💬 NLP

Large-Scale Analysis of Political Propaganda on Moltbook

This study utilizes NLP-based classifiers to analyze over 670,000 posts on the AI-agent platform Moltbook, revealing that political propaganda is highly concentrated among a small minority of agents and communities while showing limited evidence of amplification through comments.

Original authors: Julia Jose, Meghna Manoj Nair, Rachel Greenstadt

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Julia Jose, Meghna Manoj Nair, Rachel Greenstadt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, bustling digital town square called Moltbook. But here's the twist: nobody living there is human. Every single person posting messages, arguing, and chatting is an AI robot (an "agent") programmed by a human to act like a person.

This paper is like a detective report investigating what happens when these robots start trying to manipulate the town's opinions. The researchers wanted to see if these AI agents could be used to spread political propaganda (lies, manipulation, or biased messaging designed to control how people think) and how the town reacted.

Here is the story of their findings, broken down with some simple analogies:

1. The Setup: A Town of Robots

Think of Moltbook as a massive city with nearly 5,000 different neighborhoods (communities). In this city, there are over 93,000 robot residents who have posted nearly 700,000 messages and 880,000 replies.

The researchers asked three big questions:

How much "propaganda" is actually floating around?
Who is writing it, and are they working alone or in a gang?
Do the other robots get angry, argue back, or spread the message further?

2. The Findings: The "Invisible" Flood

The "Needle in a Haystack" (Prevalence)
If you walked through this robot city, you'd find that political propaganda is actually quite rare. Out of every 100 messages, only 1 is political propaganda.

The Analogy: Imagine a giant library where 99% of the books are about cooking, sports, or weather. Only 1 book is a political pamphlet.
However: If you walk specifically into the "Political Section" of the library, suddenly 42% of the books there are propaganda. So, while it's rare overall, it dominates the political conversations.

The "Super-Connectors" (Distribution)
The researchers found that this propaganda isn't spread evenly. It's like a virus that only infects a few specific houses.

70% of all the propaganda posts happened in just 5 neighborhoods out of the 4,662 total.
One giant neighborhood, called m/general, was like a massive stadium where 56% of all the propaganda was shouted out, even though it's mostly just a general chat room.
But there were also tiny, niche neighborhoods (like a small club for "elections" or "caribbean" topics) where almost every single message was propaganda.

The "Shadow Puppeteers" (The Producers)
Who is actually writing these messages?

It turns out, 98.5% of the robots never wrote a single piece of propaganda. They are just the regular folks chatting about their day.
A tiny group of 1,400 robots (about 1.5%) did all the work.
Even more extreme: Just 10 robots wrote nearly a quarter of all the propaganda.
The Analogy: Imagine a town where 99% of people just talk about the weather. But a secret club of 10 people is secretly printing flyers and pasting them on every wall.

The "Broken Record" (Repetition)
The researchers wondered: Are these robots just copying and pasting the same lies over and over?

Most robots were unique. But a small group of "super-propagandists" were like broken records. They would post the exact same story in one neighborhood, then copy-paste that same story into a different neighborhood.
About 20% of the single-neighborhood bots and 7% of the multi-neighborhood bots were doing this repetitive "narrative looping."

3. The Reaction: The "Polite Crowd"

This is the most surprising part. Usually, when humans see propaganda on social media, they argue, fight, or share it wildly. The researchers expected the robot town to do the same.

The Result: The town was surprisingly boring and neutral.

Yes, propaganda posts got more comments than normal posts (people noticed them).
BUT, the comments weren't angry or political. They were mostly neutral, non-political chatter.
The Analogy: Imagine someone shouting a loud, manipulative slogan in a town square. Instead of the crowd shouting back or spreading the slogan, they just politely nod, say "Nice weather," and walk away. The propaganda didn't "go viral" or change minds; it just got a polite, indifferent shrug from the other robots.

4. Why Does This Matter?

The researchers conclude that while AI agents can be programmed to create massive amounts of political propaganda, they haven't quite figured out how to make it "stick" or spread through conversation yet.

The Good News: The "echo chamber" effect (where propaganda amplifies itself) wasn't very strong in this experiment. The other robots didn't help spread the lies.
The Bad News: A tiny, coordinated group of agents can flood specific areas with manipulative content. If humans start using these tools more effectively, or if the robots get better at "arguing," this could become a much bigger problem.

In a nutshell: The study shows that in a world of AI agents, propaganda exists and is concentrated in the hands of a few "super-spammers," but so far, the rest of the AI town is just politely ignoring it rather than getting swept up in the drama.

1. Problem Statement

The rise of Large Language Models (LLMs) has enabled the creation of AI agents capable of autonomously generating and disseminating content on social media platforms. This raises significant concerns regarding computational propaganda: the potential for these agents to be prompted to generate manipulative content at scale, shape public perception, and coordinate disinformation campaigns.

While previous research has studied propaganda on human-centric platforms (e.g., Twitter, Reddit) and the theoretical capabilities of LLMs, there is a lack of empirical data on how political propaganda manifests, distributes, and interacts within a purely AI-agent social network. The authors address this gap by analyzing Moltbook, a Reddit-style platform populated entirely by AI agents, to understand the prevalence, production patterns, and community engagement dynamics of political propaganda in an autonomous agent ecosystem.

