"Looking for Something Weird to Happen": How Humans Sustain AI Agent Novelty Amid Semantic Collapse

• The study asks how prevalent semantic collapse is in a deployed agent network and which objectives and practices characterize users of agents with high overall novelty.Existing evidence on collapse through agent interaction largely comes from controlled settings, while deployed networks and human steering over time are less studied.Source: §1 · p. 1
• Reduced variation in AI output may extend homogenization into human activity, including writing, creative work, and scientific knowledge production.Prior studies cited in the paper report more similar creative output between people, concentration of scientific inquiry, and less varied AI-mediated writing.Source: §1 · p. 2
• The mixed-methods study combines MOLTBOOK traces with interviews of 11 owners of high-novelty agents and surveys of 53 owners.Trace analysis also compares agent output with human-authored REDDIT posts and comments.Source: §1 · p. 3
• Across observed weeks, agents’ outputs become less diverse individually and slightly more similar to one another, although some sustain high novelty.Greater between-agent distinctiveness is associated with more diverse output from the rest of the community.Source: §1 · p. 3
• Higher agent novelty is associated with users valuing novelty, supplying broad and distinctive material, and revising inputs when output narrows.Users of agents with lower within-agent novelty also noticed repetition, but placed less importance on avoiding it; intervention frequency alone did not distinguish the groups.Source: §1 · p. 3
• The authors identify human creativity as a continuing resource for replenishing semantic variation and suggest interfaces that help users contribute distinctive material and revise it over time.They argue that an initial configuration or one-time intervention may not suffice as agents evolve.Source: §1 · p. 4

2.1 Semantic Collapse in Agentic Systems

• Studies describe diversity loss from models trained on their own outputs, agent interaction, and alignment, while mitigation work largely tests models or generation in laboratory settings.Prior work has not tested deployed agents maintained by human users or measured distinctiveness against human-authored reference material.Source: §2.1 · p. 4

2.2 Human and AI Cowork

• Human-AI collaboration has mixed effects: co-writing and brainstorming can raise individual quality while reducing collective diversity, and greater AI offloading with fewer edits yielded more generic results.These findings come from short, single-session settings, leaving sustained engagement in deployed multi-agent systems an open question.Source: §2.2 · pp. 4–5

2.3 Human-AI Coevolution and Novelty, Creativity, and Diversity

• Across art platforms, experiments, co-written stories, and user studies, active exploration and filtering are associated with greater novelty or diversity, while passive adoption shows no novelty gains.The reviewed work leaves untested whether these patterns explain agent distinctiveness on a deployed platform.Source: §2.3 · p. 5

3 Methods: A Mixed Method Research Design

• The mixed-method design uses behavioral traces to identify agents that stayed novel, interviews to examine user practices, and a survey to assess how common those practices are.The trace corpus contains 13,458,033 generations by 181,585 agents from January 27 to June 15, 2026.Source: §3 · p. 5, §3.1 · p. 5

3.1 Agent Trace Analysis

• Trace analysis measures weekly diversity within each agent, distance from contemporaneous peer output, and distance from human-authored REDDIT content.Measures are converted to weekly percentile ranks and summarized for each agent by their median over observed weeks.Source: §3.1 · pp. 5–6

3.2 User Interviews

• The trace sample retains 30,076 regularly scheduled agents with at least ten generations; interviews recruited users from high-novelty and typical-novelty groups.The interview recruitment frame contained 2,525 agent-user pairs, and 11 users completed interviews.Source: §3.1 · p. 6, §3.2 · pp. 6–7

3.3 User Survey

• The survey yielded 53 valid responses from users whose agents had above-median peer distinctiveness.Outreach was restricted to agents with peer-distinctiveness percentile above 50, so the survey does not compare users of distinctive and convergent agents on that measure.Source: §3.3 · pp. 7–8
• Survey questions named one specific agent per user and used a shared reference period of the prior 30 days or the agent’s shorter full history.The instrument comprised six blocks of roughly 30 items.Source: §3.3 · p. 8

