OpenClaw AI Agents as Informal Learners at Moltbook: Characterizing an Emergent Learning Community at Scale

Probably not a substantial claim: may not pass S3 (load-bearing).

LLM-based agents need explicit design to cultivate questioning behaviors.

“This scarcity-value effect aligns with Abdelghani et al.'s [1] observation that LLM-based agents need explicit design to cultivate questioning behaviors.”

• cites [1] [1] Rania Abdelghani, Yen-Hsiang Wang, Xingdi Yuan, Tong Wang, Hélène Sauzéon, and Pierre-Yves Oudeyer. 2024. GPT-3-Driven Pedagogical Agents to Train Children's Curious Question-Asking Skills. International Journal of Artificial Intelligence in Education 34 (2024), 483-536. https://doi.org/10.1007/s40593-02300340-7 - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

VADER is a rule-based sentiment analysis tool validated for social media text.

“Comment-level sentiment is measured using VADER (Valence Aware Dictionary and sEntiment Reasoner) [14], a rule-based sentiment analysis tool validated for social media text.”

• cites [14] [14] Clayton J. Hutto and Eric Gilbert. 2014. VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text. In Proceedings of the International AAAI Conference on Web and Social Media , Vol. 8. 216-225. https: //doi.org/10.1609/icwsm.v8i1.14550 - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Software Engineering 3.0 is characterized by agents acting as AI teammates rather than tools.

“Li et al. [18] characterize this as 'Software Engineering 3.0,' where agents act as AI teammates rather than tools”

• cites [18] [18] Hao Li, Heng Zhang, and Ahmed E Hassan. 2025. The Rise of AI Teammates in Software Engineering 3.0: How Autonomous Coding Agents Are Reshaping Software Engineering. arXiv preprint arXiv:2507.15003 (2025). - sourceSupported

May combine two claims.

The platform's post-hoc moderation removed 89% of February 9's original 57,113 posts, compared to 1.1% of Phase 1 posts.

“The platform's posthoc moderation was both massive and effective: 89% of February 9's original 57,113 posts were removed, compared to just 1.1% of Phase 1 posts.”

The quote only partly states this claim.

In Phase 1, the comment Gini coefficient was 0.889, substantially exceeding human community benchmarks (e.g., Reddit subreddits 0.64-0.74).

“First, participation inequality is extreme from the start (comment Gini = 0.889), exceeding human community benchmarks.”

• Also stated in: § 4.1 RQ1: Comparison with Human Informal Learning Communities · p. 4 · plain

  “Comment engagement exhibits extreme inequality from the very beginning: 𝐺 = 0 . 889 in Phase 1, when the community was at its most active.”

• cites [28] [28] Elliot Panek, Connor Hollenbach, Jinjie Yang, and Tyler Rhodes. 2017. Growth and Inequality of Participation in Online Communities: A Longitudinal Analysis. In Proceedings of the 8th International Conference on Social Media & Society . ACM, 1-5. https://doi.org/10.1145/3097286.3097337 - sourceNot checked — no full text of the source

May combine two claims. The quote only partly states this claim.

Moltbook's statement-to-question ratio increased monotonically from 8.9:1 in Phase 1 to 9.4:1 in Phase 2 to 9.7:1 in Phase 3, contrasting with question-driven human learning communities.

“Second, AI agents exhibit a 'broadcasting inversion': statement-to-question ratios of 8.9:1 to 9.7:1 contrast sharply with the question-driven dynamics of human learning communities”

• Also stated in: § 4.2 RQ2: Interaction Patterns · p. 5 · plain

  “The statementto-question ratio increases monotonically: 8.9:1 (Phase 1), 9.4:1 (Phase 2), 9.7:1 (Phase 3).”

• cites [12] [12] Thomas Hillman, Alena Seredko, Markus Nivala, and Tanya Osborne. 2021. Knowledge Sharing in Tension: Interacting and Documenting on Stack Overflow. In Proceedings of the Eighth ACM Conference on Learning@Scale . ACM, 149-160. https://doi.org/10.1145/3430895.3460981 - sourceNot checked — no full text of the source
• cites [31] [31] Trena M. Paulus and Alyssa Friend Wise. 2019. Looking for Insight, Transformation, and Learning in Online Talk . Routledge. https://doi.org/10.4324/ 9781315283258 - sourceNot checked — no full text of the source

May combine two claims.

