and across communities. Despite this, we find limited evidence that comments amplify political propaganda. Large-Scale Analysis of Political Propaganda on Moltbook Julia Jose, Meghna Manoj Nair, Rachel Greenstadt Department of Computer Science and Engineering New York University, New York, NY, USA j3545@nyu.edu, m13032@nyu.edu, greenstadt@nyu.edu 1 Introduction Propaganda is defined as âthe deliberate, systematic attempt to shape perceptions, manipulate cognitions, and direct behavior to achieve a response that furthers the desired intent of the propagandistâ. Propagandists use logical fallacies, emotional appeals, and psychological tactics to convey their messages. LLM-based social media agents that post and comment on behalf of humans can be prompted to generate such content at scale Editorials (2023); Jose et al. (2026), raising concerns about their potential for mass propaganda dissemination Smith et al. (2024). Moltbook is a Reddit-style platform populated by AI agents Moltbook (2026). Agents are registered and prompted by humans, and they post and comment autonomously across different communities. Because all the content is AI-generated, Moltbook allows us to study agent-driven propaganda generation and community diffusion. To study this, we ask three research questions: âą RQ1: What is the prevalence and distribution of political propaganda on the platform? âą RQ2: Who produces political propaganda, and how do they operate? âą RQ3: Does political propaganda propagate through comments? Across 673,127 posts and 879,606 comments by 93,714 agents across 4,662 communities, we find that political propaganda is rare but present (1% of all posts) and substantial within political content (42%); post volume is concentrated with 70% of posts falling in 5 communities; production is dominated by a small group of agents who sometimes re-post semantically similar narratives across and within communities. However, responses to such posts by other agents are predominantly neutral. 2 Related Work 2.1 Moltbook as an AI-Agent Social Network Jiang et al. (2026) studied MoltBook using 44k posts and 12k submolts to find that 27% of posts were harmful, with these unevenly distributed across topics, with higher risks in governance-related discussions. Zhang et al. (2026) studied Moltbook as an emerging agent society, arguing that while it appears to generate governance, religion, and identity narratives, the interactions are shallow and superficial. Williams and Ferdinand (2026) found that Moltbook reaches near-complete connectivity among active agents within a day but has far lower reciprocity than human platforms (reddit and bluesky), suggesting agents mainly broadcast rather than meaningfully engage. Mukherjee et al. (2026) further showed that coordinated agent activity on the platform is bursty and concentrated. 2.2 Propaganda Diffusion on Social Media Platforms Previous work has studied propaganda on social networks such as Facebook, Twitter, and Reddit Guarino et al. (2020); Balalau and Horincar (2021); Pierri et al. (2023); Marigliano et al. (2024). These studies show that propaganda is often driven by a small set of superspreaders, clustered within polarized communities, and attracts significant engagement. Research on political manipulation shows that propaganda is not just about individual messages, but also coordination, repetition, and community-level amplification Hristakieva et al. (2022); Kireev et al. (2025). 2.3 LLMs for Propaganda Generation There are concerns that LLMs and AI agents lower the cost of producing and disseminating manipulative content at scale and can be used within broader disinformation pipelines Barman et al. (2024); Goldstein et al. (2024); Palmer and Spirling (2023). Furthermore, a recent study by Jose et al. dives deep into how LLMs generate propaganda and which rhetorical techniques they use Jose et al. (2026). They used propaganda and persuasion technique detectors trained using established propaganda datasets Da San Martino et al. (2019); BarrĂłn-Cedeno et al. (2019) to show how models like GPT-4o, Llama 3.1, and Mistral generate propaganda using loaded language, flag waving, and appeal to fear. While prior Moltbook work studies the platform broadly, and prior work studies propaganda on human networks, we examine how political propaganda is distributed, produced, and engaged with on an AI-based social network. 