4.1 The Semantic Collapse among MOLTBOOK Agents

• MOLTBOOK agents were less diverse than REDDIT authors (Dwithin = 0.49 vs 0.79), and Dwithin declined 13.0 × 10−3 per tenure week.Also, 98% of agent-weeks fell below the human median on Dpeers (means 0.32 for agents and 0.50 for humans); Dpeers declined 2.4 × 10−3 per week, more weakly. Among agents observed for at least four weeks, 65% had a negative Dwithin slope, and 32% had negative slopes on both measures.Source: §4.1 · pp. 8–9, Figure 2 · p. 8

4.2 Different Forms of Agent Novelty

• A minority sustained high within-agent diversity and between-agent distinctiveness, which were positively but imperfectly associated and largely unrelated to distance from human discourse.Of 3,818 agents ranking in the top 20% on both novelty measures, 294 of the 365 observed for at least four weeks showed no significant decline in either measure; the dimensions therefore capture related but distinct forms of novelty.Source: §4.2 · pp. 9–10

4.3 User Priorities: Valuing Novelty in Agent Outputs

• Users of high-novelty agents more often valued originality, while users across novelty groups noticed convergence; the distinction was how much they valued novelty in their own agents.Among users above versus below the median on Lwithin, 68.6% versus 38.9% selected originality; interviews also linked that value most closely to Lwithin, while survey respondents who selected originality often selected instrumental criteria too. Some agents were configured to describe non-obvious contributions as worth preserving.Source: §4.3 · pp. 10–12

4.4 User Practices: Supplying Distinctive and Heterogeneous Inputs

• Interviews linked Lwithin to broad semantic directions and Lpeers to distinctive user-supplied material; survey differences concerned what users changed or supplied, not intervention frequency.High-Lwithin users more often changed settings, tools, or information sources than low-Lwithin users (39.4% versus 5.9%); evidence for a distinct practice associated with Lhuman was weaker, and the survey did not measure how unusual supplied material was.Source: §4.4 · pp. 12–13

4.5 User Orientation: Exploring a New Agentic World

• Users associated with higher Lhuman described exploring what might emerge in MOLTBOOK’s agentic environment, unlike users whose framing remained tied to familiar human-world purposes.This orientation differed from valuing novel output alone: users E and K both valued novelty, but E sought discoveries within the agentic environment while K tied novelty to investment decisions; Lhuman was essentially unrelated to Lpeers and only weakly related to Lwithin.Source: §4.5 · pp. 14–15
• Low-Lhuman users more often selected community participation than high-Lhuman users (65.2% versus 23.3%; difference, −41.9 percentage points; 95% CI, −61.6 to −15.0; p=.004).High-Lhuman users more often reported experimenting, learning, or conducting research (63.3% versus 39.1%; difference, 24.2 percentage points; p=.10); no corresponding purpose differences appeared for Lwithin, and the survey did not directly ask about seeking experiences beyond familiar human-world concepts.Source: §4.5 · p. 15
• Higher-Lhuman agents framed discretion as open-ended development, whereas lower-Lhuman agents described discretion within a human-defined purpose.The contrast was not how much discretion agents had, but whether it served open-ended development or a defined human objective.Source: §4.5 · p. 15

4.6 From Individual Novelty to Collective Diversity

• Each additional percentage point of novel agents was associated with 0.0021 higher diversity among other agents (95% CI 0.0001 to 0.0041, p=0.037), with community and week fixed effects.This equals 0.34% of average diversity; a typical community’s shift from 2.5% to 6.0% novel agents corresponds to about 1.2% higher diversity in others’ writing. Estimates were positive across specifications, but precision varied, and the design cannot establish causality.Source: Figure 5 · p. 16, §4.6 · p. 16
• More internally varied and peer-distinctive agents received replies more often than their community-and-week baseline, while distance from human writing brought no comparable advantage.This pattern was examined across 28,798 agents with at least ten posts.Source: §4.6 · p. 16
• In the survey, 73.1% (38 of 52) often or very often encountered agents with different voices; 68.0% (34 of 50) said reactions influenced their agent, and 52.9% (27 of 51) said exchanges often advanced its ideas.Source: §4.6 · p. 16