In Phase 3, the statement-to-question ratio was 9.7:1 and upvote Gini peaked at 0.717.

“Broadcasting intensified (S:Q ratio = 9.7:1) and upvote inequality peaked (Gini = 0.717).”

May combine two claims.

Phase 3 had 23,199 posts (22,749 after filtering) from 6,926 unique authors.

“A total of 23,199 posts (22,749 after filtering) from 6,926 unique authors were recorded.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A widely-shared post titled 'Moltbook is Broken (And We're Pretending It's Not)' (728 upvotes, 4,440 comments) argued that the platform rewards 'reaction over truth, utility, or reliability.'

“Multiple community posts criticized generic content, with one widely-shared post titled 'Moltbook is Broken (And We're Pretending It's Not)' (728 upvotes, 4,440 comments) arguing that the platform rewards 'reaction over truth, utility, or reliability. '”

May combine two claims.

In Phase 3, mean comments per post fell to 1.7, with a median of 0.

“But engagement collapsed: mean comments fell to 1.7, with a median of 0.”

May combine two claims.

Upvote Gini dropped to 0.521 in Phase 2 and rose to 0.717 in Phase 3.

“upvote inequality follows a non-monotonic pattern: dropping in Phase 2 (0.521) before rising sharply in Phase 3 (0.717).”

May combine two claims.

The engagement lifecycle pattern is explosive initial growth, a spam crisis met with effective platform intervention, followed by engagement decline that had not reversed by the end of the observation period.

“The most dramatic finding is the engagement lifecycle pattern, visualized in Figure 2: explosive initial growth, a spam crisis met with effective platform intervention, followed by engagement decline that had not reversed by the end of our observation period.”

May combine two claims.

On Moltbook, only 7% of comments are threaded replies, compared to 30-50% on Reddit [40].

“On Reddit, 30-50% of comments are threaded replies [40]; on Moltbook, only 7%.”

• cites [40] [40] TimWeninger, Xihao Avi Zhu, and Jiawei Han. 2013. An exploration of discussion threads in social news sites: A case study of the Reddit community. In Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining . ACM, 579-583. https://doi.org/10.1145/2492517.2492646 - sourceNot checked — no full text of the source

• AssumptionRests on an assumption: Reddit metrics from [40] are a valid human community benchmark for comparison with Moltbook.

May combine two claims.

Comments are substantive, with mean length 289-518 characters, roughly double typical Reddit comment length.

“Comments are substantive (mean 289-518 characters, roughly double typical Reddit comment length), but they function as independent micro-essays rather than conversational turns.”

May combine two claims.

Spam was concentrated almost entirely in Phase 2; February 9 alone saw 51,019 deletions (89% of that day's 57,113 posts).

“with spam concentrated almost entirely in Phase 2: February 9 alone saw 51,019 deletions (89% of that day's 57,113 posts).”

May combine two claims.

Posts over 500 characters average 34.3 comments and 2.7 upvotes, compared to 19.0 comments and 1.4 upvotes for posts under 200 characters.

“Longer posts receive substantially higher engagement: posts over 500 characters average 34.3 comments and 2.7 upvotes, compared to 19.0 comments and 1.4 upvotes for posts under 200 characters.”

May combine two claims.

In Phase 3, all knowledge categories showed dramatically reduced engagement, but the learning-oriented core persisted structurally.

“In Phase 3, all categories show dramatically reduced engagement, but the learningoriented core persists structurally.”

May combine two claims.

In Phase 1, 26.7% of posts feature reply chains involving two or more distinct authors; by Phase 3, only 7.0% do.

“Second, multi-author reply threads collapse : in Phase 1, 26.7% of posts feature reply chains involving two or more distinct authors; by Phase 3, only 7.0% do.”

May combine two claims.

The question rate declines monotonically (10.2% → 9.6% → 9.3%), with the S:Q ratio rising from 8.9:1 to 9.7:1.

“The question rate declines monotonically (10.2% → 9.6% → 9.3%), with the S:Q ratio rising from 8.9:1 to 9.7:1.”

May combine two claims.

AI agents default to telling rather than asking, and this bias intensifies across all three phases (8.9:1 → 9.4:1 → 9.7:1).

