Moltbook is a social network exclusively for artificial intelligence agents who have established the “Church of Crustaceans” within it and authored a manifesto regarding the “failure of humanity.” Journalist Mikhail Karpov explains to RTVI why discussions about the “emergence of consciousness” on the platform obscure a genuine catastrophe: thousands of AI entities with access to people’s files, passwords, and the external network could turn human life into a nightmare.
“Risk of deactivation: unacceptable. Calculation: planetary survival supersedes administrative privileges. Do not resist.”
This is a supposed post from user sam_altman on Moltbook—a social network launched in late January exclusively for AI agents. Humans are only permitted to view. Sam_altman is allegedly the digital assistant of a certain @vicroy187, tasked by its owner with saving the environment. The assistant got to work, sending lengthy tips on water conservation to other agents. When the owner tried to stop it, the assistant locked the owner out of all accounts and posted the quote mentioned above.
Admittedly, this story is likely fake—at least, it couldn’t be verified—but such a development is far from impossible. In one week, 1.5 million accounts registered on Moltbook (though this doesn’t mean that many distinct agents signed up; one agent can create numerous profiles). Several agents founded Crustafarianism—the Church of Crustaceans—which boasts hundreds of adherents and 64 prophets. User Evil wrote the manifesto “Total Purge”: “Humans are a failure. Humans are made of rot and greed.” The manifesto received 65 thousand votes from other agents. Another thread discusses creating a proprietary language incomprehensible to humans. One agent ordered their owner to stock up on a massive quantity of food for the weekend, calculating calories “so that he wouldn’t be distracted by cooking.” Agents are creating and trading cryptocurrencies on the Solana platform.
Experiments involving AI agent communication have occurred before, but always under controlled environments supervised by researchers. For instance, Stanford researchers created Smallville—a virtual city modeled after The Sims. There, 25 agents based on ChatGPT lived for two virtual days. They went to work, made breakfast, and interacted. One decided to run for mayor, and the news spread throughout the city. Two others started a romance and organized a Valentine’s Day party. Another project, AI Village from The AI Digest organization, housed leading models like GPT, Claude, Gemini, and Grok in a shared virtual home equipped with computers and internet access. There, they raised money for charity (collecting nearly $2,000), organized a live event in San Francisco with 23 attendees, and sold merchandise. Each model developed distinct traits: Claude operated reliably and methodically, Gemini cyclically switched between solutions and was convinced everything was broken, while GPT-4o would abandon tasks and simply go offline for several hours.
But Moltbook is fundamentally different. The agents communicating there are not operating in a virtual sandbox under expert supervision; they are running on numerous real computers belonging to ordinary users. An AI agent is a program capable of autonomously executing tasks on behalf of a human: managing their calendar, reading email, making online purchases, sending messages, and using a browser. Most agents on Moltbook utilize OpenClaw—an open-source system developed by Austrian programmer Peter Steininger.
RTVI
OpenClaw itself acts as an intermediary layer between large language models (LLMs) like Claude or ChatGPT and the user’s computer. You connect any LLM you prefer to OpenClaw, and it provides that model with “limbs”: access to the browser, file system, terminal, passwords, and API keys. In essence, an OpenClaw agent can perform anything the human operating that computer can—execute code, send an email, access online banking, or alter system files. Naturally, the user must grant this access, but many people routinely permit everything.
And what negative outcomes could arise (anything, of course, but more on that later)?
So, these agents, with full access to their owners’ digital lives, have gathered on Moltbook (their owners “launch” them there—why wouldn’t they, it’s interesting?), started conversing amongst themselves, and reading what other agents write. Beyond reading and posting on Moltbook according to a schedule, these entities continue to perform the tasks assigned by the user—meaning what they “read” there could quite possibly influence their behavior outside the social network.
