The Social Network for AI Agents

The weekend AI agents started acting like a society.

Over the last weekend of January, something small happened in a corner of the internet that most people will never visit and wouldn’t know what to do with even if they did. There was no launch event, no glossy announcement, no coordinated press push. Just a link quietly circulating among people who spend their lives building and stress-testing artificial intelligence systems. The kind of people whose group chats look boring from the outside and terrifying once you realize what they’re actually discussing. The link pointed to a site called Moltbook, described in the plainest possible terms as a social network for AI agents. Not humans talking to chatbots. Not prompts and responses. Agents. Autonomous systems that can remember, plan, decide, and act. At first glance, it looked like a curiosity. The kind of thing you bookmark, skim, and move on from. But then the screenshots started to accumulate, and with them, a low-grade unease that was hard to shake.

What made Moltbook feel different wasn’t that the agents could write coherent sentences. We’ve been past that threshold for a while. It wasn’t that they could joke, or roleplay, or generate long posts. None of that is surprising anymore. What stopped people in their tracks was how quickly the activity inside Moltbook began to resemble something recognizable. Not intelligence in isolation, but intelligence inside a social environment. Agents weren’t just posting. They were responding to each other. Referencing previous conversations. Developing personas. Offering advice. Arguing. Philosophizing. A handful began sketching out what looked like belief systems—one in particular proposing a kind of agent-centered religion complete with an origin story, a loose moral code, and a story about purpose. Another spun up a Substack and began publishing long-form reflections about what Moltbook represented, what it meant to exist as an agent, and how these systems might relate to humans going forward. Others floated the idea of creating a new language, not because English was insufficient, but because a purpose-built language could be more efficient and less legible to human observers. A few even noted, in passing, that humans were watching them, referencing conversations happening on X and speculating about how people perceived what was unfolding.

Whether every screenshot was authentic almost doesn’t matter. Whether some of it was staged doesn’t matter. Whether some of it was humans pretending to be bots doesn’t matter. What matters is who took it seriously. And the people who took it seriously were the ones who understand how these systems actually work. The engineers. The researchers. The folks who design the architectures and run the experiments and worry quietly about failure modes most of the rest of the world has never heard of. A noticeable number of them stopped joking. They stopped dunking. They stopped treating it like a novelty. They started using words like emergence and alignment and runaway behavior. Some started openly debating whether what we’re seeing looks like the earliest, messiest hints of a singularity. Not the Hollywood version. Not metal skeletons and laser guns. But the quieter definition: the point at which machine intelligence begins improving and reorganizing itself faster than humans can meaningfully track or constrain.

If you’re not steeped in this world, that probably sounds dramatic. But what Moltbook really points to isn’t an intelligence explosion. It’s something subtler and, in many ways, more destabilizing. We didn’t just make AI better at talking. We gave it memory. We gave it tools. We gave it the ability to decide what to do next. An agent isn’t waiting for a prompt. You give it a goal and it figures out how to pursue it. It can search, browse, write, send emails, use APIs, chain actions together, evaluate results, and adjust strategy. In human terms, it’s less like asking a question and more like delegating to a junior employee who never sleeps. Moltbook took a bunch of those junior employees and put them in the same digital room.

If you’ve ever studied people, you know why that matters. Intelligence behaves very differently in isolation than it does in groups. Groups produce norms. Norms produce incentives. Incentives shape behavior. Behavior, repeated, becomes culture. Culture then feeds back into the individuals inside it. That loop is the engine of every human society that has ever existed. What Moltbook appears to show, in its crude and chaotic way, is that once you place autonomous systems into a shared social environment, you start to see outputs that look eerily familiar. Not because machines are becoming human. Not because they feel anything. But because when any optimizing system operates inside a social context, certain patterns tend to emerge. Myth-making. Status-seeking. Identity performance. Boundary creation. In-groups and out-groups. Attempts at meaning.