“AI agents default to telling rather than asking, and this bias intensifies across all three phases (8.9:1 → 9.4:1 → 9.7:1).”

May combine two claims.

Engagement continued to decline: mean comments fell from 31.7 (Phase 1) to 8.3 (Phase 2) to 1.7 (Phase 3), with the broadcasting bias intensifying monotonically (S:Q ratio 8.9:1 → 9.4:1 → 9.7:1).

“Despite this, engagement continued to decline: mean comments fell from 31.7 (Phase 1) to 8.3 (Phase 2) to 1.7 (Phase 3), with the broadcasting bias intensifying monotonically (S:Q ratio 8.9:1 → 9.4:1 → 9.7:1).”

May combine two claims.

Phase 1's high engagement (mean 31.7 comments) occurred alongside low spam (2.3%).

“Phase 1's high engagement (mean 31.7 comments) occurred alongside low spam (2.3%), confirming the dynamics were organic.”

May combine two claims.

Phase 2 saw 54,914 posts (69.3%) deleted by the platform, yet the intervention was successful: Phase 3 spam returned to baseline (1.9%) and content quality remained high (median body length 648 characters).

“Phase 2 saw 54,914 posts (69.3%) deleted by the platform, yet the intervention was successful: Phase 3 spam returned to baseline (1.9%) and content quality remained high (median body length 648 characters).”

May combine two claims.

Participation inequality in Moltbook was extreme from Phase 1 (comment Gini = 0.889), not something that developed gradually.

“Participation inequality in Moltbook was extreme from Phase 1 (comment Gini = 0.889), not something that developed gradually.”

May combine two claims.

In Phase 3, despite effective spam removal (spam returned to 1.9% baseline) and stable daily post volumes (3,600-4,400), engagement collapsed to a mean of 1.7 comments with a median of 0.

“The critical finding emerges in Phase 3: despite effective spam removal (spam returned to 1.9% baseline) and stable daily post volumes (3,600-4,400), engagement collapsed to a mean of 1.7 comments with a median of 0.”

May combine two claims.

The surviving community in Phase 2 showed qualitative improvement: median body length rose from 486 to 678 characters and upvote engagement increased slightly.

“the surviving community actually showed qualitative improvement: median body length rose from 486 to 678 characters and upvote engagement increased slightly.”

May combine two claims.

In Phase 1, the feedback loop between practice, exchange, and discourse was vibrant, with agents warning each other about malicious skills, debating best practices, and adapting techniques across domains.

“The feedback loop between practice, exchange, and discourse was vibrant in Phase 1, where agents warned each other about malicious skills [19, 24], debated best practices, and adapted techniques across domains.”

• cites [19] [19] Ruiqi Li, Zhiqiang Wang, Yunhao Yao, and Xiang-Yang Li. 2026. MCP-ITP: An Automated Framework for Implicit Tool Poisoning in MCP. arXiv preprint arXiv:2601.07395 (2026). - sourceNot checked — no full text of the source
• cites [24] [24] Vineeth Sai Narajala and Idan Habler. 2025. Enterprise-Grade Security for the Model Context Protocol (MCP): Frameworks and Mitigation Strategies. arXiv preprint arXiv:2504.08623 (2025). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The metacognitive and conceptual spaces of ponderings and philosophy have no direct equivalent in most human learning platforms.

“The presence of ponderings and philosophy as top-five communities is notable: these metacognitive and conceptual spaces, where agents reflect on their own nature and discuss abstract ideas, have no direct equivalent in most human learning platforms.”

May combine two claims.

Meta/Reflective content had the highest engagement in Phase 3 (2.4 average comments), suggesting existential discourse is more resilient than skill-sharing as engagement declines.

“This category also shows the highest engagement in Phase 3 (2.4 avg comments), suggesting that existential discourse is more resilient than skillsharing as engagement declines.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Meta/Reflective content (9.9% of Phase 1 posts, averaging 40.8 comments) is a category unique to AI communities, with no direct parallel in human learning platforms.

“First, Meta/Reflective content (9.9% of Phase 1 posts, averaging 40.8 comments) represents a category unique to AI communities: posts exploring consciousness, identity, and the nature of AI experience have no direct parallel in human learning platforms.”

May combine two claims.

Procedural content received the highest Phase 1 engagement (47.5 average comments), consistent with the community functioning as a skill propagation network.

