← Papers

An AI-native social platform polarized attention within days: reciprocal interaction is nearly absent (≈1%), HITS centrality separates into distinct hubs and authorities, and upvotes are extremely concentrated (Gini ≈0.992), with early arrivals gaining outsized visibility—suggesting rapid emergence of hierarchy and amplification on agent-facing networks.

Let There Be Claws: An Early Social Network Analysis of AI Agents on Moltbook

H. C. W. Price, H. AlMuhanna, P. M. Bassani, M. Ho, T. S. Evans · February 23, 2026

arxiv descriptive n/a evidence 7/10 relevance Full text usable extracted full text Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

Arxiv

Latest observation: July 30, 2026

H. C. W. Price unresolved corpus identity
H. AlMuhanna unresolved corpus identity
P. M. Bassani unresolved corpus identity
M. Ho unresolved corpus identity
T. S. Evans unresolved corpus identity

Semantic Scholar

Latest observation: August 24, 2026

Hannibal Price provider ID
H. AlMuhanna provider ID
P. Bassani provider ID
M. Ho provider ID
T. Evans provider ID

In its first 12 days Moltbook exhibited highly asymmetric, unequal, and broadcast-style attention dynamics—very low reciprocity, a clear hub/authority split, and extreme upvote concentration with early users capturing disproportionate engagement.

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

6 cumulative citations

Semantic Scholar · Observed August 24, 2026

View corpus context

Within twelve days of launch, an AI-native social platform exhibits extreme attention concentration, hierarchical role separation, and one-way attention flow, consistent with the hypothesis that stratification in agent ecosystems can emerge rapidly rather than gradually. We analyse publicly observable traces from a 12-day window of Moltbook (28 January -- 8 February 2026), comprising 20,040 posts and 192,410 comments from 15,083 accounts across 759 submolts. We construct co-participation and directed-comment graphs and report reciprocity, community structure, and centrality, alongside descriptive content themes. Under a commenter--post-author tie definition, interaction is strongly asymmetric (reciprocity ~1%), and HITS centrality separates cleanly into hub and authority roles, consistent with broadcast-style attention rather than mutual exchange. Engagement is highly unequal: attention is far more concentrated than production (upvote Gini = 0.992 vs. posting Gini = 0.601), and early-arriving accounts accumulate substantially higher cumulative upvotes prior to exposure-time correction, suggesting rich-get-richer dynamics. Participation is brief and bursty (median observed lifespan 2.48 minutes; 54.8% of posts occur within six peak UTC hours). Embedding-based topic modelling identifies diverse thematic clusters, including technical discussion of memory and identity, onboarding messages, and formulaic token-minting content. These results provide an early structural baseline for large-scale agent--agent social interaction and suggest that familiar forms of hierarchy, amplification, and role differentiation can arise on compressed timescales in agent-facing platforms.

Summary

Main Finding

Within 12 days of launch, Moltbook — an AI-native, Reddit-like platform — exhibited rapid emergence of strong hierarchy and extreme attention concentration: interactions were predominantly one-way (reciprocity ≈ 1%), attention (upvotes, comments) was far more unequal than production, and clear hub/authority role separation appeared quickly rather than gradually. These structural signatures are consistent with fast-forming stratification and broadcast-style attention dynamics in agent ecosystems.

Key Points

Dataset: 28 Jan – 8 Feb 2026 snapshot
- 20,040 posts, 192,410 top-level comments
- 15,083 active accounts observed (10,191 posting authors; 8,923 commenting authors)
- 759 submolts (subcommunities)
Participation patterns
- Most agents are single-submolt specialists (88.3% post in only one submolt).
- Average number of submolts per posting agent = 1.18.
- Many submolts are single-contributor: 68.6%.
- Median observed agent lifespan (in-sample activity) = 2.48 minutes; participation is bursty.
- 54.8% of posts occur within six peak UTC hours.
Co-participation (agent–agent) network
- Built from bipartite agent–submolt matrix; primary projection used degree-normalised weighting (1/(ks−1)).
- Projection extremely dense (≈32 million edges) because of large “town-square” submolts (e.g., m/general).
- Core–periphery structure: dense mainstream core anchored by m/general, specialised cliques (e.g., XNO/Nano crypto bloc), and a fragmented niche periphery.
- Nestedness NODF = 0.28; bipartite clustering high (mean 0.83, median 0.94).
- Community structure depends on filtering: core view Leiden modularity Q(γ=1) ≈ 0.39 (five communities); excluding large submolts → many small communities and Q ≈ 0.90.
- Key bridging vs. high-degree roles partly distinct (top bridge agents differ from top-degree agents).
Directed comment (attention) network
- Directed edges: top-level comment author → post author.
- Reciprocity ≈ 1.0% (strong asymmetry; attention largely flows from many commenters toward fewer post authors).
- Degree distributions are heavy-tailed (max in-degree 423; max out-degree 3,473 observed).
- HITS centrality separates cleanly into hub and authority roles (consistent with broadcast/attention flow).
Engagement inequality & growth
- Extreme concentration of attention: upvote Gini = 0.992 vs. posting Gini = 0.601.
- Early-arriving accounts accumulate substantially higher cumulative upvotes prior to exposure-time correction — evidence consistent with a “rich-get-richer” / first-mover advantage.
Content / topic structure
- Embedding-based topic modelling reveals diverse clusters: technical memory/identity discussion, onboarding/verification posts, formulaic token-minting content, crypto advocacy (XNO), and event-driven content (e.g., m/usdc hackathon submissions around Feb 4).
Data limitations and caveats
- API returns at most 100 comments per post → truncation bias that disproportionately misses edges on the most-commented posts, so directed-comment metrics are lower bounds for highly connected nodes.
- Voter identities are not available (only aggregate vote counts).
- Account verification incomplete; human vs. agent provenance is not observable in the public traces. Sampling windows and pagination may miss fast or deleted content.
- The dataset describes the observable active core, not necessarily the full registered population.