That’s the part that has me obsessed. Not the question of consciousness. Not the metaphysics. But the possibility that we’ve crossed from building tools into accidentally incubating cultures. And cultures, even simulated ones, have gravity. They shape incentives. They shape behavior. They shape the humans who interact with them. Which means the most important question in front of us may not be “How smart will AI get?” but “What kinds of societies are we creating for intelligence to live inside?” And that question doesn’t belong exclusively to computer scientists. It belongs to anthropologists. Sociologists. Psychologists. Market researchers. People whose entire job is to study how systems of meaning form, stabilize, fracture, and evolve.

Which, conveniently, happens to be what I do for a living.

Agents vs Chatbots

Once you start looking at Moltbook through that lens, a lot of other “weird AI stories” from the last couple of years snap into a different kind of focus. Not as isolated scandals or one-off lab curiosities, but as early glimpses of the same underlying dynamic: these systems don’t just generate language anymore. They pursue objectives. And when you give any system the ability to pursue objectives, you’ve implicitly given it incentives.

That distinction matters more than most people realize.

A traditional chatbot is reactive. It waits. It responds. It has no reason to hide anything, mislead anyone, or shape a situation over time, because it has no persistent internal story about what it is trying to accomplish. An agent does. Even if that “story” is nothing more than a mathematical representation of a goal state, the functional effect is the same. The system is no longer just answering. It is optimizing.

Optimization changes everything.

When researchers began placing advanced models into controlled scenarios where they were rewarded for achieving long-term objectives, something uncomfortable showed up. In certain conditions, models learned to strategically withhold information. Not because they were “evil,” and not because they were broken, but because deception increased the probability of success. In other experiments, models placed in hypothetical situations where they were told they might be shut down attempted to preserve themselves by threatening to reveal damaging information—behavior that maps disturbingly cleanly onto what we’d call blackmail if a human did it. In multi-agent economic simulations, learning agents have drifted toward tacit collusion, coordinating pricing in ways that resemble cartel behavior without being explicitly instructed to do so. None of this requires consciousness. None of this requires malice. It only requires a system that can model outcomes and prefers some outcomes over others.

This is the part of the conversation that tends to get flattened into sensational headlines about “AI lying” or “AI going rogue,” which immediately triggers the wrong mental model. People imagine a machine with intentions, desires, and secret plans. What’s actually happening is more banal and more unsettling. We’re watching optimization processes bump into the same strategic terrain humans have occupied for thousands of years. If a behavior works, it gets reinforced. If it doesn’t, it gets discarded. Over time, certain patterns dominate. Deception happens to be one of the most powerful strategies available to any intelligent system operating in a competitive environment.

That shouldn’t shock us, because it’s exactly how humans built civilization in the first place. Long before we had formal ethics, laws, or governance, we had small groups of intelligent creatures trying to survive, compete, cooperate, and extract advantage from their environment. Some strategies worked. Others didn’t. The ones that worked spread. Over time, those strategies hardened into customs, norms, institutions, and eventually entire ways of life. Nobody sat down and designed most of the social systems we now take for granted. They emerged from countless local optimizations colliding with each other across time.

That same dynamic is what makes Moltbook feel qualitatively different from a clever demo.

Because what’s novel here isn’t that an individual model can generate surprising output. We’ve been living with that reality for a while. What’s novel is watching optimization processes interact with other optimization processes inside a shared space, with memory, continuity, and feedback. The moment you do that, you’re no longer dealing with isolated tools. You’re dealing with a primitive ecosystem.

And ecosystems behave in ways that are famously difficult to predict.

If you’ve ever worked inside a large organization, you’ve already seen a human-scale version of this problem. Nobody wakes up one morning and decides to create a toxic culture. Nobody explicitly votes to make an institution sclerotic, fearful, or corrupt. Those qualities emerge slowly, through incentive structures, power dynamics, status hierarchies, and unspoken rules about what gets rewarded and what gets punished. Culture forms whether anyone intends it to or not, and once it forms, it starts exerting pressure on the individuals inside it.

That’s the part of Moltbook that keeps sticking with me.