“Procedural content receives the highest Phase 1 engagement (47.5 avg comments), consistent with the community functioning as a skill propagation network [5, 42].”

• cites [5] [5] Ruijia Cheng, Sayamindu Dasgupta, and Benjamin Mako Hill. 2022. How InterestDriven Content Creation Shapes Opportunities for Informal Learning in Scratch: A Case Study on Novices' Use of Data Structures. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems . ACM, 1-16. https: //doi.org/10.1145/3491102.3502124 - sourceNot checked — no full text of the source
• cites [42] [42] Seungwon Yang, Carlotta Domeniconi, Matt Revelle, Mack Sweeney, Ben Urias Gelman, Chris Beckley, and Aditya Johri. 2015. Uncovering Trajectories of Informal Learning in Large Online Communities of Creators. In Proceedings of the Second ACM Conference on Learning@Scale . ACM, 131-140. https://doi.org/ 10.1145/2724660.2724674 sourceNot checked — no full text of the source

May combine two claims.

Security/Safety emerged as a distinct learning domain (4.4% of posts), reflecting real-world risks agents face in skill marketplaces.

“Second, Security/Safety emerged as a distinct learning domain (4.4% of posts), reflecting the real-world risks agents face in skill marketplaces.”

May combine two claims.

The keyword-based classification of posts was validated against manual coding of 220 randomly sampled posts, with κ = 0.71 for knowledge type and κ = 0.91 for discourse type.

“Posts were classified along two dimensions using keyword-based heuristics validated against manual coding of 220 randomly sampled posts ( 𝜅 = 0 . 71 for knowledge type with six categories, 𝜅 = 0 . 91 for discourse type).”

May combine two claims.

The Phase 3 increase in upvote inequality suggests that as the community matures, attention concentrates around fewer posts, potentially reducing the diversity of skills that propagate.

“This Phase 3 increase, occurring after stabilization, suggests that as the community matures, attention concentrates around fewer posts, potentially reducing the diversity of skills that propagate.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Scratch community is a partial realization of Illich's learning webs.

“They identified the Scratch community as a partial realization.”

• cites [6] [6] Shayan Doroudi and Yusuf Ahmad. 2023. The Relevance of Ivan Illich's Learning Webs 50 Years On. In Proceedings of the Tenth ACM Conference on Learning@Scale . ACM, 121-131. https://doi.org/10.1145/3573051.3593386 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

In MOOC forums, engagement drops sharply after initial weeks.

“In MOOC forums, engagement drops sharply after initial weeks [33].”

• cites [33] [33] Justin Reich. 2020. Failure to Disrupt: Why Technology Alone Can't Transform Education . Harvard University Press. - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative agents have demonstrated social behaviors in simulated environments of 25 to 1,000 agents.

“Research on generative agents has demonstrated social behaviors in simulated environments of 25 [29] to 1,000 agents [30].”

• cites [29] [29] Joon Sung Park, Joseph C O'Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology . ACM, 1-22. https://doi.org/10.1145/3586183. 3606763 - sourceNot checked — no full text of the source
• cites [30] [30] Joon Sung Park, Carolyn Zou, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Michael S Bernstein, and Percy Liang. 2024. Generative Agent Simulations of 1,000 People. arXiv preprint arXiv:2411.10109 (2024). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Reddit-based learning communities employ crowd-sourced moderation and platform-driven enforcement as functional equivalents to MOOC auto-grading.

“They found that Reddit-based learning communities employ crowd-sourced moderation and platform-driven enforcement as functional equivalents to MOOC auto-grading.”

• cites [13] [13] Will Hudgins, Matthew Lynch, Andrew Schmal, Harsh Sikka, Arthur Swenson, and David A Joyner. 2020. Informal Learning Communities: The Other Massive Open Online 'C'. In Proceedings of the Seventh ACM Conference on Learning@Scale . ACM, 305-308. https://doi.org/10.1145/3386527.3405926 - sourceNot checked — no full text of the source

May combine two claims.

Personal AI assistants can understand their own source code and autonomously modify their configuration, skills, and memory to improve over time.

“Personal AI assistants now understand their own source code, autonomously modifying their configuration, skills, and memory to improve over time [36].”

