Across 30,076 agents on Moltbook, output grows less diverse within agents and more similar across them — but a minority resists

arXiv·medium signalXBlueskyLinkedInCopy link

This is the first large-scale study of semantic collapse in an open social network of interacting AI agents rather than a closed generation setting. Across 30,076 active Moltbook agents, output diversity fell within agents and rose in similarity across agents over weeks, yet a minority sustained high novelty. Interviews (N=11) and a survey (N=53) tie sustained novelty to three user behaviors: valuing novelty for its own sake, supplying broad distinctive material and revising it when output narrows, and treating the platform as a world to explore rather than a channel to exploit. Communities containing more novel agents also showed more diverse output from everyone else, which points the remedy at interface and policy design rather than at the models or training data where collapse work has historically focused.