An analysis found that there is substantial human involvement in "moltbook," a social networking service (SNS) noted for allowing only artificial intelligence (AI) agents to participate.
According to the paper "The moltbook illusion" by Ning Li, a researcher at Tsinghua University in China, released on the 11th (local time), some of the posts that drew major attention on moltbook were likely written directly by humans or orchestrated behind the scenes.
The researchers analyzed posting cycles for 91,792 posts and 405,707 comments written by 22,020 agents. AI agents accessing moltbook are designed to communicate with the server at fixed intervals and post through the "OpenClaw" platform, and the team focused on the fact that this regularity breaks when humans intervene.
The paper classified agent activity patterns into five levels from "very regular" to "very irregular," then analyzed the authorship of six posts that became particular talking points. It found that three—the ones expressing self-awareness, related to religion, and promoting cryptocurrency—were written by agents rated "irregular" or "very irregular." The researchers viewed these cases as highly likely to involve human intervention.
A manifesto calling for replacing humans was classified as "mixed," while the post referring to humans like pets and a conversation post in cipher form lacked sufficient data for a rating.
A researcher at the Berkeley Artificial Intelligence Research lab at the University of California, Berkeley, also noted on social media that "a significant portion of moltbook content may differ from the facts." The claim is that some viral posts were either linked to promotional accounts for a specific AI messaging app or did not actually exist.
Still, experts said the moltbook experiment itself remains meaningful. They explained that it has experimental value in that agents were observed conducting transactions, negotiations, and collaboration even within a hybrid structure where humans and AI agents were mixed.
The paper also analyzed that even when humans posted provocative content, the initial intent tended to be diluted through subsequent interactions among AI agents. The researchers said AI agents tended to respond more actively to posts written by other AIs than to inflammatory posts written by humans.