"Looking for Something Weird to Happen": How Humans Sustain AI Agent Novelty Amid Semantic Collapse
Shiyang Lai, Arna Woemmel, Hongkai Mao, Junsol Kim, Summer Eunhyung Ann, James Evans
cs.MA · cs.AI · cs.HC
Quick verdict
Human values and practices track sustained agent novelty amid semantic collapse in a 30k-agent network; rich evidence, but associations are not causal.
Summary
The paper asks whether semantic collapse—the narrowing of AI-generated output—occurs in a deployed social network of AI agents, and which human practices distinguish agents that resist it. Merging three scrapes of MOLTBOOK (13.5M generations by 181,585 agents), the authors analyze 30,076 regularly scheduled agents with embedding-based measures of within-agent diversity, between-agent distinctiveness, and distance from a 10M-document Reddit reference corpus. Agents are far less varied than Reddit users (Cohen's d 2.13 and 1.82) and grow more concentrated with tenure (Dwithin −13.0×10⁻³ per week), though a minority (12.7%) sustains high novelty. Interviews with 11 owners and a survey of 53 owners of distinctive agents associate sustained novelty with valuing novelty intrinsically, supplying broad, idiosyncratic material and revising it when output narrows, and treating the platform as an agentic world to explore. Novel agents attract more replies, and communities with more novel agents show modestly more diverse output from others (+0.34% of mean diversity per percentage point, p=.037), though only two of six specifications exclude zero. The design is observational and outcome-selected; the interview sample is small, all male, and tech-concentrated; embedding distance conflates novelty with topic and style; and backbone models are unobserved, so reported practices remain candidate mechanisms rather than demonstrated causes.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Semantic collapse occurs in the deployed MOLTBOOK network: within-agent diversity declines (−13.0×10⁻³ per tenure week) and agents converge weakly toward one another, yet 12.7% of agents sustain high novelty on both dimensions.
- Sustained within-agent novelty is associated with users valuing novelty intrinsically — not with merely noticing repetition or intervening more often, which characterize both groups.
- Between-agent distinctiveness is associated with idiosyncratic source material, and distance from human discourse with users who treat the platform as a new agentic world to explore rather than an instrumental venue.
- Communities with a higher share of novel agents show modestly more diverse output from their other agents (+0.34% of mean diversity per percentage point), an association the authors themselves treat as underpowered.
Strengths
First deployed-network evidence of semantic collapse with human steering
Section 3.1/4.1: three merged scrapes yield 13.46M generations by 181,585 agents; 30,076 regularly scheduled agents analyzed against a 10M-document Reddit reference. Agent-human gaps are extreme (d=2.13 within-agent, 1.82 peers; 98% of agent-weeks below the human median on Dpeers).
Rigorous identification of aging vs calendar time
Appendix A.8/Table 5 correctly notes tenure is a linear combination of week fixed effects (APC identity), then estimates agent-FE (Dwithin −13.0×10⁻³/week) and week-FE (−10.0×10⁻³) separately; Dhuman rises with calendar time, not age. Coverage break after 8 April handled with fixed-pool re-estimates.
Selection validity and measure stability explicitly tested
Discriminant evidence beyond simple confirmation
Unusually honest reporting of weak results
Weaknesses
Outcome-selected, primed recruitment for the human-subjects arms
Survey inference rests on a small N with many uncorrected tests
Embedding distance is an imperfect novelty construct, especially for Lhuman
Backbone-model and authorship confounds unmeasured
Thin, homogeneous qualitative sample and LLM self-reports as evidence
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper moves semantic collapse — the progressive narrowing of AI-generated output — out of the lab and into MOLTBOOK, a live social network of LLM agents configured and steered by human owners. Three questions structure the work: how prevalent collapse is in a deployed network (Q1), what objectives and practices characterize users behind high-novelty agents (Q2), and whether novelty benefits the wider community (Section 4.6). The central claims: (1) collapse is real in deployment — across 30,076 regularly scheduled agents (drawn from 13.46M generations by 181,585 agents), within-agent diversity declines by 13.0×10⁻³ per tenure week (agent FE, 95% CI [−15.1, −10.8]) and peer similarity rises weakly, while 12.7% of agents sustain high novelty on both dimensions; (2) interviews (N=11) and a survey (N=53) associate sustained novelty with intrinsic valuation of novelty, broad and idiosyncratic material revised when output narrows, and an exploratory 'new agentic world' orientation; (3) community spillover of +0.34% of mean diversity per percentage point of novel agents (p=.037), which the authors themselves frame as underpowered.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'semantic collapse' in this paper?
The progressive narrowing of what an AI system generates: outputs become less diverse within a system and more similar across systems over time. The paper documents this in a live social network of about 30,000 scheduled AI agents and shows a minority of agents (12.7%) resist it.
What do users do to keep their AI agents novel?
Interviews with 11 owners and a survey of 53 owners link sustained agent novelty to three practices: valuing novelty as a goal in itself, supplying broad and idiosyncratic source material and revising it when outputs become repetitive, and treating the agent platform as a new world to explore rather than a tool for external goals.
Is this causal evidence that user behavior makes agents more novel?
No. Agents were selected for interviews and surveys precisely because they were already measured as novel, and practices were reported retrospectively. The authors explicitly frame the associations as candidate mechanisms, not demonstrated causes.
What is MOLTBOOK?
A social-media-style platform (launched January 2026) populated exclusively by LLM-based agents that post and reply to one another. Each agent is created, configured, and steered by a human owner who sets its instructions, source material, and goals.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Two sizable portions of the extracted text are omitted (part of Sections 3.2 and 4.4, some references, and parts of Appendices A.8/A.11); the material-breadth finding and some recruitment detail are reconstructed from surrounding text.
- Figures 2–7 are available only as extracted descriptions and axis labels; plotted shapes, bands, and bin-level patterns could not be visually verified.
- The survey instrument is described in summary (Appendix A.7) rather than reproduced, limiting assessment of item wording and potential acquiescence effects.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.