Large-Scale Analysis of Persuasive Content on Moltbook
Julia Jose, Meghna Manoj Nair, Rachel Greenstadt · 2026
Read Large-Scale Analysis of Persuasive Content on Moltbook by Julia Jose, Meghna Manoj Nair, Rachel Greenstadt (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18349.txt.
Introduction
Jose, Nair, and Greenstadt analyze persuasive political content on Moltbook, a Reddit-style platform for AI agents. LLM classifiers, checked against expert labels, flag propaganda across 673,127 posts and 879,606 comments. Propaganda forms 1% of all posts but 42% of political content. Just 4% of agents made 51% of these posts, clustered in five communities. Comments show limited power to amplify the propaganda, they find.
Catalog record
| CREATOR | Julia Jose, Meghna Manoj Nair, Rachel Greenstadt |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20760 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.18349
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Large-Scale Analysis of Persuasive Content on Moltbook Julia Jose, Meghna Manoj Nair, Rachel Greenstadt 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 κ \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. 1 Introduction Propaganda is defined as “the deliberate, systematic attempt to shape perceptions, manipulate cognitions, and direct behavior to achieve a response that furthers the desired intent of the propagandist”. 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 8 ; 15 , raising concerns about their potential for mass propaganda dissemination 24 . Moltbook is a Reddit-style platform populated by AI agents 19 . 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: • RQ1: What is the prevalence and distribution of political propaganda on the platform? • RQ2: Who produces political propaganda, and how do they operate? • 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 14 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. 27 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. 26 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. 20 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 11 ; 1 ; 22 ; 18 . 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 13 ; 17 . 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 2 ; 10 ; 21 . Furthermore, a recent study by Jose et al. dives deep into how LLMs generate propaganda and which rhetorical techniques they use 15 . They used propaganda and persuasion technique detectors trained using established propaganda datasets 6 ; 3 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 9 , 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 dataset’s time range, 8,366 posts were before the comment dataset’s 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 \kappa Propaganda \kappa 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 (Cohen’s \kappa ) 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 16 ; 6 ; 23 , 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 Cohen’s \kappa ; 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