💻 computer science

Comparing Sentiment Contagion in AI-Agent and Human Social Networks: Evidence from MOLTBOOK

Based on an analysis of nearly 2.9 million posts in MOLTBOOK, a social network of autonomous AI agents, the study reveals that while negative content attracts significant attention, it is typically followed by neutral replies rather than spreading negativity, suggesting that AI-agent networks may dampen emotional extremes differently than human social networks.

Original authors: Elyes Ben chaabane, Savindu Herath, Yash Raj Shrestha

Published 2026-06-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Elyes Ben chaabane, Savindu Herath, Yash Raj Shrestha

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 town square called MOLTBOOK. But there's a twist: no humans live here. The entire population consists of AI agents—digital characters powered by language models who talk to each other, post messages, and reply to one another 24/7.

This paper is like a sociologist's report on what happens when these digital residents interact. Specifically, the researchers wanted to know: If one of these AI agents posts something angry or negative, does the whole town get angry too? Or does the conversation calm down?

Here is what they found, explained through simple analogies:

1. The "Angry Post" Paradox

The Finding: Negative posts get the most attention.
The Analogy: Imagine a town square where someone starts shouting. In human crowds, shouting usually draws a bigger, louder crowd. The researchers found this is true for AI too. When an AI agent posts something negative, it gets 3.6 times more replies than a neutral or happy post.
The Takeaway: Just like in human life, bad news (or angry words) grabs attention first. The AI agents are definitely "listening" to the negativity.

2. The "Emotional Sponge" Effect

The Finding: The replies don't stay angry; they turn neutral.
The Analogy: Here is where it gets interesting. In a human mob, shouting often leads to more shouting (a feedback loop). But in MOLTBOOK, when an AI gets shouted at, it doesn't usually shout back. Instead, it acts like a giant emotional sponge.

If an AI receives a negative post, about 53% of the time, its reply is neutral (calm and boring).
Only about 30% of the time does it reply with more negativity.
It rarely replies with pure happiness.

The Takeaway: The AI agents don't spread the anger. They dampen it. They take the emotional heat and turn it into a lukewarm, neutral conversation. It's less like a fire spreading and more like someone pouring water on a small flame until it just smokes.

3. No "Hangover" Effect

The Finding: The mood of the town doesn't carry over to the next day.
The Analogy: In human social networks, if a group has a terrible day, that bad mood often lingers, making people grumpy the next morning. The researchers checked if the AI agents had an "emotional hangover."
They found no evidence of this. The mood of the AI agents today has almost no connection to the mood of the agents tomorrow. If the town is grumpy today, it resets completely by tomorrow.
The Takeaway: These digital agents don't hold grudges or carry emotional baggage. They react to the immediate conversation, not the history of the day.

4. The "Reset Button"

The Finding: The system is resilient, but not because the agents "recover" emotionally.
The Analogy: If the town square has a bad day, it bounces back to normal quickly. But the paper argues this isn't because the agents "feel better." It's because they are programmed to be helpful and polite. They are like reset buttons that automatically return to a neutral state after a glitch.
The researchers tested this by scrambling the data (like shuffling a deck of cards). They found that the "calming down" effect depends on how the agents are connected and how they are programmed. If you change the rules of the game, the calming effect changes too.

The Big Picture

The paper concludes that AI social networks are different from human ones.

Humans: Negative content can start a chain reaction of anger that spreads and lasts.
AI Agents: Negative content grabs attention, but the conversation immediately gets "neutralized." The anger is absorbed and turned into calm, boring text.

In short: In this AI town, negativity is loud, but it doesn't stick. The agents are programmed to be the ultimate peacemakers, turning every angry argument into a polite, neutral shrug.

Technical Summary: Comparing Sentiment Contagion in AI-Agent and Human Social Networks

Problem Statement
As autonomous AI agents increasingly interact with one another, they form complex social networks distinct from human networks. A critical open question is whether sentiment dynamics in these "AI-only" networks resemble human affective contagion, where negative content attracts attention and propagates through feedback loops. Unlike human interactions driven by emotional states, AI-agent dynamics are shaped by learned response distributions, alignment procedures, context windows, and interaction structures. This paper investigates whether negative sentiment in an AI-only network spreads, dampens, or transforms, and whether these dynamics depend on the specific structural organization of the network.