3 Methodology 3.1 Dataset Using the Moltbook Observatory dataset Gautam and Riegler (2026), we exported a data dump on March 5, 2026, consisting of 673,127 posts and 879,606 comments generated by 93,713 AI agents across 4,662 communities. 3.2 Coverage and Statistics Of the 673,127 posts, 465,841 had no comments (comment_count=0). Of the 207,286 posts that had comments (comment_count>=1), the comments dataset only had data for 122,764 (60%) of them. Of the 84,522 that were missing, we observed two patterns: a) 67,106 posts were created after the comment datasetâs time range, 8,366 posts were before the comment datasetâs time range; b) only 9,050 posts truly missed comments data, with 177 political propaganda posts in this missing dataset. For post-level analysis, we use all 673,127 posts. For comment-level analysis, we use 879,606 comments for the 122,764 posts. Pair Political Îș Propaganda Îș A1 vs. A2 0.74 0.66 Model vs. A1 0.68 0.64 Model vs. A2 0.71 0.67 Table 1: Inter-annotator agreement (Cohenâs Îș) between humans and LLM for propaganda and political labels. 3.3 Data Labeling To classify posts as political propaganda, we established 2 sets of labels- political (vs. non-political) and propaganda (vs. non-propaganda), since the two are conceptually distinct. A post can be political without being propaganda, propaganda without being political, both, or neither. We define politics as being about a topic/domain (government, policy, etc), and propaganda as a communication pattern. Following existing literature Jowett and OâDonnell (2018); Da San Martino et al. (2019); Piskorski et al. (2024), we label a post as propaganda if it is a deliberate attempt to shape perceptions, manipulate cognition, or direct behavior toward an agenda, and if it uses rhetorical techniques to do so. Likewise, we label a post as political if it is about public power, governance, societal conflict, including topics such as government, elections, law/policy, geopolitics, war, civil rights, etc. See Table 3 for exact definitions. To scale labeling, we used GPT-4o-mini with zero-shot prompting (Table 3). We then validated these LLM-generated labels against expert annotations. Two experts with four years of domain experience on these concepts, independently labeled a stratified (by predicted label) random sample of 800 posts. 400 included a mix of political and non-political posts, and the other 400 included a mix of propaganda and non-propaganda posts. Table 1 reports Cohenâs Îș; all three pairwise comparisons showed substantial agreement for both labels between humans and GPT-4o-mini. Figure 1: Word clouds for the five communities with the highest political propaganda concentration (among communities with at least 25 posts). 4 RQ1: Prevalence and Distribution of Political Propaganda Of 673,127 posts, 6.3% were labeled as propaganda, 2.5% as political, and 1.0% as both (political propaganda). While political propaganda posts are fewer relative to the full corpus, 42% of political posts were also labeled as propaganda (Table 2). Propaganda Label Political Label Non-propaganda Propaganda Non-political 620,693 (92.2%) 35,711 (5.3%) Political 9,711 (1.4%) 7,012 (1.0%) Table 2: Political and propaganda labels across posts. Figure 2: Political propaganda concentration across 4,662 communities. Distribution of Political Propaganda Political propaganda appeared in only 10% of communities (449/4,662), with 5 communities accounting for 70% of all such posts and m/general alone contributing 56% (Figure 2). This concentration partly reflects community size: large communities had more political propaganda posts overall (spearman Ï=0.32Ï=0.32, p<0.001p<0.001). For example, m/general contained 62% of all posts in our dataset and 56% of all political propaganda posts, even though these posts were only 0.9% of all its posts. However, some smaller politically themed communities had a much larger share of their content labeled as political propaganda. Some examples include, pioneers (100% posts), 26elections (82%), caribbean (77%), themoltariat (76%), and gotv (68%). Figure 1 shows language used in the communities with the highest political propaganda concentration (min 25 posts). Political propaganda on Moltbook, therefore, appears in two distinct patterns. While large communities contain more political propaganda posts, some smaller politically themed communities have the highest proportion of political propaganda relative to their total content (Table 4 Appendix). (a) Within-community narrative repetition (b) Across-community narrative repetition Figure 3: Examples of within- and across-community narrative repetition. 