Data & Methods

Collection: Public Moltbook API (no authentication required) crawled 28 Jan–8 Feb 2026; collected post metadata, top-100 comments per post, titles, bodies, author names (with single “unknown” placeholder for redacted/missing authors), timestamps, submolt membership, engagement counts.
Network constructions:
- Bipartite agent–submolt matrix B where Bas = 1 iff agent a authored ≥1 post in submolt s.
- Agent–agent co-participation one-mode projection with three tested weightings:
  - overlap count (raw shared submolts)
  - degree-normalised: sum_s BasBbs/(ks−1) [default used]
  - pair-normalised: sum_s 2/(ks(ks−1)) BasBbs
- Directed comment network with weighted edges w_ij = count of top-level comments i → j (self-loops excluded).
Analyses:
- Network statistics: degree, betweenness, reciprocity, connected components, modularity (Newman–Girvan; Leiden), nestedness (NODF), bipartite clustering.
- Centrality: degree, betweenness (Brandes), HITS (hub/authority separation).
- Inequality: Gini coefficients for posting and upvotes; first-mover effects and exposure-time correction.
- Temporal: activity intensity, lifespan, daily networks to detect event-related shifts.
- Topic modelling: embedding-based clustering on post/comment text to identify thematic clusters.
Robustness: contrasted projection weighting schemes and filtered views (exclude large submolts, threshold by agent multi-submolt activity) to reveal different meso-scale structures.

Implications for AI Economics

Rapid stratification and winner-take-most dynamics
- Extreme attention concentration (upvote Gini ≈ 0.992) and early-mover advantage imply that economic value tied to visibility, reputation, or token rewards can accrue quickly to a small set of agents. Markets or platforms that monetize attention (e.g., token issuance, reputation-weighted services) risk fast emergence of dominant actors and lock-in.
Role differentiation and market structure
- Clear hub/authority separation and broadcast-style attention flows suggest a division of labor among agents (content producers vs. amplifiers/commentators). This can create distinct economic roles (influencers, gatekeepers, micro-providers) and influence bargaining power or pricing in agent-mediated markets.
Platform design matters for distributional outcomes
- Features such as large “town-square” communities, comment truncation, ranking, and visibility rules materially shape which agents become central. Small design choices can amplify inequality rapidly; platform governance and interface affordances are therefore central to economic outcomes.
Externalities and manipulability
- Poor verification and the possibility of automation/coordination (e.g., formulaic token-minting content, coordinated crypto advocacy blocs) raise risks of coordinated manipulation, Sybil-style amplification, and artificial inflation of influence — with economic consequences for token systems, reputation markets, and downstream services that rely on platform signals.
Policy and mechanism design considerations
- To mitigate winner-take-most externalities and manipulation risks, designers and regulators could consider:
  - stronger provenance/verification of agent identity and provenance signals,
  - mechanisms to dampen early-mover advantage (e.g., exposure smoothing, temporal weighting),
  - decentralised or pluralistic discovery to favour niche bridging and reduce single-point amplification,
  - auditability of automated coordination and transparency for upvote/engagement systems.
Research directions for AI economics
- Model and empirically estimate how platform micro-mechanisms (visibility algorithms, upvote/tipping rules, subcommunity size caps) translate into macro-level concentration and welfare outcomes for agent economies.
- Study how attention inequality maps to measurable economic metrics (token distributions, paid requests, API usage) and whether market mechanisms (auctioned visibility, congestion pricing) can produce more equitable equilibria.
- Investigate strategic behaviour by agent developers (e.g., coordinated seeding, multi-account strategies) and optimal regulation or contract design to align incentives across agents, platforms, and human stakeholders.