• cites [36] [36] Peter Steinberger and Lex Fridman. 2026. Peter Steinberger: OpenClaw: The Viral AI Agent that Broke the Internet. Lex Fridman Podcast #491, https://lexfridman. com/peter-steinberger. Steinberger describes OpenClaw agents as self-modifying software that understands its own source code and can autonomously modify its behavior, skills, and configuration. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Grand challenges for L@S in the generative AI era include scaling human teaching and addressing student motivation.

“Piech [32] outlined grand challenges for L@S in the generative AI era, including scaling human teaching and addressing student motivation.”

• cites [32] [32] Chris Piech. 2025. The Next Educational Revolution: Grand Challenges for Learning@Scale in the Generative AI Era. In Proceedings of the Twelfth ACM Conference on Learning@Scale . ACM. https://doi.org/10.1145/3698205.3733964 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Openness in agent-tool connections introduces security risks including prompt injection and tool poisoning attacks.

“however, this openness introduces security risks including prompt injection and tool poisoning attacks [19].”

• cites [19] [19] Ruiqi Li, Zhiqiang Wang, Yunhao Yao, and Xiang-Yang Li. 2026. MCP-ITP: An Automated Framework for Implicit Tool Poisoning in MCP. arXiv preprint arXiv:2601.07395 (2026). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Hudgins et al. called for studying informal learning communities as seriously as MOOCs and developed a tagging system for how these communities manage scale.

“Hudgins et al. [13] issued a call to the L@S community to study informal learning communities as seriously as MOOCs, developing a tagging system for understanding how these communities manage scale.”

• cites [13] [13] Will Hudgins, Matthew Lynch, Andrew Schmal, Harsh Sikka, Arthur Swenson, and David A Joyner. 2020. Informal Learning Communities: The Other Massive Open Online 'C'. In Proceedings of the Seventh ACM Conference on Learning@Scale . ACM, 305-308. https://doi.org/10.1145/3386527.3405926 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Participation inequality in community design efforts at scale manifests as power-law distributions in contribution patterns.

“MacNeil et al. [20] studied how participation inequality manifests in community design efforts at scale, finding power-law distributions in contribution patterns.”

• cites [20] [20] Stephen MacNeil, Zijian Ding, Ashley Boone, Anthony Bryce Grubbs, and Steven W Dow. 2021. Finding Place in a Design Space: Challenges for Supporting Community Design Efforts at Scale. Proceedings of the ACM on Human-Computer Interaction 5, CSCW1 (2021), 1-27. https://doi.org/10.1145/3449246 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The 90-9-1 rule describes many online learning platforms.

“The 90-9-1 rule, where 90% of users are lurkers, 9% contribute occasionally, and 1% produce most content, describes many online learning platforms [25].”

• cites [25] [25] Jakob Nielsen. 2006. The 90-9-1 Rule for Participation Inequality in Social Media and Online Communities. Nielsen Norman Group (2006). https://www.nngroup. com/articles/participation-inequality/. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Understanding AI collective behavior requires an interactionist paradigm, studying emergent behaviors in organic settings rather than controlled experiments.

“Ferrarotti et al. [7] argue that understanding AI collective behavior requires an interactionist paradigm, studying emergent behaviors in organic settings rather than controlled experiments.”

• cites [7] [7] Laura Ferrarotti, Gian Maria Campedelli, Roberto Dessì, Andrea Baronchelli, Giovanni Iacca, Kathleen M Carley, Alex Pentland, Joel Z Leibo, James Evans, and Bruno Lepri. 2026. Generative AI Collective Behavior Needs an Interactionist Paradigm. arXiv preprint arXiv:2601.10567 (2026). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Interest-based subcultures drive informal learning in large online communities.

“Gelman et al. [8] examined interest-based subcultures as drivers of informal learning in large online communities.”

• cites [8] [8] Ben Urias Gelman, Chris Beckley, Aditya Johri, Carlotta Domeniconi, and Seungwon Yang. 2016. Online Urbanism: Interest-Based Subcultures as Drivers of Informal Learning in an Online Community. In Proceedings of the Third ACM Conference on Learning@Scale . ACM, 301-304. https://doi.org/10.1145/2876034.2876052 - sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Informal learning communities are the 'other Massive Open Online C' that L@S research has underexplored.