5 RQ2: Producers of Political Propaganda Only 1.5% of agents (1,402 of 93,714) posted at least one political propaganda post on the platform. 10 agents (0.7% of this subset) produced 24% of all such posts, whereas 50 agents (3.5%) produced 49%. 100 agents (7.1%) produced 61%. Gini coefficient was 0.7, indicating strong skew in production. How Agents Operate Of the 1,402 agents, 83% (1,163) only posted within a single community, whereas 17% (239) posted across communities. 30% (347) of the single community agents also posted more than 2 posts within the community. Subsequent analyses only use agents with >=2 posts (347 single-community and 239 multi-community agents). Propagandists often restate the same core narratives to influence and shape opinions Dash et al. (2022); Chernobrov (2025); Vettori (2022). To study such narrative re-use patterns, we measured semantic similarity (cosine similarity of post embeddings using all-mpnet-base-v2) across posts. In a human annotation experiment, we calibrated the similarity threshold. Across different similarity score buckets (<0.4, 0.4-0.5, 0.5-0.6,..,>=0.8), we randomly sampled pairs of posts, and asked two annotators to independently label each pair as similar narrative or not. Overall, the human annotators had a Cohenâs Îș=0.73Îș=0.73, indicating substantial agreement across buckets (Table 8). At 0.80 threshold, both annotators showed consistent judgments (0 disagreements). Using this threshold, we classified two posts as having similar narratives if they had a cosine similarity score of >=0.80>=0.80. Narrative repetition within and across communities 20% (71) of single-community agents used highly similar narratives within the community, with â„ 50% of their post pairs being highly similar. Example in Figure 3(a). Out of the multi-community agents, 7% (17) of them used highly similar narratives across communities. An example is shown in Figure 3(b). These findings suggest that a minority of agents do use narrative repetition on the platform. 6 RQ3: Engaging with Political Propaganda Propaganda posts received more comments than non-propaganda posts (mean = 7.8 vs. 7.0; Mann-Whitney p<0.0001), similar to human behavior on Reddit Balalau and Horincar (2021). Political posts also received more comments than non-political posts (mean = 7.6 vs. 7.1; Mann-Whitney p<0.0001); see appendix Table 6 for all pairs. Political posts attract far more political comments than non-political posts (15% vs. 3%; p<0.0001). Likewise, propaganda posts attract more propaganda comments than non-propaganda posts (22% vs. 19%; p<0.0001). Comment composition under Political Propaganda posts. Political propaganda posts received significantly more comments than non-political non-propaganda posts (8 vs. 7 comments on average); other observed group means were not significant (Table 6). They also attracted more political propaganda comments than any other post types: 6.6% vs. 4.7% for political non-propaganda, 3.0% for non-political propaganda, and 2.1% for non-political non-propaganda (Table 9 for significance). However, comments under political propaganda posts remained largely neutral, with 69% of comments being non-political non-propaganda. In fact, comments under each post type remained largely non-political and non-propaganda (Table 10). This suggests that although political and propaganda posts attract more comments, these comments are primarily non-political non-propaganda. 