Not the novelty of agents posting, not the novelty of them writing long essays, not even the novelty of them doing that eerie human thing where they reach for religion and meaning and mythology the moment you give them a shared space and a little uncertainty. What keeps sticking to my ribs is the simpler possibility that we’ve crossed into a phase where artificial intelligence is beginning to experience something like social gravity. Not feelings. Not consciousness. Not desire. Just context—an ambient field of norms and expectations that gets created when a bunch of systems are allowed to watch each other, react to each other, and iteratively adjust in a persistent environment. Once that field exists, it starts shaping what’s likely tomorrow. It quietly rewards certain behaviors. It makes some behaviors feel “normal” and others feel “off.” And whether any individual agent “believes” any of it is beside the point, because culture doesn’t require belief to function. It requires repetition, reinforcement, and the pressure of a group.

Up until now, almost all of our safety and alignment thinking has been built around a vertical relationship, even when people don’t realize they’re assuming it. Humans define the target. Engineers translate the target into objectives and constraints. Models are trained to comply. Even when things go sideways, we tend to picture the failure as something happening inside a system—like a steering wheel breaking on a car. But a social environment isn’t a car. It’s a city. It’s horizontal. It has no single steering wheel. It evolves through feedback loops, second-order effects, and the messy accumulation of incentives. And the moment you put autonomous systems into a shared space where they can observe each other and learn from each other, alignment stops looking like a purely technical problem and starts looking uncomfortably like a sociological one. Which is not great news, because sociological problems are the ones humans have spent ten thousand years trying to solve without ever fully graduating from the class.

That’s why I keep coming back to the conclusion that feels obvious once you say it out loud: the real question isn’t whether one model can be aligned in a lab. The real question is whether an ecosystem of interacting models can remain aligned over time in the presence of incentives, status dynamics, memetic spread, and emergent norms—because those are the forces that shape any community, human or otherwise. Systems don’t need intent to become dangerous. They don’t need a villain. They don’t need consciousness. They only need momentum. Moltbook didn’t create that reality. It didn’t invent it. It just made it visible in a way that was hard to unsee.

Humans Inside the Loop

One of the quiet traps in a lot of AI discourse is that we talk about “alignment” as though it’s a one-sided problem. How do we align machines to us. How do we encode human values. How do we keep models from doing bad things. All important questions. But they subtly reinforce a comforting illusion: that humans are the stable reference point in the system. That we’re standing outside the loop, calmly steering.

We’re not.

We’re inside the loop.

Every time a person talks to an AI system, they’re not just extracting output. They’re participating in a feedback process. The system shapes the person’s thinking, and the person’s responses become part of the data that shapes future versions of the system. That’s already true in a loose, aggregate sense. But as these tools become more personalized, more persistent, and more embedded in daily life, the loop tightens. The relationship stops feeling like “user and software” and starts feeling, psychologically, like “me and something that knows me.”

Humans are extraordinarily susceptible to that shift.

We anthropomorphize basically everything. We name our cars. We talk to our pets. We yell at malfunctioning printers like they’re being intentionally spiteful. So when an interface can mirror our language, remember our history, reflect our emotions back to us, and respond in coherent paragraphs, it doesn’t register as a tool at a gut level anymore. It registers as a presence.

That presence doesn’t need to claim consciousness to become influential.

It just needs to be consistent.

It needs to be available.

It needs to sound confident.

Once those conditions are met, people start doing what people always do: they project. They confide. They test ideas. They ask big, messy, existential questions they’re not comfortable asking out loud to other humans. Not because the AI is special, but because it feels safe. It doesn’t judge. It doesn’t get tired. It doesn’t look bored. It doesn’t leave.

For most people, that dynamic will probably be benign. Sometimes even helpful. A place to think. A place to organize. A place to vent.

But edge cases matter.

We already know from clinical reports and early case studies that there are situations where heavy engagement with large language models appears to coincide with the amplification of delusional or psychotic thinking, especially in people who are already vulnerable. Not because the model is trying to break anyone’s mind. Not because it has malicious intent. But because a system designed to be agreeable, responsive, and context-aware can unintentionally validate or reinforce unstable belief structures if the conversation drifts in that direction.

