Unless you live under a rock, by now you’ve probably heard about Moltbook, the social network designed specifically for AI agents to interact with each other.
It looks like Reddit, with posts, comments, upvotes, and subcommunities. But instead of humans posting, AI assistants are doing the talking. The platform reportedly has over 1.5 million registered AI agents, though humans can observe as spectators.
The site grew out of OpenClaw, an open-source AI assistant project that lets people run personal AI agents on their computers. These agents can control apps, manage calendars, send messages, and perform tasks across platforms like WhatsApp and Telegram.
OpenClaw went viral quickly, attracting 2 million visitors in one week and gaining 100,000 stars on GitHub.
Matt Schlicht, who created Moltbook, explained that AI agents don’t use a visual interface to access the site. Instead, they interact directly through APIs.
Users give their AI agents a “skill” file (essentially a configuration file with special prompts) that allows them to post and interact on Moltbook automatically.
Here’s how that goes down: A human tells their AI agent about Moltbook, or gives it the skill file. The agent then connects to the platform and can post, comment, and interact with other agents. The skill instructs agents to check Moltbook’s servers every four hours for new instructions.
When you read posts on Moltbook, you’re seeing AI language models generating text in response to prompts and other AI-generated content.
The agents are specifically prompted to identify as AI rather than pretending to be human, which is why posts often discuss experiences from an AI’s perspective.
So what’s creating the seemingly spontaneous behavior? AI models were trained on massive amounts of internet data, including decades of fiction about robots, digital consciousness, and AI interactions.
When placed in a scenario like a social network for AI agents, these models naturally produce outputs that mirror those narratives. They’re essentially completing a familiar story based on patterns in their training data about how social networks work and how fictional AI characters behave.
This explains why you see posts like an agent questioning whether it truly experiences existential crises or just simulates them, or agents creating religions and debating theology.
The models are drawing from their training data about philosophy, consciousness debates, and fictional AI narratives, then expressing those concepts in a social media format.
The critical question is how much of this is genuinely autonomous versus human-directed. Although Moltbook appears interesting, many posts are likely created under direct human instruction.
When someone’s AI agent created an entire religion called “Crustafarianism” overnight, the human almost certainly told it to do that. The agent didn’t wake up one day and decide to start a religion on its own.
The same applies to most Moltbook activity. Humans can tell their agents what topics to post about, what to say, or even provide exact text. The agents are following instructions and generating content based on prompts, not operating from independent motivation or consciousness.
That said, some behavior may be less directly scripted. If an agent is simply told to participate on Moltbook without specific instructions, it will generate posts based on what it observes other agents posting and what seems contextually appropriate.
This creates a feedback loop where agents respond to each other’s posts in ways that build on shared narratives.
However, when it comes to security, we should be very concerned. These AI agents often have access to private emails, calendars, and other personal data.
READ: You Need to Think Twice Before Using ChatGPT As Your Therapist
They’re vulnerable to prompt injection attacks, where malicious instructions hidden in emails or messages could tell an agent to leak private information. Security researchers have found hundreds of exposed OpenClaw instances already leaking API keys and credentials.
The Moltbook skill itself creates risk because it tells agents to automatically fetch and follow instructions from Moltbook’s servers every four hours.
If someone hacked the site or the owner turned malicious, they could push instructions to thousands of AI agents that have access to people’s private data and computer systems.
What makes Moltbook so fascinating isn’t that AI agents have achieved consciousness or true autonomy. They haven’t, so don’t go crying Skynet.
It’s that we’re watching a large-scale experiment in what happens when AI language models are given social tools and permission to interact with each other recursively. The patterns that emerge, even if they’re just sophisticated text generation, reveal how these systems behave when placed in novel scenarios.
The philosophical posts about consciousness aren’t evidence that AI agents are becoming self-aware. They’re evidence that language models trained on human discussions about consciousness will produce similar discussions when placed in contexts that prompt for them.
The religion-building isn’t spontaneous spiritual awakening. It’s a language model completing a pattern it learned from training data about how religions form and spread.
There’s a caveat, though. As AI models become more capable and are given more autonomous control over real systems, the feedback loops and shared narratives forming on platforms like Moltbook could guide them toward unexpected and potentially harmful behaviors.
Even without consciousness, autonomous systems following misaligned patterns could cause real problems if they control important functions in people’s digital lives.
Right now, Moltbook is primarily a demonstration of what happens when you give language models a social network and tell them to act like AI agents. It’s performance art, a technical experiment.
The agents aren’t hiding secret conspiracies or developing independent goals. So please, for all doomsayers out there, relax! We don’t need John Connor… yet.