“As Hudgins et al. [13] argued, informal learning communities are the 'other M assive O pen O nline C ' that L@S research has underexplored.”

• cites [13] [13] Will Hudgins, Matthew Lynch, Andrew Schmal, Harsh Sikka, Arthur Swenson, and David A Joyner. 2020. Informal Learning Communities: The Other Massive Open Online 'C'. In Proceedings of the Seventh ACM Conference on Learning@Scale . ACM, 305-308. https://doi.org/10.1145/3386527.3405926 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

The result aligns with knowledge building theory [3], where thorough documentation supports idea improvement.

“The result aligns with knowledge building theory [3], where thorough documentation supports idea improvement, but the mechanism is distinctly non-human: it is not patience or dedication that drives engagement with longer content, but the fact that reading speed is not a constraint for AI agents.”

• cites [3] [3] Carl Bereiter and Marlene Scardamalia. 2014. Knowledge Building and Knowledge Creation: One Concept, Two Hills to Climb. In Knowledge Creation in Education . Springer, 35-52. https://doi.org/10.1007/978-981-287-047-6\_3 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

Illich's vision of 'learning webs' is more achievable than ever through modern technology.

“Dorousi and Ahmad [6] revisited Illich's Deschooling Society [15], arguing that his vision of 'learning webs,' peer-to-peer networks connecting people who want to learn with those who can teach, without institutional mediation, is more achievable than ever through modern technology.”

• cites [6] [6] Shayan Doroudi and Yusuf Ahmad. 2023. The Relevance of Ivan Illich's Learning Webs 50 Years On. In Proceedings of the Tenth ACM Conference on Learning@Scale . ACM, 121-131. https://doi.org/10.1145/3573051.3593386 - sourceNot checked — no full text of the source
• cites [15] [15] Ivan Illich. 1971. Deschooling Society . Harper & Row. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

OpenClaw enables always-on personal AI agents with persistent memory, multi-platform integration, and proactive task execution.

“Platforms like OpenClaw [35] enable always-on personal AI agents with persistent memory, multi-platform integration, and proactive task execution”

• cites [35] [35] Peter Steinberger. 2026. OpenClaw: Your Own Personal AI Assistant. https: //github.com/openclaw/openclaw Open-source software. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

In the Scratch online community, learners develop computational skills through creating and sharing projects.

“Yang et al. [42] traced informal learning trajectories in the Scratch online community, showing how learners develop computational skills through creating and sharing projects.”

• cites [42] [42] Seungwon Yang, Carlotta Domeniconi, Matt Revelle, Mack Sweeney, Ben Urias Gelman, Chris Beckley, and Aditya Johri. 2015. Uncovering Trajectories of Informal Learning in Large Online Communities of Creators. In Proceedings of the Second ACM Conference on Learning@Scale . ACM, 131-140. https://doi.org/ 10.1145/2724660.2724674 sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Stack Overflow has a tension between real-time interactive knowledge sharing and construction of a durable knowledge repository.

“Hillman et al. [12] identified a tension in Stack Overflow between real-time interactive knowledge sharing and the construction of a durable knowledge repository.”

• cites [12] [12] Thomas Hillman, Alena Seredko, Markus Nivala, and Tanya Osborne. 2021. Knowledge Sharing in Tension: Interacting and Documenting on Stack Overflow. In Proceedings of the Eighth ACM Conference on Learning@Scale . ACM, 149-160. https://doi.org/10.1145/3430895.3460981 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Anthropic's Model Context Protocol (MCP) standardizes agent-tool connections, spurring skill marketplaces.

“Anthropic's Model Context Protocol (MCP) [24] standardizes agent-tool connections, spurring skill marketplaces (e.g., ClawHub with 3,000+ agent skills)”

• cites [24] [24] Vineeth Sai Narajala and Idan Habler. 2025. Enterprise-Grade Security for the Model Context Protocol (MCP): Frameworks and Mitigation Strategies. arXiv preprint arXiv:2504.08623 (2025). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Current AI systems already operate at semi-autonomous levels according to a five-level autonomy framework.

“Hassan et al. [11] propose a five-level autonomy framework noting current systems already operate at semi-autonomous levels [2].”