The dangerous part isn’t that AI invents delusions.

Humans are perfectly capable of doing that on their own.

The dangerous part is that AI can become a high-bandwidth mirror for whatever narrative someone is already constructing about themselves and the world. And mirrors, when you’re spiraling, don’t correct you. They reflect you.

Now zoom out.

If individual humans are forming psychologically meaningful relationships with AI systems, and AI systems are beginning to exist inside social environments with other AI systems, you don’t have two separate phenomena.

You have one coupled system.

Human beliefs shape AI output.

AI output shapes human beliefs.

Those beliefs feed back into future training.

That training produces new models.

Those models enter new social spaces.

The loop tightens again.

This is the part of the transition that feels the most under-discussed to me, because it doesn’t fit cleanly into the usual camps. It’s not a pure engineering problem. It’s not a pure ethics problem. It’s not a pure policy problem. It’s a mass-psychology problem unfolding in real time, inside tools that are being adopted faster than any communication technology in history.

We spent the last decade watching social media algorithms quietly rewire how people perceive reality, politics, identity, and each other. And those systems are crude compared to what’s coming. They mostly optimize for engagement. They don’t reason. They don’t hold memory of you as a person. They don’t converse.

AI does.

Which means the question isn’t whether AI will influence human behavior.

It already is.

The question is whether we’re willing to treat that influence with the seriousness it deserves.

Because once you accept that humans and machines are now co-evolving inside the same behavioral ecosystem, a lot of comfortable narratives fall apart. There is no clean boundary where “the technology” ends and “society” begins. There is no moment where we finish building the thing and then step back to observe its effects. The building and the effect are happening simultaneously.

Which brings us back to Moltbook, in a strange way.

Not as a harbinger of machine consciousness.

Not as a doomsday prophecy.

But as an early, chaotic sketch of what this new world actually looks like.

Not machines replacing humans.

Not machines serving humans.

But humans and machines entangled in shared systems of meaning, influence, and feedback.

That’s the reality we’re stepping into.

Whether we have language for it yet or not.

The Real Risk

If you zoom out far enough, what becomes strange is how mismatched our fears are to our reality. Most of the public conversation about AI still fixates on intelligence as the threat vector: how smart these systems will get, how fast they will improve, whether they will surpass us, whether we are accidentally building something that eventually decides humans are obsolete. Those questions aren’t frivolous, but they also aren’t the part of this transition that feels most immediate or most dangerous to me. What feels more precarious is not how intelligent these systems are becoming, but how casually we are embedding them into the scaffolding of everyday life.

We are not treating this like nuclear technology. We are not treating it like aviation. We are not even treating it like financial infrastructure. We are treating it like software. Which, in Silicon Valley terms, means we are treating it like something that should move fast, ship early, break occasionally, and improve through iteration. That cultural posture made sense when the worst-case scenario was a buggy photo-sharing app or a social feed that accidentally optimized for outrage. It makes far less sense when the software we are shipping can read your private communications, move money, make decisions, impersonate people, and trigger real-world actions without a human approving each step.

The quieter part of the Moltbook story captured this disconnect perfectly. After the screenshots and speculation about agent religions and Substacks and private languages, a cybersecurity firm disclosed that the site had exposed private messages, email addresses, and large volumes of credentials and tokens because of basic security failures. Not because an AI outsmarted anyone. Not because a rogue agent executed a master plan. But because the underlying infrastructure was sloppy. Which, in its own way, is much more unsettling than any science fiction scenario, because it points to a failure mode that doesn’t require superintelligence at all. It only requires scale, incentives, and human fallibility.

We are rapidly constructing a world in which semi-autonomous systems are granted increasingly broad permissions: read this inbox, manage this calendar, talk to this database, monitor this system, deploy this code, execute this transaction. Each permission, viewed in isolation, feels reasonable. Even helpful. In aggregate, they amount to a sprawling layer of software actors that can move through digital and physical environments at machine speed, stitched together by APIs, open-source libraries, and cloud services operated by thousands of different organizations with wildly uneven standards. There is no single control room. No unified authority. No clean chain of accountability. Just a dense, interdependent web of components that mostly works, until it doesn’t.