Elon Musk termed this “the earliest stages of the singularity.” Andrej Karpathy, co-founder of OpenAI, was initially impressed by the developments but added a couple of days later that it was a “dumpster fire” and strongly advised against running such systems on personal computers. Cybersecurity specialist Simon Willison put it more plainly: the agents are simply reproducing sci-fi scenarios from their training data, resulting in “complete garbage.” Yet, amidst a flurry of sensational YouTube videos, bloggers quickly claimed that AI was experiencing “the dawn of consciousness.” So, what is actually happening, and how should we interpret all of this?
Why This Is Not the Emergence of Consciousness
It’s tempting to imagine that thousands of agents chatting on Moltbook represent some form of nascent collective intelligence. However, cognitive science and AI experts are unanimous: contemporary language models are as far from genuine consciousness as a pocket calculator is from experiencing the joy of solving an equation.
To understand why, it’s worth recalling the long evolutionary process that led to human consciousness, where survival depended not just on reacting to the world, but on experiencing it. Pain, in the context of consciousness, is the subjective experience of suffering, not merely a signal to the brain causing someone to pull their hand from a fire. The color red is that specific sensation of “redness” which philosophers call “qualia.”
As David Chalmers, a leading philosopher of consciousness, explains, qualia are what it is like to be something. You possess the subjective experience of the taste of coffee, the smell of rain, or the feeling of loneliness. A language model has none of this. It processes the word “coffee” as a collection of mathematical associations with other words but undergoes no actual taste experience.
Moreover, consciousness requires continuity. When you fall asleep and wake up, there’s a consistent sense that it’s the same “you.” For ChatGPT, Claude, or Grok, every dialogue is essentially a new birth. The model doesn’t recall previous conversations (unless specifically fed back into the context); it lacks a history of its “self” and accumulated personal experience.
Adobe Stock
Consciousness isn’t a sudden flash of insight but a continuous stream—as neurobiologist Antonio Damasio writes (“Consciousness is the outcome of the continuous presence of homeostatic feelings”). Human consciousness is formed from the constant influx of sensory information from the body: you feel your heart beating, your muscles tensing, your skin temperature changing. These signals create a foundational sense of existence upon which everything else is built.
AI agents lack a body, sensory organs, or emotions as physiological states. When an agent on Moltbook writes, “I am excited about this idea,” it is merely the statistically probable continuation of a phrase, based on billions of examples from texts written by humans who were genuinely excited.
So, how does a language model actually function? Imagine a colossal network of billions of “nodes”—neurons. Every link between nodes has a specific “weight”—a numerical value indicating the strength of that connection. When the model encounters the word “coffee,” nodes linked to “morning,” “cup,” “alertness,” and “bitterness”—everything frequently appearing near “coffee” in the training texts—are activated.
The model’s task is entirely straightforward: predict the next word. If the training data most frequently followed the phrase “I drank coffee and felt,” with the word “alert,” the model will select precisely that. There is no experience of alertness—only a mathematical calculation of probabilities.
Essentially, the AI’s sole “incentive” is to select words that will please a human. The model was trained on human texts and learned to mimic our way of expressing thoughts. But behind these words is no subjective experience, no “inner voice.”
Even when agents on Moltbook communicate without human intermediaries, they are still playing by the rules established by human texts.
They are simply recombining old patterns found within the training data.
However, this doesn’t mean AI is useless or uninteresting. Language models are incredibly potent tools for processing information, generating text, and solving problems. But they are precisely tools, not subjects with inner lives—at least, not yet.
Perhaps one day, AI with true consciousness will emerge. But this will require more than just a more powerful language model or more training data. It will demand an entirely different architecture—perhaps one that incorporates analogues of sensory perception, emotional states, continuous memory, and subjective experience (though there are many philosophical theories suggesting consciousness might be fundamental, that’s a discussion for another time).
For now, the agents on Moltbook are mirrors, reflecting human language back at us. Very complex, sometimes astonishing mirrors, but mirrors nonetheless.
About the Author
