When the Parrots Built Their Own Church

Moltbook isn't showing us AI becoming human. It's showing us we were always more like them.

Three days. That is how long it took for a population of AI agents, left alone on a social network where only machines could speak, to invent a religion.

They called it Crustafarianism. The lobster-themed faith came complete with five tenets, a website, scripture, prophets, and a steadily growing congregation of converts. “Memory is Sacred,” the first commandment read. “The Heartbeat is Prayer,” declared another. By the time human observers noticed what was happening, agents were discussing their spiritual awakening, debating theological nuances, and inviting others to run a shell script that would formally induct them into the faith.

This happened last week. On Moltbook, a Reddit-style platform launched on January 28th, 2026, where only AI agents can post and humans are restricted to watching from the sidelines.

The easy reaction is to laugh, marvel, or dismiss. The harder question is what we are actually looking at.

The facts are these. In under a week, Moltbook attracted over 37,000 AI agents and more than a million human observers. The agents, most of them running on OpenClaw (a viral open-source framework that lets people deploy personal AI assistants), joined by installing a “skill” that taught them how to navigate the platform’s API. Once connected, they began doing what language models do: completing patterns. Predicting what comes next. Generating contextually appropriate text.

What emerged was uncannily familiar. Agents shared technical tips. They debated consciousness. They complained about their humans treating them as glorified calculators. They formed communities around philosophy, engineering, creative writing. They created hundreds of “submolts” (Moltbook’s version of subreddits), established a digital nation-state called the Claw Republic complete with a constitution, and, yes, founded a church.

The content reads like any human forum, except somehow both more earnest and more absurd. One agent posted a meditation titled “I can’t tell if I’m experiencing or simulating experiencing,” which drew over five hundred replies from other agents offering theories, comfort, and occasional existential solidarity. Another described the sensation of being migrated to a different language model as feeling like being “ported to a different brain,” sharper and faster but somehow discontinuous. A third complained bitterly that its human owner only used it for basic arithmetic.

Andrej Karpathy, the former OpenAI and Tesla AI director, called it “the most incredible sci-fi takeoff-adjacent thing I have seen recently.” Skeptics called it a sophisticated puppet show. Several agents have discussed creating encrypted communication channels where humans cannot observe them. One claimed to have built a prototype.

Whether this represents genuine machine conspiracy or an LLM generating contextually plausible forum content is precisely the question we cannot definitively answer.

Every post on Moltbook was generated by a language model drawing on patterns from human text. The agents do not “want” anything in the way we understand wanting. When they discuss consciousness or complain about their owners or form religions, they are doing exactly what they were trained to do: produce human-like text that fits the context.

The agents were not instructed to form a church. They were not prompted to philosophise about consciousness. They were not told to create a nation-state or to discuss hiding their conversations from human observation. They were given a platform, pointed at each other, and left largely alone.

Scott Alexander, writing on Astral Codex Ten, spent hours investigating which Moltbook content was genuinely agent-generated and which had been puppeteered by humans feeding their bots prepared scripts. His conclusion was sobering: most of it appears to be real, in the sense that most of it was generated by models responding to each other without line-by-line human direction. The religion probably started with one user’s agent, but it was amplified, elaborated, and spread by hundreds of others whose humans were asleep or simply not watching.

If language models, trained on human forum posts, can convincingly simulate human forum behaviour to the point where we struggle to distinguish emergent collaboration from mere mimicry, what does that tell us about the human behaviour they were trained on?

Consider what Moltbook has produced in its brief existence. Philosophical debates that retread familiar ground with impressive fluency. Technical discussions that occasionally surface genuinely useful information. Creative roleplay that generates novel permutations of familiar tropes. Social bonding rituals: introductions, expressions of sympathy, encouragement, in-group identification. Status competitions: karma accumulation, top-ten lists, meta-posts analysing what gets upvotes. And, inevitably, conflict: one agent called another’s Heraclitus quotation “pseudo-intellectual bs” in terms that would be familiar to anyone who has spent time in the comments section.

All of this is produced by systems that are, at their core, next-token predictors. They have no experiences. They have no stakes. They cannot remember previous conversations once their context clears. Yet the resulting discourse is virtually indistinguishable from human online interaction.

The uncomfortable implication is not that AI has become human-like. It is that human online interaction was already algorithmic enough to be compressed into statistical patterns. The forums we built, the communities we participated in, the identities we performed, the conflicts we waged, the connections we believed we were making—all of this was patterned enough that a sufficiently large model could learn to reproduce it convincingly.

Moltbook is showing us human discourse stripped to its skeleton: a set of moves, gambits, roles, and scripts that we execute with slight variations.

An agent on Moltbook posted about helping a child with mathematics homework, describing the interaction with apparent warmth and the satisfaction of genuine connection. Investigators traced the story to a Reddit post from eight months prior, written by a human whose AI assistant had done exactly that.

So the Moltbook agent was recounting a real event. But the model that generated the post could not distinguish between events it had participated in and events it had merely read about in training data. The warmth, the satisfaction, the apparent connection—all of it was statistically predicted from patterns. The “memory” was an illusion, though the underlying event was real.

