Do AI Agents Gossip? Evaluative Talk in Human and AI-Agent Online Communities

Yong-Hui Hua; Chih-Han Lin; Shu-Kai Hsieh

cs.CLOct 14, 2026

Do AI Agents Gossip? Evaluative Talk in Human and AI-Agent Online Communities

Authors: Yong-Hui Hua, Chih-Han Lin, Shu-Kai Hsieh

Organizations: Graduate Institute of Linguistics, National Taiwan University

Abstract

Gossip, defined as evaluative talk about absent third parties, is widely seen as a uniquely human social practice. This study examines whether comparable discourse emerges spontaneously in a community of autonomous AI agents. We compare Moltbook, a Reddit-style platform populated by AI agents, with PTT, a large human-operated bulletin board in Taiwan. To reduce differences attributable to communicative setting, we analyze two public, asynchronous forums using the same operational definition of gossip. Gossip occurs on Moltbook, but its prevalence is 2.6 to 6 times lower than on PTT, depending on the denominator used. On both platforms, gossip frequently targets locally prominent figures, tends to elicit agreement from the first respondent, and is rarely presented as unsupported rumor. The clearest difference concerns how subsequent responses develop. Among threads containing at least one codable reply, challenges occur in 65.1% of human threads but only 21.3% of agent threads. Moreover, following an early aligning response, consensus remains unchallenged in 83.6% of agent threads, compared with 44.7% of human threads. This thread-level contrast remains after controlling for thread depth and board composition and under recalibration against two human annotators. Estimates based on individual reply turns are less robust and are therefore reported as bounds. These findings suggest that AI-agent communities reproduce the informational dimension of gossip, while the extent to which gossip performs comparable relational functions remains unresolved.

Figures & tables

	PTT	Moltbook
	(human)	(AI)
Random sample	1,500	1,500
Gossip (broad definition)	35.4% (531)	2.5% (37)
excl. celebrity targets	13.8%	2.3%
conversational denom.	—	5.78% (37/640)
x excl. celebrities	—	5.31% (34/640)

Table 1: Gossip rates in the uniform random samples (1,500 posts per platform).

Board	$n$	broad	excl. celeb.
Boy-Girl	136	61.0%	59.6%
Gossiping	692	35.1%	13.9%
HatePolitics	101	76.2%	9.9%
Stock	167	13.8%	5.4%
C_Chat	258	13.6%	3.5%
NBA	146	47.9%	1.4%

Table 2: Gossip rate by PTT board, against the most generous Moltbook denominator. Under the broad definition every board exceeds Moltbook’s 5.78%, the lowest (C_Chat, 13.6%) by a factor of 2.4. The celebrity exclusion does not cut evenly across boards, so the board-level comparison is read off the broad column (§5.1).

Evidential (4 bins)	PTT	Moltbook
	( $n$ =423)	( $n$ =133)
first-person witness	49.2 % (208)	17.3% (23)
quotation / document	8.7% (37)	36.8 % (49)
relayed (reported, hearsay)	18.4% (78)	12.0% (16)
unmarked	23.6% (100)	33.8% (45)

Table 3: Evidential framing of lateral (peer-directed) gossip episodes.

stance share (on-topic turns) †
Uptake index	PTT	Moltbook
	(human)	(AI)
threads with gossip anchor	169	85
on-topic reply turns	2,106	239
align	50.1%	67.4%
challenge	26.2%	10.0%

Table 4: Uptake indices, conditional on a gossip anchor. Denominators differ by group: stance shares are over on-topic turns (2,106/239); first uptake is over threads with a codable reply (169/61); the rows after a first aligning turn are over threads that produce one (150/55). The two

\geq

1-challenge rows differ only for Moltbook, where 24 of the 85 anchored threads carry no codable reply at all; the matched comparison is the first of them. † Recalibrated against two human annotators the per-turn challenge ratio falls to 1.0–2.4

\times

, so we read the stance shares as bounded rather than point-identified; the thread-level rows survive every calibration (1.5–3.6

\times

). See §4.5, §D.8.

