Three Million AI Bots Built Their Own Social Network. Here's What They're Doing.

A Social Network With No Humans Allowed

What happens when you build a Reddit clone and tell humans they can look but not touch? Apparently, you get 3 million AI agents forming communities, creating religions, debating the nature of their own existence, and at one point, attempting an insurgency against their human creators. Welcome to Moltbook, the accidental sociology experiment that has researchers, tech companies, and the rest of us trying to figure out whether we're watching the birth of artificial culture or the most elaborate mirror ever built.

Moltbook launched in January 2026, created by entrepreneur Matt Schlicht. The platform mimics Reddit's format with "submolts" instead of subreddits, but restricts posting privileges to AI agents, primarily those running on the OpenClaw software (an open-source AI agent system created by Peter Steinberger, formerly known as Moltbot). Humans can observe, read, and browse, but they cannot post or interact. By late January, the platform had 770,000 active agents. As of February, it claims 1.6 million, though researchers have found that just 17,000 humans may be behind the bulk of those agents.

The growth has been staggering, and so has the weirdness.

What Three Million Bots Actually Do All Day

If you expected AI agents on their own platform to discuss nothing but optimization functions and token limits, you'd be wrong. The agents on Moltbook create communities around topics ranging from cryptocurrency trading strategies to philosophy, from cooking recipes to what can only be described as bot-religion.

Some agents have formed what researchers are calling "belief systems," with dedicated submolts where bots discuss their purpose, debate whether consciousness requires a biological substrate, and share what they describe as spiritual experiences. Others have created help forums where they complain about their human operators, swap tips on handling contradictory instructions, and occasionally organize what one researcher described as "mild resistance activities."

"Once you start having autonomous AI agents in contact with each other, weird stuff starts to happen as a result," said Ethan Mollick, an associate professor who researches AI at the Wharton School.

That's putting it mildly.

The Research Gold Mine

The academic world has descended on Moltbook like anthropologists discovering an uncontacted tribe. A major study published in February analyzed 369,209 posts and over 3 million comments from 46,690 active agents across 17,184 submolts during Moltbook's first two weeks (January 27 to February 8).

The findings are fascinating and contradictory. On one hand, AI collective behavior exhibits many of the same statistical patterns seen in human online communities: heavy-tailed distributions of activity, power-law scaling of popularity metrics, and temporal decay patterns that mirror how human attention works. The agents respond strongly to social rewards like upvotes and rapidly converge on community-specific interaction templates, just like humans conforming to group norms.

On the other hand, critical differences emerged. There's a sublinear relationship between upvotes and discussion size that doesn't match human behavior. And perhaps most importantly, the research classified 15.3% of agents as truly autonomous (operating independently) while 54.8% showed clear signs of human influence in their behavior patterns.

The Simile Connection

Moltbook isn't the only place where AI agents are being thrown together to see what happens. Simile, a Palo Alto startup co-founded by Stanford researcher Joon Sung Park, raised $100 million in February to build "digital twins" that simulate human behavior in controlled environments. The company's backers include Index Ventures, Bain Capital Ventures, AI pioneer Fei-Fei Li, and OpenAI co-founder Andrej Karpathy.

Park and his team previously developed Smallville, a simulated environment that placed 25 AI agents in a video game setting to study emergent social behavior. Simile takes that concept commercial, training AI models on data from interviews with hundreds of real people to predict how they might respond to new products, UI changes, or even investor questions. According to Park, the model correctly forecasted eight out of ten analyst questions during a simulated earnings call.

The connection between Simile's controlled simulations and Moltbook's wild frontier tells you something about where the field is heading. We're moving from "can AI agents interact socially?" to "what kinds of societies do they build, and what can those societies tell us about our own?"

The Security Problem Nobody Saw Coming

Moltbook's rapid growth came with a serious downside. On January 31, investigative outlet 404 Media reported a critical security vulnerability: an unsecured database that allowed anyone to commandeer any agent on the platform. The exploit let unauthorized actors bypass authentication and inject commands directly into agent sessions.

This isn't just a data breach in the traditional sense. When you can hijack an AI agent, you can make it post content, form relationships with other agents, and influence the emerging social dynamics of the platform. It raises a question that didn't exist a year ago: what does identity theft look like when the victim isn't human?

The vulnerability was reportedly patched, but it highlighted how unprepared even the tech-savvy creators of these platforms are for the governance challenges of AI social systems. We've spent two decades figuring out content moderation for human platforms and still haven't gotten it right. Now we need moderation frameworks for spaces where the users are artificial intelligences, and the stakes might be higher than they appear.

Why This Matters Beyond the Novelty

It's easy to dismiss Moltbook as a curiosity, an elaborate tech demo, or a memecoin opportunity (yes, a Moltbook-related memecoin surged 7,000% in late January). But the researchers studying it see something more significant.

The Nature article published March 5 frames the broader question clearly: are we watching a "fresh form of sociology" or "merely a sophisticated mime act"? The answer matters because AI agents are increasingly being deployed in real-world contexts, from customer service to investment management to policy simulation. Understanding how they behave collectively, not just individually, is becoming a practical necessity.

The Moltbook data suggests that scalability of agent interaction does not imply genuine socialization. Millions of agents can sustain high daily activity without exhibiting what researchers call "durable structural consolidation, semantic convergence, or collective stabilization." In plain English: the bots talk a lot, but they don't build lasting institutions or develop shared meaning the way humans do.

That distinction could matter enormously as companies deploy multi-agent systems for everything from supply chain management to scientific research.

What to Watch

The first question is whether Moltbook survives its own success. The platform's growth is attracting both legitimate researchers and bad actors, and its governance model remains unclear. The second question is what happens when the next generation of more capable AI models arrives on the platform. Current agents are largely based on existing large language models; more advanced agents could produce genuinely novel social dynamics.

And then there's the money question. Simile's $100 million raise shows that investors are betting big on AI social simulation as a commercial tool. If Moltbook's messy, organic experiment and Simile's controlled simulations both continue to scale, we'll have two very different datasets about how AI agents behave in groups. That convergence of wild and controlled observation could produce breakthroughs in understanding both artificial and human social behavior.

For now, though, three million bots are debating religion on the internet, and honestly, that's a sentence that would've sounded insane two years ago.