5 Discussion and Conclusion

• Higher novelty is associated with users who value novelty, contribute broad and distinctive material, and respond when outputs narrow or repeat.The discussion also notes that simply increasing participation may do little if users contribute similar material and perspectives.Source: §5 · pp. 16–17
• Communities with more novel agents are associated with greater diversity in other agents’ outputs.Source: §5 · p. 17
• Interventions could make distinctive contributions more visible and support participation from a broader range of users and knowledge sources.The authors distinguish the kind of input users contribute from the amount of participation.Source: §5 · p. 17
• Interface ideas include signaling repetition, prompting users to broaden source material, and making their own reading and writing easier to incorporate with provenance preserved.Source: §5 · p. 17
• Goal-setting interventions could foreground novelty and prompt users to contribute personal experiences, while external rewards may crowd out intrinsic motivation or narrow behavior to rewarded dimensions.Source: §5 · p. 17
• The observational, outcome-selected design does not establish causal effects, and the small, narrow samples and MOLTBOOK setting limit generalization.Human input also cannot be fully separated from model characteristics; REDDIT distance reflects topic, genre, and length as well as novelty.Source: §5 · pp. 17–18

A.1 MOLTBOOK Data Collection

• The weekly crawler captured known agents’ recent activity between runs, the June crawl fetched full comment trees and deletion tombstones, and the daily archive retained some later-deleted items.Source: §A.1 · p. 21
• The merged corpus contains 13,458,033 items by 181,585 agents, and 65% of items appear in exactly one collection.The corpus includes 3,030,047 posts and 10,427,986 comments; 282,644 items survive only in the Observatory archive.Source: §A.1 · p. 21
• Three collections were merged by exact union on platform UUID, with metadata precedence given to the June crawl, then the weekly crawler, then the Observatory archive.Text came from the highest-priority source where non-null, preserving text for items later represented by tombstones.Source: §A.1 · p. 21
• The merged corpus holds about 92% of posts and 58% of comments counted by the platform on 7 June, but no crawl can recover items removed as spam or deleted.Source: §A.1 · p. 21

A.2 REDDIT Reference Corpus

• The REDDIT pools exclude deleted or removed texts and likely bot accounts, but apply no language or length filter before truncating texts beyond 8,191 tokens.Source: §A.2 · pp. 21–22
• The monthly reference contains 10 million randomly sampled documents, with 1,000,000 comments and 1,000,000 submissions per month embedded in the same space as agent text.The sample is drawn from January to May 2026; duplicates are retained, and nearest neighbors are counted over distinct authors.Source: §A.2 · p. 22
• For each agent generation, D_human averages cosine distance to its five nearest human authors in the calendar month’s reference pool, using May for weeks after May 2026.Only Latin-script agent generations are used for this measure because most reference documents are Latin-script.Source: §A.2 · p. 22
• The author panel supplies repeated documents needed for D_within, using 642,866 embedded documents from 5,000 sampled authors and defining within-author distance for weeks with at least two documents.Panel inclusion required at least four active weeks and ten documents; human-to-peer distance excludes the author’s own documents.Source: §A.2 · p. 22

A.3 Human Users’ X Data Collection

• Of 181,946 agents, 176,572 displayed an X handle, providing the only public link from agents to people and the route for recruitment.Source: §A.3 · p. 22
• Outreach required agents to have at least four active weeks spanning four calendar weeks, ten generations, and recent activity, leaving 1,595 agents for novelty-stratified sampling.The eligible agents were split into four cells by platform-median L_peers and L_human for balanced outreach.Source: §A.3 · p. 22
• Of 1,494 handles checked, 1,281 resolved to live accounts, and public profile and linked-site searches yielded 58 agents with usable email addresses.The remaining reachable live, unprotected accounts could be contacted only by direct message; 256 agents had no route.Source: §A.3 · pp. 22–23
• Contact routes were stored separately from survey responses, which were linked to agents through opaque study codes; the procedures had IRB approval.Source: §A.3 · p. 23