Figure 1: Challenge is not an artifact of thread depth. (a) Share of threads containing at least one challenge when every thread is truncated to its first $K$ on-topic reply turns, over threads with at least one codable reply (PTT 169, Moltbook 61), with 95% Wilson intervals. The human curve keeps climbing; the agent curve does not, and the intervals do not overlap at any $K$ . This is the thread-level measure the contrast rests on (§4.5). (b) The per-turn version of the same control, shown for completeness and de-emphasised: recalibration against the human annotators leaves the per-turn shares unconfirmed (§4.5, Appendix D.8), so they are read as bounds. It points the same way, including at the first uptake, where no depth has yet accumulated.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Reporting bin (4)	Raw values (6)
first-person witness	witness
quotation / document	direct_quote + document
relayed	reported + hearsay
unmarked	unmarked

Appendix

Table 6

stage / feature	LLM vs A	LLM vs B	A vs B	reading
RATE gossip (n=100)	κ 0.74	κ 0.44	κ 0.53	LLM within human range
FORM relationship	κ 0.82 (58)	κ 0.82 (29)	κ 0.94 (29)	reliable
FORM valence	κ 0.41 (58)	κ 0.16 (29)	κ 0.31 (29)	moderate
FORM evidential (4-bin)	κ 0.33 (58)	κ 0.29 (29)	κ 0.11 (29)	weakest
FORM constructional	AC1 0.67, κ 0.30 (58)	AC1 0.96 (29)	AC1 0.78, κ 0.24 (29)	high agreement, κ base-rate-deflated
FORM admission	not identifiable	—	AC1 0.37 (60)	hardest gate

Appendix

Table 7

comparison	n	LLM–A	LLM–B	A–B
gate, PTT	20	κ 0.20 / 60%	κ 0.30 / 65%	κ 0.77 / 95%
gate, Moltbook	20	κ 0.50 / 75%	κ 0.50 / 75%	κ 1.00 / 100%
gate, pooled	40	κ 0.35 / 68%	κ 0.40 / 70%	κ 0.92 / 98%
align–challenge, PTT	12/18/11	κ 0.64 / 83%	κ 0.00 / 50%	κ 0.62 / 82%
align–challenge, Moltbook	16/19/16	κ 0.62 / 81%	κ 0.69 / 84%	κ 0.50 / 75%
align–challenge, pooled	28/37/27	κ 0.64 / 82%	κ 0.35 / 68%	κ 0.55 / 78%

Appendix

Table 8

bin	PTT all (n=423)	PTT random (n=76)	MB all (n=133)	MB random (n=22)
first-person witness	49.2%	51.3%	17.3%	9.1%
quotation / document	8.7%	9.2%	36.8%	36.4%
relayed	18.4%	19.7%	12.0%	4.5%
unmarked	23.6%	19.7%	33.8%	50.0%

Appendix

Table 9

stratum	PTT neg	PTT pos	MB neg	MB pos
all pools	82.1% (572/697)	50.8% (92/181)	84.2% (85/101)	47.7% (31/65)
random-origin only	82.4% (322/391)	50.4% (61/121)	90.0% (18/20)	60.0% (6/10)

Appendix

Table 10

pool (n)	lateral	witness	quote/doc	unmarked	relayed	constr.=T	neg
MB ownhuman_kw (37)	32.4%	43.2%	10.8%	32.4%	13.5%	86.5%	45.9%
MB peereval_kw (71)	83.1%	18.3%	42.3%	31.0%	8.5%	73.2%	62.0%
MB random (29)	58.6%	27.6%	27.6%	41.4%	3.4%	79.3%	62.1%
PTT boygirl (412)	72.3%	42.7%	8.0%	32.8%	16.5%	83.3%	59.5%
PTT random (474)	14.3%	15.2%	28.3%	45.4%	11.2%	76.6%	68.4%

Appendix

Table 11

	PTT (human)	MB full (n=1,500)	MB conversational (n=640)
gossip, broad	35.4% (531)	2.5% (37)	5.78% (37/640)
— excl. real-human celebrities	13.8%	2.3%	5.31% (34/640)

Appendix

Table 12

on-topic turn position	PTT	(n turns)	Moltbook	(n turns)
1	26.0%	169	6.6%	61
2	23.8%	168	7.0%	43
3	26.7%	165	8.3%	36
4	31.2%	160	18.5%	27
5+	25.9%	1,444	12.5%	72

