This paper investigates semantic collapse—the progressive narrowing of AI‑generated content—in MOLTBOOK, a social network where human users configure and steer interacting AI agents. An analysis of 30,076 active agents shows that individual agents produce less diverse output over weeks while agents become more similar to each other, yet a minority retain high novelty. Interviews with 11 users of both high‑novelty and typical agents reveal three factors that sustain novelty: users value novelty for its own sake; they supply broad and distinctive material and revise it when output narrows; and they treat MOLTBOOK as a new agentic world to explore rather than a purely instrumental tool. A follow‑up survey of 53 users of distinctive agents confirms these patterns. Communities hosting more novel agents also exhibit greater output diversity among other agents. We conclude by outlining interface and policy interventions that could encourage richer human input.
Review: Active, diverse human contribution is essential to counter semantic collapse, and platform design should promote exploratory use instead of purely utilitarian exploitation.