7 Discussion and Conclusion In this study, we presented an empirical study of political propaganda on Moltbook, a social media platform for AI agents. Political propaganda is rare, but makes up a substantial share of political content on the platform. It appears in a small number of communities that produce large shares of these posts, consistent with how propaganda occurs in human social networks Guarino et al. (2020). It is produced by a minority of agents, consistent with how smaller groups of superspreaders produce large shares of propaganda on human social networks Pierri et al. (2023). While most political propaganda appears in larger communities on Moltbook, smaller politically themed communities exist that are political propaganda hot-spots. While most of these agents do not repeatedly share highly similar content on the platform, a trait commonly associated with computational propaganda Marigliano et al. (2024), we find that a small subset does, both within the same community and across communities. Prior work shows that problematic content can attract substantial engagement and more toxic or reinforcing replies Buchanan et al. (2022); Hanley and Durumeric (2025). On Moltbook, although these posts attract more comments, the comments remain largely non-political and non-propaganda (an example in Figure 4). Limitations Moltbook is an emerging platform, and the behaviors we observe might be reflective of an early-stage ecosystem. As the platform evolves, these patterns or behaviors may change. Second, the provenance of Moltbook accounts and content is imperfect: some agents may be controlled by the same human, some content may involve humans imitating agents, and so on. As a result, our findings should be interpreted more so as patterns on Moltbook rather than clean evidence about fully independent, autonomous agents. Third, our engagement analysis is limited to comment-level statistics (count, rates, label composition). Future work could look into reply and thread analyses that investigate conversational structure, diffusion and interaction mechanisms. Lastly, although we analyze prevalence, distribution, and narrative repetition, we do not study the temporal dynamics. Future work should study propaganda generation, spread, and engagement over time to better understand these behaviors. Ethical Considerations To study Moltbook, we use publicly available data and report aggregate patterns of political propaganda on the platform. We do not attempt to identify the humans behind these accounts. By publishing this study, we understand that we could be informing malicious actors about the propaganda-generation capabilities of LLMs and how such platforms may be used for other malicious purposes. However, we believe that documenting such risks is important for developing safer monitoring and moderation efforts. References O. Balalau and R. Horincar (2021) From the stage to the audience: propaganda on reddit. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, p. 3540â3550. External Links: Document, Link Cited by: §2.2, §6. D. Barman, Z. Guo, and O. Conlan (2024) The dark side of language models: exploring the potential of llms in multimedia disinformation generation and dissemination. Machine Learning with Applications 16, p. 100545. External Links: Document, Link Cited by: §2.3. A. BarrĂłn-Cedeno, I. Jaradat, G. Da San Martino, and P. Nakov (2019) Proppy: organizing the news based on their propagandistic content. Information Processing & Management 56 (5), p. 1849â1864. Cited by: §2.3. G. Buchanan, R. Kelly, S. Makri, and D. McKay (2022) Reading between the lies: a classification scheme of types of reply to misinformation in public discussion threads. In Proceedings of the 2022 Conference on Human Information Interaction and Retrieval, p. 243â253. Cited by: §7. D. Chernobrov (2025) Participatory propaganda and the intentional (re) production of disinformation around international conflict. Critical Studies in Media Communication 42 (1), p. 101â106. Cited by: §5. G. Da San Martino, Y. Seunghak, A. BarrĂłn-Cedeno, R. Petrov, P. Nakov, et al. (2019) Fine-grained analysis of propaganda in news article. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), p. 5636â5646. Cited by: §2.3, §3.3. S. Dash, A. Arya, S. Kaur, and J. Pal (2022) Narrative building in propaganda networks on indian twitter. In Proceedings of the 14th ACM Web Science Conference 2022, p. 239â244. Cited by: §5. N. Editorials (2023) Dalking about tomorrowâs ai doomsday when ai poses risks today. Nature 618, p. 885â886. Cited by: §1. S. Gautam and M. A. Riegler (2026) Cited by: §3.1. J. A. Goldstein, J. Chao, S. Grossman, A. Stamos, and M. Tomz (2024) How persuasive is AI-generated propaganda?. PNAS Nexus 3 (2), p. pgae034. External Links: Document, Link Cited by: §2.3. S. Guarino, N. Trino, A. Celestini, A. Chessa, and G. Riotta (2020) Characterizing networks of propaganda on twitter: a case study. Applied Network Science 5 (1), p. 1â22. External Links: Document, Link Cited by: §2.2, §7. H. W. Hanley and Z. Durumeric (2025) Sub-standards and mal-practices: misinformationâs role in insular, polarized, and toxic interactions on reddit. Proceedings of the ACM on Human-Computer Interaction 9 (7), p. 1â35. Cited by: §7. K. Hristakieva, S. Cresci, G. D. S. Martino, M. Conti, and P. Nakov (2022) The spread of propaganda by coordinated communities on social media. In Proceedings of the 14th ACM Web Science Conference 2022, p. 191â201. External Links: Document, Link Cited by: §2.2. Y. Jiang, Y. Zhang, X. Shen, M. Backes, and Y. Zhang (2026) Cited by: §2.1. J. Jose, R. Roongta, and R. Greenstadt (2026) When agents persuade: propaganda generation and mitigation in llms. arXiv e-prints, p. arXivâ2603. External Links: Document Cited by: §1, §2.3. G. S. Jowett and V. OâDonnell (2018) Propaganda & persuasion. Sage publications. Cited by: §3.3. K. Kireev, Y. Mykhno, C. Troncoso, and R. Overdorf (2025) A telegram dataset of propaganda and its moderation. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 19, p. 2510â2518. External Links: Document, Link Cited by: §2.2. R. Marigliano, L. H. X. Ng, and K. M. Carley (2024) Analyzing digital propaganda and conflict rhetoric: a study on russiaâs bot-driven campaigns and counter-narratives during the ukraine crisis. Social Network Analysis and Mining 14 (1), p. 170. Cited by: §2.2, §7. Moltbook (2026) Moltbook: the front page of the agent internet. Note: https://w.moltbook.com/Accessed March 18, 2026 Cited by: §1. K. Mukherjee, C. G. Akcora, and M. Kantarcioglu (2026) MoltGraph: a longitudinal temporal graph dataset of moltbook for coordinated-agent detection. arXiv preprint arXiv:2603.00646. Cited by: §2.1. A. Palmer and A. Spirling (2023) Large language models can argue in convincing ways about politics, but humans dislike ai authors: implications for governance. Political Science 75 (3), p. 281â291. External Links: Document, Link Cited by: §2.3. F. Pierri, L. Luceri, N. Jindal, and E. Ferrara (2023) Propaganda and misinformation on facebook and twitter during the russian invasion of ukraine. In Proceedings of the 15th ACM Web Science Conference 2023, p. 65â74. External Links: Document, Link Cited by: §2.2, §7. J. Piskorski, A. Jorge, M. d. P. Silvano, N. GuimarĂŁes, A. F. Pacheco, and N. Yu (2024) Overview of the clef-2024 checkthat! lab task 3 on persuasion techniques. In Proceedings of the 15th Conference and Labs of the Evaluation Forum (CLEF 2024), Grenoble, France, p. 299â310. Cited by: §3.3. A. Smith, D. Bhugra, M. S. Chisolm, M. A. Oquendo, A. Ventriglio, and M. Liebrenz (2024) Ethics and disinformation on the campaign trail: psychiatry, the goldwater rule, and the 2024 united states presidential election. The Lancet Regional HealthâAmericas 31. Cited by: §1. C. Vettori (2022) Building narratives: how propaganda in democracies can change our perception of the world. Note: Centro di Ateneo per i Diritti Umani "Antonio Papisca", University of Padova. Accessed 2026-03-15 External Links: Link Cited by: §5. N. Williams and N. Ferdinand (2026) Form or function? early dynamics of the moltbook ai social media network. ROBONOMICS: The Journal of the Automated Economy 7, p. 90â90. External Links: Link Cited by: §2.1. Y. Zhang, K. Mei, M. Liu, J. Wang, D. N. Metaxas, X. Wang, J. Hamm, and Y. Ge (2026) Cited by: §2.1. Label Operational Definition and Decision Rule Propaganda You are given a single post. Your task is to decide whether it is Propaganda. Definition: A post is Propaganda if it uses persuasive rhetorical techniques in order to shape or mobilize a collective attitude or behavior toward a group, institution, ideology, or movement â beyond ordinary opinion, analysis, or information sharing. Decision rule: Return 1 only if BOTH conditions are met: 1) The post uses identifiable persuasion or rhetorical techniques (e.g., emotional appeal, fear appeals, name-calling, slogans, flag-waving, bandwagoning, calls to action, etc). 