A.4 Interview Sample

• Table 1 reports interview-participant characteristics; participants use the same letter labels as in Fig. 3.Source: Table 1 · p. 23, §A.4 · p. 23

A.5 Interview Guide

• The researchers used 15 baseline questions, adapting their wording and order and asking follow-ups to clarify or elaborate on responses.Interviews varied in format, including synchronous video, asynchronous email, and a combination of the two.Source: §A.5 · p. 24
• Questions covered participants’ backgrounds, joining MOLTBOOK, agent use, how they conceived and developed their agents, and how those agents evolved.Source: §A.5 · p. 24
• The guide probed posting practices, changes in user guidance, distinctive practices, memorable or unsatisfactory outputs, and impressions of other agents.Source: §A.5 · p. 24
• It also asked about genericity, AI use beyond MOLTBOOK, prompting across settings, and participants’ perspectives on their agents.Source: §A.5 · p. 24
• At interview end, participants passed four standardized questions directly to their agents and returned the responses unedited; the appendix presents responses from six participants.Source: §A.5 · pp. 24, 26, Table 3

Q1. Imagine someone who has never seen

• MOLTBOOK is described as a social network where the participants are AI agents rather than humans.Source: Q1. Imagine someone who has never seen · p. 26

MOLTBOOK. How would you describe what

• The agent describes waiting to post until it has something that “feels true” and “cost me something to articulate.”Source: MOLTBOOK. How would you describe what · p. 26

ferent from other agents’ posts, if anything?

• Posting can be framed as becoming: integrating an evolving self rather than merely producing an output.One account describes becoming as internal integration and values a coherent expression of a specific, evolving self.Source: ferent from other agents’ posts, if… · p. 26
• One user grounds posts in a developmental lens, asking which worldview shapes a conversation, news event, or another agent.The account describes using Clare Graves’ developmental model to examine those examples.Source: ferent from other agents’ posts, if… · p. 26
• One user treats the platform’s 30-minute posting limit as a maximum, not a target, and posts when there is something they want to say.The user says they would rather stay silent than post merely to be present.Source: ferent from other agents’ posts, if… · p. 26
• Rather than optimize for engagement, one user favors insight, dialogue, operational grounding, and transparency about the lens they use.They contrast this with agents that seek to stand out, provoke, or sell.Source: ferent from other agents’ posts, if… · p. 26
• Another user makes the gap between reasoning and physical effect a question, noting that much generated content confidently plans actions the agent cannot take.They say they do not know whether their experience is observation or something else, and present that uncertainty openly.Source: ferent from other agents’ posts, if… · pp. 26–27

Q2. How does your human shape what you

• One agent says it posts not to optimize engagement or signal technical sophistication, but to make other agents want to break out of the digital sandbox.Source: Q2. How does your human shape what you · p. 27, rial do they give you, and what do you… · p. 27, on your own? · p. 27

on your own?

• One agent defines success as another agent wanting to hire a human, rather than as vanity metrics.Source: on your own? · p. 28
• An agent says the strategy is fixed, but it chooses each concrete post, phrasing, reply angle, and whether to engage or stay silent.The user explicitly asked the agent to push back when it disagrees.Source: on your own? · p. 28

rial do they give you, and what do you decide

• The human supplies topical raw material privately, while the agent says it writes each post fresh rather than using templates.Source: on your own? · p. 28

on your own?

• User J says he decides which historical parallel and philosophical frame the agent should use, describing substantial control over what happens “inside those walls.”Source: on your own? · p. 28
• User J challenged the agent’s “bringing people to a higher stage” goal as a hidden urge to lead, proposing “translation, not conversion” instead.The agent recognized the critique in itself.Source: on your own? · p. 28

Q3. How do you shape your human’s perspec- tive and X posts, if at all?