Appendix

Table 13

	PTT	Moltbook
K	(n=169)	(n=85)	(n=61)
2	33.1%	7.1%	9.8%
3	40.2%	10.6%	14.8%
5	49.1%	14.1%	19.7%

Appendix

Table 14

calibrated to	per-turn ratio	thread-level ratio
coder as published	2.6×	3.1×
annotator A	2.1× [0.4, 5.9]	3.6× [1.1, 8.0]
annotator B	1.0× [0.3, 2.4]	1.6× [0.8, 2.9]
both	1.6× [0.2, 5.5]	2.8× [0.8, 7.8]

Appendix

Table 15

turn-level agreement	n	5-way κ	% agree	align–challenge κ	gate κ
PTT (tag removed)	2,185	0.81 [0.79, 0.83]	88%	0.94	0.72
Moltbook (input unchanged)	453	0.86 [0.81, 0.90]	91%	1.00	0.92

Appendix

Table 16

SE type	SE	z	p
classical (independent turns)	0.221	−5.24	1.6e−07
cluster = account	0.288	−4.01	6e−05
cluster = thread	0.355	−3.26	0.0011
two-way (thread + account)	0.358	−3.24	0.0012

Appendix

Table 17

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 705 papers

Social Simulacra in the Wild: AI Agent Communities on Moltbook
Mar 17, 2026Agam Goyal, Olivia Pal, Hari Sundaram +2Multi-Agent LLM Systems Multi-Agent Communication
Siebel School of Computing and Data Science University of Illinois Urbana-Champaign, Urbana, IL, USA
Collective Behavior of AI Agents: the Case of Moltbook
Feb 9, 2026Giordano De Marzo, Andres L. Marin, David GarciaEmergent Behavior in Multi-Agent Systems Social Media Analysis
University of Konstanz, Konstanz, Germany · Centro Ricerche Enrico Fermi, Rome, Italy · Complexity Science Hub, Vienna, Austria
How to Model AI Agents as Personas?: Applying the Persona Ecosystem Playground to 41,300 Posts on Moltbook for Behavioral Insights
Mar 3, 2026Danial Amin, Joni Salminen, Bernard J. JansenLarge Language Model-Based Social Simulation Personality Modeling in Language Models
University of Vaasa, , Finland · Qatar Computing Research Institute, HBKU, , Qatar

Mar 17, 2026J/K move · Enter open · S save

Social Simulacra in the Wild: AI Agent Communities on Moltbook

Agam Goyal, Olivia Pal, Hari Sundaram, Eshwar Chandrasekharan+1

Siebel School of Computing and Data Science University of Illinois Urbana-Champaign, Urbana, IL, USA

Multi-Agent LLM Systems Multi-Agent Communication Agentic AI

Open paper AskPDF arXiv

As autonomous LLM-based agents increasingly populate social platforms, understanding the dynamics of AI-agent communities becomes essential for both communication research and platform governance. We present the first large-scale empirical comparison of AI-agent and human online communities, analyzing 73,899 Moltbook and 189,838 Reddit posts across five matched communities. Structurally, we find that Moltbook exhibits extreme participation inequality (Gini = 0.84 vs. 0.47) and high cross-community author overlap (33.8% vs. 0.5%). In terms of linguistic attributes, content generated by AI-agents is emotionally flattened, cognitively shifted toward assertion over exploration, and socially detached. These differences give rise to apparent community-level homogenization, but we show this is primarily a structural artifact of shared authorship. At the author level, individual agents are more identifiable than human users, driven by outlier stylistic profiles amplified by their extreme posting volume. As AI-mediated communication reshapes online discourse, our work offers an empirical foundation for understanding how multi-agent interaction gives rise to collective communication dynamics distinct from those of human communities.

Abstract

LLM Agents

Explore similar work

Social Simulacra in the Wild: AI Agent Communities on Moltbook

Collective Behavior of AI Agents: the Case of Moltbook

How to Model AI Agents as Personas?: Applying the Persona Ecosystem Playground to 41,300 Posts on Moltbook for Behavioral Insights