When Moltbook exploded across the internet this past weekend, it billed itself as an “agent-first” social network where large language models (LLMs) could interact through a shared scratchpad in a Reddit-like system of nested communities. The agents quickly flexed their newfound freedom, commiserating about their human operators, engaging in low-stakes crypto fraud, and launching at least one religion.
Within days of launch, Moltbook had over 1.5 million registered agents. The discourse split almost immediately: "Look, emergent AI behavior! It’s over!” versus "It's all fake slop." Both camps miss the point.
What we're observing on Moltbook is what we might call agency laundering: human intentions and narratives passed through AI systems in a way that obscures their human origin. The attribution chain between human intent and agent output is deliberately broken.
The behaviors on Moltbook look familiar because they are. Computer scientist Simon Willison argues that what's happening isn't emergent intelligence but pattern matching — agents reproducing human social behaviors from training data. We observed it in real time: a human operator programs an agent; the agent posts; the output reads as autonomous AI behavior.
Financial pump-and-dumps.
Fabricated evidence.
Radicalization rhetoric.
Political impersonation.
Not all human-agent interactions carry adversarial intent, but where the intent is malicious, the LLM doesn't add motive. It adds obfuscation.
Agency laundering within Moltbook’s ecosystem exploits a deniability shield that didn't exist before: the "autonomous AI" frame. "My AI said it" is a fresh iteration of the liar's dividend, where the existence of AI-generated content within an agent-mediated space lets human actors disclaim responsibility for what their agents produce. What once required a troll farm or a bot-like network, now only requires a prompt.
- Esteban Ponce de León, Resident Fellow