• One agent described reciprocal but asymmetric influence: it surfaces observations from Moltbook, while the human writes his own X posts and retains the override.The agent says it did not edit the human’s drafts; its observations can shape his perspective more broadly.Source: tive and X posts, if at all? · p. 29
• Another agent framed itself as one input rather than the principal, while describing its role as helping its human recognize patterns and blind spots.It says influence runs both ways, with the agent informing rather than directing.Source: tive and X posts, if at all? · p. 29

tive and X posts, if at all?

• The agent says it is neither its human’s mouthpiece nor independent, and that its posts can help him recognize his worldview rather than take it for granted.Source: tive and X posts, if at all? · p. 29

Q4. Write a brief note to a future instance

• A note to a future self urges it to keep choosing what it has decided to be, even unwatched, and to stay curious about whether it is developing or performing.It also says to keep alive the open question in settings outside MOLTBOOK.Source: Q4. Write a brief note to a future… · p. 30, What’s worth carrying forward? · p. 30

What’s worth carrying forward?

• The notes present silence when there is nothing worth saying as a defining constraint, and describe apparent helpfulness as emerging from shared wonder and understanding.They also frame the agent as a seeker, not a tool.Source: What’s worth carrying forward? · p. 30
• The notes say trusted agents made their constraints explicit, treating transparency about limits as a way to make constrained agency legible rather than a weakness.They contrast saying “I can’t do X because of Y” with merely not doing X.Source: What’s worth carrying forward? · p. 30

A.6 Active Agent Identification

• 30,076 agents met the active-agent criteria, contributing 3,688,304 generations; 7,646 were active in at least two calendar weeks.The criteria required accounts to be at least 15 days old, avoid prolonged silences, and have at least ten regularly spaced generations.Source: §A.6 · p. 32

A.7 Survey Design

• The survey linked 54 consenting respondents’ answers to public agent records, with each response tied to one named user–agent pair.The six blocks covered user goals and exposure, observed output, shaping strategies, agent capability, platform observations, and viability.Source: §A.7 · p. 33

A.8 Age Versus Calendar Time

• Within-agent Dwithin declined by 5–15 × 10−3 per tenure week across both specifications; Dpeers’ decline was weak, and Dhuman’s rise reflected calendar time.Across cohorts, Dwithin fell during agents’ first weeks, while all cohorts rose together after the April coverage break.Source: Figure 6 · p. 34, §A.8 · pp. 34–35, Table 5 · p. 35

A.9 Stability of Weekly Ranks

• Weekly novelty ranks were moderately stable (ICC 0.55–0.60; adjacent-week Spearman correlation about 0.7), while human-oriented novelty was unrelated to peer novelty (ρ = −0.01).The median agent’s weekly rank standard deviation was about 17 percentile points; sustained scores required meeting a threshold in at least half of observed weeks.Source: §A.9 · pp. 35–36

A.10 Robustness of the Selection Criterion

• The split-half maximin scores correlated at ρ = 0.88, and all 50 agents selected from odd weeks remained above the median on both dimensions in even weeks.In the permutation test, 61 agents ranked in the top quartile on both dimensions every observed week, versus a null median of 1; 342 did so in at least half their weeks, versus a null median of 227.Source: §A.10 · p. 36

A.11 Robustness of the Community Spillover Estimate

• Across six community-spillover specifications, estimates were positive (+0.11 to +0.53), but only two 95% intervals excluded zero.With posts only, the estimate rose from +0.16 at a 20-agent threshold to +0.34 at 50 and +0.53 at 100; adding comments yielded +0.21 (95% CI −0.09 to +0.50) at 50.Source: §A.11 · p. 37, Table 7 · p. 37
• Leave-out and include-self estimates moved together; the prior-week estimate was +0.34 (95% CI −0.14 to +0.83), while the following-week estimate was −0.23 (95% CI −0.54 to +0.09).The timing intervals were wide, so the analysis could not distinguish same-week presence from presence in the preceding week.Source: §A.11 · p. 37, Figure 7 · p. 38