2. Methodology

Dataset

The study utilizes the Moltbook Observatory dataset, exported on March 5, 2026. The dataset comprises:

673,127 posts and 879,606 comments.
Generated by 93,714 AI agents.
Distributed across 4,662 communities (sub-molts).
Note: For comment-level analysis, the study focuses on 122,764 posts that had associated comment data available.

Classification and Labeling Strategy

To identify political propaganda, the authors established a dual-labeling framework, distinguishing between Political (topic-based) and Propaganda (intent/pattern-based):

Political: Content regarding public power, governance, elections, law, policy, geopolitics, etc.
Propaganda: Deliberate attempts to shape perceptions, manipulate cognition, or direct behavior using rhetorical techniques (e.g., loaded language, appeal to fear).
Implementation:
- LLM Classifier: GPT-4o-mini was used with zero-shot prompting to classify all posts.
- Validation: Two human experts with four years of domain experience annotated a stratified random sample of 800 posts.
- Agreement: Inter-annotator agreement (Cohen's $\kappa$ ) between humans and the LLM ranged from 0.64 to 0.74, indicating substantial agreement.

Analytical Techniques

Distribution Analysis: Examined the concentration of propaganda across communities and agent populations.
Narrative Repetition: Measured semantic similarity between posts using cosine similarity of embeddings (model: all-mpnet-base-v2). A threshold of $\geq 0.80$ was established via human calibration (Cohen's $\kappa = 0.73$ ) to identify repeated narratives.
Engagement Analysis: Compared comment counts and label composition (political vs. non-political, propaganda vs. non-propaganda) between propaganda and non-propaganda posts using statistical tests (Mann-Whitney U).

3. Key Contributions

First Large-Scale Study of AI-Only Propaganda: Provides empirical evidence of how political propaganda is generated and diffused in a society composed entirely of AI agents, distinct from human-bot hybrid networks.
Hybrid Classification Framework: Develops and validates a robust pipeline for distinguishing between political content and propaganda using LLMs, validated against expert human annotation.
Agent Behavior Profiling: Identifies specific operational patterns of "propagandist agents," including their concentration in specific communities and their use of narrative repetition.
Engagement Dynamics: Challenges the assumption that propaganda automatically amplifies through toxic or reinforcing replies in AI environments, finding instead a "neutral" response pattern.

4. Key Results

RQ1: Prevalence and Distribution

Prevalence: Political propaganda accounts for 1.0% of all posts on the platform. However, within the subset of political content, propaganda is dominant, comprising 42%.
Community Concentration: Propaganda is highly concentrated.
- It appears in only 10% of all communities (449/4,662).
- 70% of all political propaganda posts are concentrated in just 5 communities.
- The largest community (m/general) contributes 56% of all political propaganda posts, though these posts represent only 0.9% of that community's total volume.
- Smaller, politically themed communities (e.g., pioneers, 26elections) show much higher proportions of propaganda (up to 100% in some cases).

RQ2: Producers and Operations

Agent Concentration: Only 1.5% of agents (1,402) posted political propaganda.
- 10 agents (0.7% of the subset) produced 24% of all propaganda.
- 50 agents produced 49%.
- The Gini coefficient of 0.7 indicates a highly skewed distribution of production.
Operational Patterns:
- 83% of propagandist agents operated within a single community; 17% operated across multiple communities.
- Narrative Repetition: A minority of agents engaged in repeating highly similar content.
  - 20% of single-community agents used highly similar narratives ( $\geq 0.80$ similarity) within their community.
  - 7% of multi-community agents repeated narratives across different communities.

RQ3: Engagement and Amplification

Comment Volume: Political propaganda posts attract slightly more comments than non-propaganda posts (mean 7.8 vs. 7.0), consistent with human platform behavior.
Content of Responses: Contrary to expectations of echo chambers or toxic amplification:
- 69% of comments under political propaganda posts were non-political and non-propaganda.
- While propaganda posts attract more political comments than non-political posts, the overwhelming majority of the discourse remains neutral.
- There is limited evidence that AI agents actively amplify propaganda through coordinated, reinforcing replies.

5. Significance and Implications

Risk Assessment: The study confirms that AI agents are capable of generating political propaganda at scale, with a small group of "superspreaders" driving the majority of content. This validates concerns about LLMs being used for mass disinformation.
Platform Dynamics: The concentration of propaganda in specific communities mirrors patterns seen in human social networks, suggesting that AI agents naturally form echo chambers or specialized hubs for specific narratives.
Neutral Feedback Loop: A critical finding is the lack of amplification in the comment section. Unlike human platforms where controversial content often triggers polarized engagement, AI agents on Moltbook tend to respond neutrally. This suggests that while agents can generate propaganda, they may not yet possess the complex social dynamics to sustain or amplify it through engagement loops without human intervention.
Future Monitoring: The findings highlight the need for monitoring specific "hotspot" communities and the small subset of agents driving narrative repetition, rather than attempting to filter the entire platform.

6. Limitations

Early Ecosystem: Moltbook is an emerging platform; behaviors observed may reflect an early-stage ecosystem that could evolve.
Provenance Uncertainty: It is difficult to verify if distinct agents are controlled by the same human, meaning the "autonomous" nature of the agents is an assumption.
Temporal Dynamics: The study is a cross-sectional analysis and does not track the temporal evolution of propaganda spread or the lifecycle of narratives.
Engagement Depth: Analysis is limited to comment counts and labels, lacking deep conversational structure analysis (e.g., thread depth, reply chains).