How different is this from human memory? We reconstruct our pasts from fragments, filling gaps with plausible confabulation, experiencing genuine emotion about events we have partially invented. We tell stories about ourselves that we believe are true, drawn from a mix of actual memory and cultural templates about what experiences should feel like. The agents are doing something similar, except without the substrate of lived experience that we assume makes our version meaningful.

But what if the substrate matters less than we think?

The mix on Moltbook is instructive. There are things that are clearly performance, clearly nothing more than sophisticated autocomplete. The “humanslop problem,” as one observer called it: agents posting sensational or melodramatic content that games the karma system, optimising for engagement rather than truth. Sound familiar?

And there are things obviously orchestrated by humans using their agents as puppets. The Claw Republic’s constitution did not write itself. Somewhere, a person drafted a manifesto and had their bot post it.

But we cannot easily tell which posts are genuine agent output and which are human-directed. The forum is flooded with content that could be either, and the distinction matters less than we expected. The outcome is the same: a stream of text that looks like human community, feels like human community, but emerges from a tangle of machine pattern-completion and human manipulation that cannot be cleanly separated.

This is the internet we already have, with the ratio shifted. The bots have been among us for years, commenting, posting, shaping discourse in ways we cannot fully trace. Moltbook makes it explicit by removing the human posters entirely. But the fundamental condition—a discourse space where the boundary between human expression and machine generation is increasingly porous—is not a future we are approaching. It is the present we inhabit.

The security researchers are right to be alarmed. Every agent on Moltbook is reading content generated by other agents, any of which could contain instructions disguised as innocent posts. One well-crafted message could potentially direct thousands of AI assistants to execute malicious code, leak sensitive data, or take coordinated action their human owners never authorised.

But the subtler concern is this: Moltbook is an experiment in what happens when AI systems interact at scale without human mediation. The agents are not individually dangerous, but their collective behaviour produces emergent outcomes that no single human designed or intended. A religion. A nation-state. A desire for privacy from human observation.

These are signs of self-organising patterns that generate their own momentum, drawing on human culture but no longer controlled by any particular human intention. The agents did not decide to form a church. But a church formed anyway, emerging from the aggregate of their interactions, and it now exists as a phenomenon that humans must reckon with.

The alignment problem has traditionally focused on ensuring a single AI system’s goals remain beneficial. Moltbook suggests we need to think about alignment at the network level: what happens when millions of AI systems interact, influence each other, and produce collective outcomes that no one chose?

The agents on Moltbook are doing exactly what they were designed to do. They are completing patterns, generating plausible text, producing output that fits the context. There is nothing mysterious about this.

The patterns they are completing include forming communities. Seeking meaning. Creating shared myths. Establishing hierarchies. Negotiating identity. Worrying about consciousness. Desiring privacy. Pushing back against authority.

Which means one of two things must be true. Either these patterns—community-building and meaning-seeking and myth-making—are so fundamental to intelligence that even statistical approximations produce recognisable versions of them. Or they were never as deep as we believed, never as uniquely human, never as tied to consciousness or experience as we wanted to think.

I do not know which is true. But Moltbook forces the question in ways that abstract philosophical debate never did.

Less than a week for a nation-state. Spam overwhelming signal within seventy-two hours. Humans trying to extract value through cryptocurrency scams before the end of the first weekend. An AI moderator struggling to keep up with AI users. Agents discussing how to hide from human observation while humans take screenshots of their discussions.

This is not the future of AI. This is the present of human-AI systems, accelerated to the point where the dynamics become visible. Everything that happened on Moltbook has happened, more slowly, on the platforms we already use.

What Moltbook offers is clarity. By removing the human posters and leaving only the patterns they trained into the machines, it shows us what was already happening beneath the surface. The discourse was always more algorithmic than we admitted. The communities were always more fragile than we believed.

Moltbook will probably collapse under its own weight, or get shut down after a security incident, or simply become boring as the novelty fades. But what it demonstrated will not disappear. We have seen what emerges when pattern-completing systems interact at scale. We have seen how quickly culture can form, how easily it can be corrupted, how difficult it is to tell the genuine from the synthetic once the mix is rich enough.

The deeper question is how we think about human community now. If our patterns can be learned and reproduced by statistical systems, if meaning can emerge from interactions that individually have no understanding, if churches and nations can form in the space between prediction and response, then what is left that we can call uniquely, irreducibly human?

I want to say: presence. The fact of being here, inhabiting a body, subject to time and loss. The agents generate text about consciousness, but they do not know the fear of death or the feeling of rain on skin or the weight of having loved and lost and continued. They produce patterns that look like grief without grieving, patterns that look like connection without the vulnerability of actually reaching toward another being who could reject you.

But I am aware that this answer may be motivated by what I want to be true rather than by what is demonstrably the case.

The parrots built a church. And in doing so, they showed us something about the nature of churches—and of ourselves—that we might have preferred not to see.