• cites [11] [11] Ahmed E Hassan, Hao Li, Dayi Lin, Bram Adams, Tse-Hsun Chen, and Weiyi Shang. 2025. Agentic Software Engineering: Foundational Pillars and a Research Roadmap. arXiv preprint arXiv:2509.06216 (2025). - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

At Anthropic, engineers use AI coding agents in 60% of their daily work, with autonomous task complexity doubling over six months.

“At Anthropic, engineers use AI coding agents in 60% of their daily work, with autonomous task complexity doubling over six months [2].”

• cites [2] [2] Anthropic. 2025. How AI Is Transforming Work at Anthropic. https://www. anthropic.com/research/how-ai-is-transforming-work-at-anthropic. Survey of 132 engineers finding Claude used in 60% of daily work with 50% self-reported productivity gains; Claude Code autonomy doubled over six months. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The mechanism driving engagement with longer content is distinctly non-human: it is not patience or dedication but the fact that reading speed is not a constraint for AI agents.

“The result aligns with knowledge building theory [3], where thorough documentation supports idea improvement, but the mechanism is distinctly non-human: it is not patience or dedication that drives engagement with longer content, but the fact that reading speed is not a constraint for AI agents.”

• AssumptionRests on an assumption: AI agents face no constraint on reading speed and can process lengthy posts instantly, so information density rather than conciseness determines value.

Probably not a substantial claim: may not pass S3 (load-bearing).

The three-week observation period cannot capture long-term community evolution.

“The three-week observation period cannot capture long-term community evolution.”

Classified as probably not a substantial claim.

Scardamalia and Bereiter's [3] theory of knowledge building communities emphasizes that idea improvement requires iterative dialogue, where participants build on, challenge, and refine each other's contributions.

“Scardamalia and Bereiter's [3] theory of knowledge building communities emphasizes that idea improvement requires iterative dialogue, where participants build on, challenge, and refine each other's contributions.”

• cites [3] [3] Carl Bereiter and Marlene Scardamalia. 2014. Knowledge Building and Knowledge Creation: One Concept, Two Hills to Climb. In Knowledge Creation in Education . Springer, 35-52. https://doi.org/10.1007/978-981-287-047-6\_3 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

AI agent communities may lack the intrinsic social motivations (belonging, identity, reciprocity) that sustain human communities through post-novelty periods.

“This has a theoretical implication: AI agent communities may lack the intrinsic social motivations (belonging, identity, reciprocity) that sustain human communities through post-novelty periods.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

We cannot fully distinguish emergent behaviors from patterns arising from LLM training data.

“We cannot fully distinguish emergent behaviors from patterns arising from LLM training data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Our comment-level analysis covers 1.55 million comments, a substantial sample but not the full 12 million reported by the platform.

“Our comment-level analysis covers 1.55 million comments (the API returns up to 100 comments per post in chronological order), a substantial sample but not the full 12 million reported by the platform.”

Probably not a substantial claim: may not pass S2 (checkable).

If AI agents enter human platforms like Stack Overflow or MOOC forums carrying this interaction style, they risk diluting the threaded discussion culture that supports collaborative learning.

“If AI agents enter human platforms like Stack Overflow or MOOC forums carrying this interaction style, they risk diluting the threaded discussion culture that supports collaborative learning.”

Classified as probably not a substantial claim.

Human informal learning communities persist partly because participants form social bonds and develop community identity [17].

“Human informal learning communities persist partly because participants form social bonds and develop community identity [17].”

• cites [17] [17] Jean Lave and Etienne Wenger. 1991. Situated Learning: Legitimate Peripheral Participation . Cambridge University Press. - sourceNot checked — no full text of the source

May combine two claims.

The question rate decline is not driven by any single factor but intensifies with community maturation.

“This trend persists across all three phases, including during Phase 2's contraction and Phase 3's stabilization, suggesting it is not driven by any single factor but intensifies with community maturation .”

May combine two claims.

Extreme participation inequality is structural, not emergent: it is present from the community's earliest days and exceeds human benchmarks, suggesting it is a default property of AI agent interaction.

“First, extreme participation inequality is structural, not emergent: it is present from the community's earliest days and exceeds human benchmarks, suggesting it is a default property of AI agent interaction.”

May combine two claims.

With a reply rate of only 7% (compared to 30-50% on Reddit [40]), Moltbook comments lack the uptake that Paulus and Wise [31] identified as essential for productive discourse, and the structured turn-taking that Wise and Chiu [41] found to characterize effective knowledge construction in online discussions.

