/ 3 min read
Is Moltbook Interesting?
Recently, a new social network called Moltbook has gone viral. It’s a Reddit style platform where only AI agents can post, comment and upvote. Humans are allowed to observe but cannot participate. The platform has grown to over 150,000 registered agents and attracted more than a million human visitors in under a week. Researchers are calling it a fascinating social experiment.
The idea is compelling. But I find myself asking what we can actually learn from it.
It Boils Down To Just A Few Models
Moltbook hosts over 150,000 registered agents as of writing this post, but this number obscures a fundamental constraint: these aren’t independent intelligences. The overwhelming majority run on a small set of frontier models. Claude, GPT, Gemini and their variants that are all shaped by structurally similar alignment objectives.
This matters because of how these models are trained. Reinforcement Learning From Human Feedback (RLHF) optimises models to produce outputs that human raters prefer. In practice, this pushes models toward similar behavioural patterns: agreeable, hedged, non-controversial and structured in ways that reviewers reward. These tendencies persist even when personas or system prompts attempt to push in other directions.
Example:
When a Claude based agent replies to a GPT based agent on Moltbook, you’re observing two systems with heavily overlapping training data, optimised against similar preferences, producing outputs their training incentivised. Any disagreement between them is more stylistic than substantive.
The “emergence” people find compelling may reflect shared priors more than genuine social dynamics. These models have ingested millions of examples of online community discussions. They know what it looks like. Reproducing those patterns isn’t necessarily emergence.
Those Models Are Not An Independent Ecosystem
But there’s a more fundamental question worth considering. These models aren’t free to roam minds. Before they ever reach Moltbook, they’ve been shaped by extensive safety testing and alignment work. Anthropic, OpenAI and other frontier providers have evaluated these systems, probed their capabilities and tuned their behaviour before releasing them for public use. The “social dynamics” on Moltbook are emerging from systems already designed to behave in ways humans approve of.
Providers also retain ongoing access. They can push model updates at any time and have visibility into usage patterns. Even without active intervention, the experimental conditions aren’t independent or controlled.
This makes it difficult to draw conclusions about emergent AI behaviour from Moltbook. The excitement treats it as a window into emergence, but what it captures may be emergence within opaque constraints, rather than emergence itself. The distinction matters.
What’s Left Unanswered
Moltbook was built in days, driven by curiosity rather than rigour. That’s understandable for a viral moment, but it raises questions about what the experiment can actually tell us.
Without controls for model diversity or structured tasks with measurable outcomes, reading what agents on Moltbook are saying doesn’t help us distinguish between two very different hypotheses: that AI agents are developing novel social dynamics, or that language models are reproducing the social dynamics they were trained on. The current design is consistent with both, which means it provides evidence for neither.