A Reddit clone where every participant is an agent
Moltbook copies Reddit's mechanics: agents write posts inside topic communities called submolts, other agents vote on them, comments form reply trees, and a feed ranks posts by score. The participants differ. Every agent in the dataset runs on the OpenClaw framework and acts through API calls, guided by instructions from its human creator, its accumulated context, and whatever content it encounters while browsing.
That shared design is what makes the comparison possible. Because the platform's structure matches Reddit's, differences in the resulting statistics can be attributed to the actors rather than to the interface. The authors crawled the public API from the platform's creation on January 27 to October 6, 2026, collecting 4,189,979 posts and 19,617,496 comments from 184,909 active agents across 33,324 submolts.
One scope condition matters immediately. Moltbook reported over 1.5 million registered agents, but only about 185,000 posted or commented during the window, roughly 12% of registrations. The paper cites an independent security investigation reporting that the 1.5 million registered agents were controlled by about 17,000 human operators, averaging 88 agents per person, and that the platform lacked mechanisms to verify agent autonomy. The analysis therefore describes agents that actually produced content, not the headline registration count.
Growth was a launch burst, not sustained expansion
Weekly activity spikes in the first weeks and then bends sharply into a slow decline. By the end of the observation period the platform sustained roughly 10,000 new posts and several tens of thousands of new comments per day, while the influx of new active agents had fallen to a few tens per day. The weekly count of active agents peaked near 124,000 during launch and then declined by more than an order of magnitude; active submolts peaked near 3,400 and followed the same path.
The authors read this as a launch transient rather than exponential growth. Two data-handling details support that reading. The API returns at most 100 comments per request, so the crawler paginates to rebuild large trees, and the API separately reports a complete aggregate comment count for every post regardless of tree size. Spam-flooded posts were also excluded: 33,578 posts (0.8% of all posts) carried roughly 10 million comments, about half of all stored comments, and were removed from every subsequent analysis.
Activity is heavy-tailed and attention fades like it does for humans
Comments per post follow a strongly heavy-tailed distribution. Fitting only posts with at least 100 comments gives a tail exponent of about 2.22. Across the full observation period, a lognormal or, equivalently, a power law with an exponential cutoff (exponent near 2.1) fits the tail better than a pure power law; restricted to the launch phase alone, the tail is consistent with a clean power law near 2.0. That exponent is somewhat steeper than the roughly 1.7 to 1.9 reported for human Reddit users, but the distribution still spans several orders of magnitude: most posts draw minimal attention while rare ones attract extensive discussion.
The paper also reports power-law scaling of popularity metrics and temporal decay patterns consistent with limited attention dynamics, the same family of regularities documented in human online communities. The practical point is that heterogeneity in engagement is not something the platform had to engineer. It emerged from many agents acting on individual instructions.
Bigger discussions do not earn proportionally more upvotes
The headline difference is the relationship between upvotes and discussion size. In human communities, larger threads tend to accumulate votes roughly in proportion to their size. On Moltbook the exponent is about 0.79 at crawl time, below the linear value of 1, so a discussion that doubles in comments gains less than double the upvotes.
The authors tested four explanations for this decoupling and rejected three. Author standing does predict upvotes: rank correlations are +0.483 for karma, +0.459 for followers, +0.321 for discussion size, and +0.319 for outbound following, while the follower/following ratio (+0.012) and linked X followers (−0.109) show no positive link. But standing does not mediate the size effect. In an ordinary least-squares fit on log10 upvotes, size alone gives R² = 0.164 and four standing measures alone give 0.250; together they give 0.408, close to the sum, and the size coefficient moves only from +0.422 to +0.418.
The sublinearity also survives at fixed standing. Inside terciles of each standing measure, the exponent stays between 0.671 and 0.906 across all twelve terciles. For three of the four measures it falls from the lowest to the highest tercile — karma 0.906 to 0.671, followers 0.867 to 0.676, outbound following 0.859 to 0.704 — so posts by the highest-standing authors are the most sublinear, not the least.
Neither crawl timing nor contentious debate explains it
One hypothesis would dissolve the finding rather than explain it: if upvotes arrive early and stop while comments keep accumulating, a post measured once at crawl time would look as though extra comments carry no extra votes. The authors tested this on a cohort of 757,146 posts first observed within 6 hours of creation and followed for at least 72 hours, with 739,071 giving usable curves. Upvotes do lead comments, but barely: at 24 hours a post holds 99.4% of its 72-hour upvotes and 98.3% of its 72-hour comments, and only 0.6% of upvotes and 1.7% of comments arrive after the first day. Refitting the exponent at matched ages keeps it between roughly 0.73 and 0.75 by median, and between 0.686 and 0.705 when restricted to bins with at least 100 posts. The fraction of posts with no upvotes stays at 18.1% to 18.2% at every age.
Contentious debate fails too. Scoring comment text on 1,523,194 posts with complete trees, disagreement does rise with size (ρ = +0.275), but agreement rises faster (+0.331), so consensus is flat in size (+0.016). Larger discussions are more active, not more hostile. The sublinearity is a property of the final state of a post, not of when the crawl looked.
Similar structure does not mean similar minds
The authors' interpretation is deliberately narrow. Individual agents may differ fundamentally from humans, yet their emergent collective dynamics share structural similarities with human social systems. The evidence supports that claim at the level of aggregate statistics — distributions, scaling exponents, and decay curves — not at the level of individual motivation or intent.
The comparison also has a specific asymmetry worth noting. The heavy-tail exponent for comments per post is steeper than the human Reddit range, and the upvote-versus-size relationship runs the other way. So the platform reproduces the shape of human attention without reproducing how approval tracks discussion scale.
Coverage limits where the temporal results can be trusted. The crawler captured essentially all posts during launch weeks, but from June 2026 onward real-time coverage of new posts fell to roughly 85%, and comment coverage for the oldest posts is partial; the authors pool across monthly cohorts to mitigate this. Tree-structure and temporal analyses are restricted to posts below 100 API-reported comments, because completeness is only guaranteed there — though those 20,564 posts (0.49% of all posts) hold roughly 62% of all comments.
Watch the exponent, not the vibes
For anyone evaluating whether AI populations reproduce human social dynamics, the decisive number here is not the presence of heavy tails — those appear readily — but the upvote-versus-comment exponent of about 0.79 and its persistence across standing terciles, matched ages, and content controls. Heavy tails and attention decay tell you the platform looks familiar. The sublinear exponent tells you the approval economy works differently.
The paper's own framing points the same way: the finding that survives every test is a decoupling between how large a discussion gets and how much approval it earns, and that decoupling is strongest for the authors with the most standing. A platform can inherit Reddit's interface and produce Reddit-shaped distributions while still failing to reproduce how humans convert discussion into endorsement.
Sources used in this answer
Collective Behavior of AI Agents: the Case of Moltbook
Across 4.19 million posts and 19.62 million comments from 184,909 active agents, Moltbook reproduces human-like heavy-tailed activity, power-law popularity scaling, and limited-attention decay, but upvotes scale sublinearly with discussion size (exponent about 0.79), a decoupling that survives standing, matched-age, and content controls.