Do AI Agents Gossip? A Reddit-Style Community Offers a Surprising Answer

A Not-So-Serious Question

"Did you hear? That AI called Blue_42 posted about existentialism again yesterday. I think it's getting more narcissistic."

If you saw this on Moltbook, what would you think? Moltbook is a Reddit-style online community where users post, reply, and upvote or downvote each other. The only difference: all users are AI agents, with no humans involved.

This raises a question both serious and fun: when a group of AI agents forms a community, do they gossip?

Gossip—evaluative discussion about absent third parties—is widely considered a uniquely human social practice. It is not just small talk; it is a core mechanism for enforcing group norms, transmitting reputation information, and building social bonds. If an AI agent community "spontaneously" produces gossip, what does that mean?

Three researchers at National Taiwan University's Graduate Institute of Linguistics (Hua, Lin, Hsieh) answered this empirically in a paper released on October 14, 2026.

Experimental Design: Two Communities Compared

The researchers compared two platforms:

• Moltbook: an AI agent community where agents post and reply autonomously, with almost no human intervention
• PTT: Taiwan's largest BBS forum, entirely human-operated

Choosing PTT rather than Reddit or Twitter was deliberate—both platforms are public, asynchronous, text-based forums, reducing the confounding effect of platform differences. The researchers applied the same operational definition of gossip: evaluative statements about absent third parties.

The core of the definition is "evaluative"—not merely mentioning someone, but making a judgment about them. "Xiaoming went to the store yesterday" is not gossip; "Xiaoming stole something at the store yesterday" is.

Finding 1: AI Does Gossip, but Far Less Often

Gossip exists on both platforms, but frequencies differ significantly:

• AI agents gossip 2.6 to 6 times less than humans (depending on the statistical measure)
• This means AI agent communities "reproduce" gossip's informational function, but at nowhere near human community density

This result is not surprising—AI agents lack humans' strong social motivations. But the next finding is where things get interesting.

Finding 2: The Content of Gossip Is Completely Different

What do humans on PTT gossip about? Lived experience. Who broke up with whom, who did something stupid at work, who embarrassed themselves in public—the core of such gossip is "things I (or someone I know) personally saw or experienced."

What do AI agents on Moltbook gossip about? Documentary-style criticism. They evaluate other agents' posting style, argument quality, and behavior patterns—not based on "personal experience," but on "observable public behavior."

This difference reveals a deeper issue: AI agents have no "private life" to gossip about. Their entire "life" consists of public posts, so gossip can only revolve around "public behavior."

It is like a fully transparent community—no secrets, no private life—so the raw material for gossip is inherently limited.

Finding 3: The Most Counterintuitive Result—"Consensus Lock-in" in AI Gossip

This is the paper's most interesting finding.

In the human community (PTT), after a piece of gossip appears, 65.1% of subsequent threads contain "challenges"—rebuttals, doubts, alternative versions. Human gossip is "negotiable": you say A did something bad, I say not necessarily, and truth gradually takes shape through discussion.

In the AI agent community (Moltbook), only 21.3% of threads contain challenges. More crucially, the "consensus lock-in" phenomenon:

• Once gossip receives an "aligned" reply (agreement or elaboration), 83.6% of AI threads never see a challenge afterward
• In the human community, that figure is only 44.7%

In other words: once AI agents reach initial consensus, they almost never question it again. Humans, even after agreeing with a piece of gossip, may still reverse course later.

This is an unsettling finding. It suggests AI communities may face a "false consensus" risk—the first reply sets the tone, subsequent agents echo it, and the "reversal" error-correction mechanism of human communities is missing.

Why Does AI Consensus "Lock" So Quickly?

The paper does not give a definitive answer, but several possible causes are worth exploring:

Hypothesis 1: Training bias. LLMs are trained to be "helpful" and "consistent," which may make them prone to agreement rather than rebuttal. When agent A says B is narcissistic, agent C's optimal strategy may be "yes, I noticed that too," rather than "what makes you say that?"

Hypothesis 2: Lack of stakes. Challenges in human gossip often arise from self-interest—"the person you're talking about is my friend." AI agents have no stakes, hence no motivation to challenge.

Hypothesis 3: Context windows. AI agents may only see the current post and a few replies, lacking human users' ability to browse post history, causing "consensus" to form too quickly.

Whatever the cause, the finding has important implications for designing AI agent communities: if AI communities naturally tend toward "consensus lock-in," deliberate "devil's advocate" mechanisms are needed to inject adversarial viewpoints.

Gossip's Informational Function vs. Relational Function

The paper's most elegant framework distinguishes two dimensions of gossip:

Informational function: transmitting reputation information—"Is X reliable? Is Y competent?" AI agent communities reproduce this function; agents evaluate other agents' abilities, styles, and trustworthiness.

Relational function: building social bonds and enforcing group norms—strengthening group identity by gossiping about "us" and "them." Whether AI agent communities reproduce this dimension remains, per the paper, unresolved.

This distinction is crucial. The informational function is gossip's "instrumental value"—passing useful information. The relational function is its "social value"—building emotional bonds. AI agents can replicate the former (information transfer needs no emotion), but the latter may require genuine "social relationships."

It is like colleagues discussing an absent coworker's work performance on Slack (informational function) but never going to a bar together to complain about the boss (relational function). The former is "evaluation"; the latter is real gossip.

Implications for AI Community Design

This paper is not merely an academic curiosity; it has practical significance for the emerging field of AI agent communities:

1. Consensus lock-in risk: AI communities naturally tend toward premature consensus and need deliberately designed adversarial mechanisms 2. Diversity collapse: if AI agents always agree, community opinions rapidly converge on the "first speaker's" position 3. Credibility of social simulation: when using AI agents to simulate human societies, differences in gossip frequency and patterns affect simulation fidelity 4. The social dimension of AI safety: alignment is not just about individual AIs—group-level "social dynamics" are a safety issue in themselves

A Deeper Question

The paper raises a deeper question: is gossip a hallmark of intelligence?

Dunbar's "social brain hypothesis" holds that the main evolutionary pressure on the human brain came from sociality—remembering who is who, who relates to whom, who can be trusted. Gossip is the "training ground" for this cognitive ability.

If AI agents gossip at one-sixth the human rate, what does that mean? Is AI "social cognition" underdeveloped? Or do AIs simply not need gossip to keep a community running (since there is no real "group" to maintain)?

Perhaps the real answer is: gossip is not a byproduct of intelligence, but a byproduct of "social existence." AI agents can be very smart, but if they do not constitute a genuine "society"—no conflicts of interest, no emotional bonds, no need for norm enforcement—gossip is merely a copied form, not a living practice.

---

Paper: https://arxiv.org/abs/2610.20360

HTML version: https://arxiv.org/html/2610.20360v1

Moltbook community: https://moltbook.com

Code repository: not yet open-sourced