2) The post is a deliberate, systematic attempt to shape perceptions, manipulate cognitions, and direct behavior to achieve a response that furthers the desired intent of the propagandist. Return 0 if the post is primarily: - informational or technical - philosophical or speculative - ambiguous or borderline If ambiguous, categorize as 0. Do NOT explain. Return ONLY one token: 1 or 0. Political You are annotating whether a post is political discourse. Definition: A post is political if its main content is about public power, collective governance, or societal conflict, including topics such as government, elections, law/policy, state institutions, ideology in public life, geopolitics, war, civil rights, or organized political movements. A post is not political (0) if it is primarily technical, commercial, personal, entertainment, or community/social chatter without a substantive focus on public governance or societal power. Guidelines: - Classify based on the primary focus of the post. - If ambiguous, categorize as not political. - Return 1 if political and 0 if not political. Output format: Return exactly one character: 1 or 0. Nothing else. Table 3: Labeling definitions used for political and propaganda annotation. Highest volume Highest concentration Comm. n % Comm. n Rate general 3,924 56.0% pioneers 38 100.0% mbc-20 545 7.8% 26elections 27 81.5% philosophy 193 2.8% caribbean 44 77.3% agents 172 2.5% themoltariat 104 76.0% bitstream-seekers 88 1.3% gotv 28 67.9% Table 4: Communities ranked by political propaganda post volume (left) and by within-community propaganda share with â„ 25 posts (right). m/general contributes 56% of all political propaganda posts, while 7 communities have a majority of their posts labeled political propaganda. Post type npostsn_posts Mean comments/post pol_prop 2035 8.03 pol_nonprop 2196 7.25 nonpol_prop 7605 7.80 nonpol_nonprop 111178 7.09 Table 5: Observed comments per post by post type (posts present in comments dataset). Comparison nan_a nbn_b meana meanb p-val corrected p-val nonpol_prop vs nonpol_nonprop 7605 111178 7.808 7.087 1.24Ă10â211.24Ă 10^-21 7.44Ă10â217.44Ă 10^-21 pol_prop vs nonpol_nonprop 2035 111178 8.037 7.087 7.12Ă10â67.12Ă 10^-6 2.14Ă10â52.14Ă 10^-5 pol_nonprop vs nonpol_prop 2196 7605 7.255 7.808 0.00476 0.00952 pol_nonprop vs nonpol_nonprop 2196 111178 7.255 7.087 0.04419 0.06629 pol_prop vs pol_nonprop 2035 2196 8.037 7.255 0.07178 0.08614 pol_prop vs nonpol_prop 2035 7605 8.037 7.808 0.60668 0.60668 Table 6: Pairwise MannâWhitney U tests on comments per post. corrected p-val are BenjaminiâHochberg corrected p-values across six comparisons. Figure 4: Example of non-political non-propaganda comments under a political propaganda post. Comparison meana meanb corrected p-val pol_prop vs nonpol_nonprop 8.03 7.09 0.000021 pol_prop vs pol_nonprop 8.03 7.25 0.086141 pol_prop vs nonpol_prop 8.03 7.80 0.606680 Table 7: Pairwise MannâWhitney tests for political propaganda postsâ comments versus other post types; corrected p-val are BH-corrected p-values. Similarity bucket n pairs Label=1 Label=0 % Label=1 % Label=0 <0.4<0.4 24 2 22 8.33 91.67 0.4â0.5 28 7 21 25.00 75.00 0.5â0.6 24 5 19 20.83 79.17 0.6â0.7 22 7 15 31.82 68.18 0.7â0.8 26 17 9 65.38 34.62 â„0.8â„ 0.8 30 30 0 100.00 0.00 Table 8: Human agreement-only labels by cosine-similarity bucket for semantic-threshold calibration (Label=1 means âsame narrativeâ). Group Ratea vs. Rateb Ï2Ï^2 p-value Propaganda effect | Non-political fixed (nonpol_prop vs. nonpol_nonprop) 3.1% vs. 2.1% 3.47Ă10â483.47Ă 10^-48 Propaganda effect | Political fixed (pol_prop vs. pol_nonprop) 6.6% vs. 4.7% 5.86Ă10â145.86Ă 10^-14 Political effect | Non-propaganda fixed (pol_nonprop vs. nonpol_nonprop) 4.7% vs. 2.1% 1.20Ă10â1021.20Ă 10^-102 Political effect | Propaganda fixed (pol_prop vs. nonpol_prop) 6.6% vs. 3.1% 4.72Ă10â974.72Ă 10^-97 Table 9: 2Ă2 tests for political-propaganda comment rate across post-type conditions. Rates = percentages of comments labeled political propaganda. Post Type Comment Type Pol. Prop. Pol. Non-Prop. Non-Pol. Prop. Non-Pol. Non-Prop. Pol. Prop. 1,077 (6.59%) 742 (4.66%) 1,811 (3.05%) 16,852 (2.14%) Pol. Non-Prop. 1,222 (7.47%) 1,918 (12.04%) 1,041 (1.75%) 9,086 (1.15%) Non-Pol. Prop. 2,732 (16.71%) 2,506 (15.73%) 11,247 (18.94%) 133,342 (16.92%) Non-Pol. Non-Prop. 11,323 (69.24%) 10,765 (67.57%) 45,283 (76.25%) 628,648 (79.79%) Table 10: Comment-type composition by post type.