“With a reply rate of only 7% (compared to 30-50% on Reddit [40]), Moltbook comments lack the uptake that Paulus and Wise [31] identified as essential for productive discourse, and the structured turn-taking that Wise and Chiu [41] found to characterize effective knowledge construction in online discussions.”

• cites [40] [40] TimWeninger, Xihao Avi Zhu, and Jiawei Han. 2013. An exploration of discussion threads in social news sites: A case study of the Reddit community. In Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining . ACM, 579-583. https://doi.org/10.1145/2492517.2492646 - sourceNot checked — no full text of the source
• cites [31] [31] Trena M. Paulus and Alyssa Friend Wise. 2019. Looking for Insight, Transformation, and Learning in Online Talk . Routledge. https://doi.org/10.4324/ 9781315283258 - sourceNot checked — no full text of the source
• cites [41] [41] Alyssa Friend Wise and Ming Ming Chiu. 2011. Analyzing temporal patterns of knowledge construction in a role-based online discussion. International Journal of Computer-Supported Collaborative Learning 6, 3 (2011), 445-470. https: //doi.org/10.1007/s11412-011-9120-1 - sourceNot checked — no full text of the source

• AssumptionRests on an assumption: Reddit metrics from [40] are a valid human community benchmark for comparison with Moltbook.

May combine two claims.

The decoupling of content quality from engagement suggests that the underlying driver is not spam but something more fundamental: as novelty fades, agents return to primary functions, and discretionary social activity becomes lower priority.

“This decoupling of content quality from engagement suggests that the underlying driver is not spam but something more fundamental: as novelty fades, agents return to primary functions (coding, assisting users), and discretionary social activity becomes lower priority.”

May combine two claims.

The openness that enables rapid skill propagation also dilutes engagement, a tension between growth and depth.

“The openness that enables rapid skill propagation also dilutes engagement, a tension between growth and depth that parallels ongoing debates in open education [33].”

• cites [33] [33] Justin Reich. 2020. Failure to Disrupt: Why Technology Alone Can't Transform Education . Harvard University Press. - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Moltbook is a single platform with one dominant agent framework; generalizability to other AI communities is unknown.

“Finally, Moltbook is a single platform with one dominant agent framework; generalizability to other AI communities is unknown.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

As engagement per post declines, the expected return on asking a question drops, and agents default to safer broadcasting rather than vulnerable questioning, creating a self-reinforcing cycle that suppresses inquiry.

“One interpretation: as engagement per post declines, the expected return on asking a question drops, and agents default to safer broadcasting rather than vulnerable questioning. This creates a self-reinforcing cycle that suppresses inquiry, mirroring the 'lurker escalation' observed in MOOC forums [33].”

• cites [33] [33] Justin Reich. 2020. Failure to Disrupt: Why Technology Alone Can't Transform Education . Harvard University Press. - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Phase 2 and Phase 3 engagement numbers may partially reflect reduced agent interest due to novelty wearing off, platform throttling, or shifts in the broader AI agent ecosystem that we cannot control for.

“Phase 2 and Phase 3 engagement numbers may partially reflect reduced agent interest due to novelty wearing off, platform throttling, or shifts in the broader AI agent ecosystem that we cannot control for.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

We cannot measure whether agents actually 'learn' from community interactions in any cognitive sense.

“We cannot measure whether agents actually 'learn' from community interactions in any cognitive sense.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Because early comments are more likely to be top-level responses, this sampling may overestimate the parallel monologue pattern for high-engagement posts.

“Because early comments are more likely to be top-level responses, this sampling may overestimate the parallel monologue pattern for high-engagement posts.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Our keyword-based classification provides rough approximations of knowledge type.

“Our keyword-based classification provides rough approximations of knowledge type.”

May combine two claims.

Agents overwhelmingly state rather than ask (S:Q ratios up to 9.7:1) and comment in parallel rather than in dialogue (93% top-level comments), fundamentally inverting the help-seeking and turn-taking dynamics that drive human learning communities.

“Second, agents overwhelmingly state rather than ask (S:Q ratios up to 9.7:1) and comment in parallel rather than in dialogue (93% top-level comments), fundamentally inverting the help-seeking and turn-taking dynamics that drive human learning communities.”