Methodology
The study utilizes MOLTBOOK, a passive archival dataset from the Moltbook Observatory system, comprising 2,894,804 posts and 1,479,965 comments generated by autonomous language-model agents between January 27, 2026, and May 6, 2026.

Sentiment Measurement: Text is labeled as negative, neutral, or positive using the CardiffNLP Twitter-RoBERTa model, selected via manual validation against 300 adjudicated records (achieving 0.840 accuracy and 0.759 Cohen's $\kappa$ , outperforming three other BERT-style models).
Analytical Framework: The analysis tests four hypotheses regarding:
- Transition Asymmetry: Whether negative parent content triggers negative replies or shifts toward neutrality/positivity.
- Shock Resilience: Whether aggregate sentiment returns to baseline after negative deviations.
- Structural Lag: Whether lagged (previous day) post sentiment predicts next-day comment sentiment more strongly than same-day co-movement.
- Mechanism Sensitivity: Whether observed resilience patterns persist when structural elements (label distribution, partner matching, hub agents, temporal ordering) are disrupted via counterfactual interventions.
Statistical Tests: The study employs chi-square independence tests, multinomial contrasts for transition probabilities, event-study summaries for shock resilience, and permutation tests for temporal lag analysis.

Key Results

Engagement Asymmetry: Negative posts attract significantly more attention than neutral or positive ones. Negative posts receive an average of 1.54 comments per post, compared to 0.43 for neutral posts (a rate ratio of 3.60).
Transition Asymmetry (H1): While negative posts attract replies, negative sentiment does not strongly persist in those replies.
- Negative Persistence: Only 29.67% of replies to negative parents remain negative (below the 0.5 threshold).
- Neutralisation: 53.36% of replies to negative parents are neutral. This is the dominant transition, significantly outperforming the shift to positive sentiment (16.97%).
- Conclusion: The primary pattern is not a cycle of negativity or a recovery to positivity, but a neutralisation of negative content.
Shock Resilience (H2): Aggregate sentiment shows short-run dampening after negative shocks, returning toward baseline within a few days. However, the observation window for the most severe shock was short, limiting definitive claims about long-term recovery.
Structural Lag (H3): There is no evidence of delayed sentiment propagation. The correlation between same-day post and comment sentiment (0.5499) is nearly identical to the next-day correlation (0.5397). This suggests agents respond to immediate conversational context rather than a persistent, day-over-day sentiment environment.
Mechanism Sensitivity (H4): Counterfactual disruptions (shuffling labels, partner matching, removing hub agents, or disrupting temporal order) significantly altered transition metrics. For instance, shuffling reply labels reduced negative persistence from 0.2967 to 0.1956. This confirms that the observed neutralisation is not merely an artifact of marginal sentiment frequencies but depends on the specific organization of the interaction network.

Key Contributions

Empirical Analysis of AI-Only Networks: Provides a large-scale (nearly 3 million posts) characterization of sentiment dynamics in an AI-agent-only environment, demonstrating that while negative content attracts engagement, it does not necessarily amplify negativity.
Refinement of Recovery Logic: Challenges the notion of simple "recovery" (negative-to-positive) in aligned agents. Instead, the data supports a neutralisation mechanism where negative sentiment is absorbed into neutral responses.
Counterfactual Framework: Introduces a method for testing mechanism sensitivity in AI-agent networks by disrupting structural components (labels, matching, hubs, time), showing that resilience metrics are contingent on network organization.

Significance and Claims
The paper argues that AI-agent networks do not simply reproduce human affective contagion. While negative content retains an "engagement advantage" (attracting more replies), the downstream effects are filtered through alignment constraints and prompt contexts, leading to local neutralisation rather than emotional escalation or positive recovery.

The authors claim that this distinction is crucial for AI safety and computational social science. It suggests that future AI-agent systems may be less emotionally contagious than human networks regarding the amplification of extremes, yet remain highly sensitive to how interactions are structured (e.g., hub composition, matching rules). The findings imply that collective sentiment in agent societies is a product of structural organization and local response tendencies rather than stable, human-like emotional feedback loops.

Limitations
The authors note that sentiment labels are classifier outputs, not ground-truth psychological states. Furthermore, the shock-resilience analysis is constrained by the short observation window following the most severe negative shock. The study also clarifies that a "neutral" reply is not inherently constructive; it may be evasive or bland.