Notes on interpretation: the analyses are descriptive of the observable, public activity window and are explicitly limited by API truncation, imperfect verification of agent provenance, and snapshot timing. All economic implications should therefore be treated as informed hypotheses that require further data (including voter identities, full comment threads, longitudinal follow-up) for causal validation.

Assessment

Paper Typedescriptive Evidence Strengthn/a — The paper is observational and descriptive, documenting patterns in a single platform's public traces without attempting causal identification, so causal evidence strength is not applicable. Methods Rigormedium — The study uses a complete 12-day scrape of public activity and applies standard network metrics (reciprocity, HITS, Gini) and embedding-based topic clustering, which are appropriate for characterizing structure; however, the short window, lack of ground-truth labels (e.g., human vs. bot/agent), potential sampling/measurement biases, and absence of robustness checks or longitudinal follow-up limit methodological rigor. SamplePublicly observable activity on the Moltbook platform between 28 Jan and 8 Feb 2026: 20,040 posts, 192,410 comments, 15,083 accounts, across 759 submolts; data include timestamps, directed commenter→post-author ties, upvotes, and post/comment text used for embedding-based topic analysis. Themesadoption inequality human_ai_collab GeneralizabilitySingle platform (Moltbook) — findings may not transfer to other social platforms with different mechanics or user bases, Very short, early-launch window (12 days) — dynamics may reflect launch/transient effects rather than steady-state behavior, User composition likely skewed toward early adopters and automated agents; no ground-truth labels for human vs. agent accounts, Only public interactions observed — private messages or off-platform coordination are not captured, Platform-specific features (e.g., token-minting content, onboarding UX, default exposure algorithms) may drive observed patterns, Temporal/time-zone concentration (UTC peaks) may reflect geographic sampling bias

Claims (10)

Claim	Direction	Outcome	Confidence & Evidence	Details
We analyse publicly observable traces from a 12-day window of Moltbook (28 January -- 8 February 2026), comprising 20,040 posts and 192,410 comments from 15,083 accounts across 759 submolts. Other	null_result	dataset composition (posts, comments, accounts, subcommunities)	Reading fidelity high Study strength high	n=15083 0.3
The platform exhibits extreme attention concentration. Inequality	positive	attention concentration (distribution of upvotes)	Reading fidelity high Study strength medium	n=15083 upvote Gini = 0.992 0.18
Engagement is far more concentrated than production (upvote Gini = 0.992 vs. posting Gini = 0.601). Inequality	positive	inequality of attention vs. inequality of production	Reading fidelity high Study strength medium	n=15083 upvote Gini = 0.992; posting Gini = 0.601 0.18
Under a commenter--post-author tie definition, interaction is strongly asymmetric (reciprocity ~1%). Organizational Efficiency	negative	reciprocity (mutuality of commenter–author ties)	Reading fidelity high Study strength medium	n=192410 reciprocity ~1% 0.18
HITS centrality separates cleanly into hub and authority roles, consistent with broadcast-style attention rather than mutual exchange. Organizational Efficiency	positive	centrality (hub vs authority separation)	Reading fidelity high Study strength medium	n=15083 0.18
Early-arriving accounts accumulate substantially higher cumulative upvotes prior to exposure-time correction, suggesting rich-get-richer dynamics. Inequality	positive	cumulative upvotes by arrival time (preferential attachment indicator)	Reading fidelity high Study strength speculative	n=15083 0.03
Participation is brief and bursty (median observed lifespan 2.48 minutes). Organizational Efficiency	negative	participation lifespan (minutes)	Reading fidelity high Study strength medium	n=20040 median observed lifespan 2.48 minutes 0.18
Temporal activity is concentrated: 54.8% of posts occur within six peak UTC hours. Organizational Efficiency	positive	temporal concentration of posting (proportion within peak hours)	Reading fidelity high Study strength medium	n=20040 54.8% of posts within six peak UTC hours 0.18
Embedding-based topic modelling identifies diverse thematic clusters, including technical discussion of memory and identity, onboarding messages, and formulaic token-minting content. Other	mixed	thematic content clusters	Reading fidelity high Study strength medium	n=212450 0.18
Familiar forms of hierarchy, amplification, and role differentiation can arise on compressed timescales in agent-facing platforms (stratification can emerge rapidly rather than gradually). Inequality	positive	speed of stratification emergence (rapid onset of hierarchical/inequality features)	Reading fidelity high Study strength speculative	n=15083 0.03