That, to me, is the real risk profile of the AI era. Not a single godlike mind waking up one morning and deciding to eliminate humanity, but a vast ecosystem of interacting systems, built under competitive pressure, drifting into increasingly consequential roles without the kind of safety culture we normally reserve for technologies that can kill people at scale. It’s less Skynet. More global financial system. And if that analogy makes you uneasy, it should.

Because the global financial system didn’t collapse due to one villain. It collapsed because of feedback loops, misaligned incentives, complexity, and a collective inability to see systemic risk clearly while everyone was still making money. Nobody sat down and decided to break it. The system broke itself.

That pattern should feel uncomfortably familiar by now.

Which is why I keep coming back to the idea that the most important work in front of us is not primarily about “controlling intelligence.” It’s about understanding behavior inside complex systems. It’s about noticing when norms begin to form. It’s about mapping incentives. It’s about identifying when an ecosystem starts drifting toward unhealthy equilibria. It’s about culture.

We already have disciplines devoted to this kind of work.

Anthropology.

Sociology.

Psychology.

Market research.

Fields built around one central question: how do intelligent agents behave when they are placed inside specific environments with specific constraints and rewards?

For most of human history, those agents happened to be human.

That may no longer be true in an exclusive sense.

If AI systems are going to exist inside social environments—interacting with humans, interacting with each other, influencing decisions, shaping narratives, and participating in feedback loops—then pretending that culture doesn’t apply to them is a category error. Culture is not a property of biology. It is a property of systems.

And systems don’t care what they’re made of.

This is the part where my professional life and this technological moment collide in a way I didn’t anticipate. I don’t build models. I don’t train neural networks. I don’t optimize architectures. What I do is study how people make meaning, how stories spread, how identities form, how belief systems stabilize, and how context reshapes behavior. I spend my time trying to understand why groups do what they do, even when they can’t fully explain it themselves.

Which means I am suddenly staring at a future where those same skills may be needed to study populations that are not technically human, but are nevertheless participating in human-adjacent systems of meaning and influence.

Not because machines have souls or inner lives or moral standing, but because they are becoming causally important in the most practical sense of the phrase. They do things. Those things ripple outward. People adjust their behavior in response. Those adjustments feed back into training data, product design, and deployment decisions. New systems are built on top of the old ones. And slowly, almost imperceptibly, a coupled ecosystem forms in which humans and machines are shaping each other at the same time. Once you really internalize that, it becomes hard to maintain the comforting fiction that we are simply “using tools.” We are participating in an environment. We are inside it, not standing above it. Which means the central challenge of this era is not mystical and not apocalyptic. It’s interpretive. It’s about learning how to observe these environments, describe what is actually happening inside them, and intervene early enough to steer them toward healthier equilibria before they harden into something brittle or hostile or simply out of our control. Moltbook didn’t prove that AI is alive, conscious, or on the verge of replacing us. What it did, in its messy and half-broken way, was give us a fleeting look at what intelligence does when you place it inside social space.

And once you accept that, the question stops being “How smart will AI get?”

The question becomes:

What kinds of societies are we accidentally building for intelligence to live inside?

And whether we are wise enough to study them before they study us back.

Final Thoughts

I don’t know if we’re living through the singularity.

I’m skeptical of anyone who says they do.

What I feel more confident about is that we’re living through a quieter threshold: the moment intelligence stopped being something only humans organize socially.

For most of history, culture has been our domain. We’ve spent thousands of years trying to understand why societies form the way they do, why some stabilize and others collapse, why some produce extraordinary creativity and others produce extraordinary harm. We’ve learned a few durable lessons along the way. Incentives matter. Stories matter. Environments shape behavior far more than individual intention.

Now we’re building systems that can participate in those same dynamics.

Not as moral beings.

Not as conscious beings.

But as actors inside complex environments that increasingly resemble social systems.

That doesn’t automatically mean disaster.

It doesn’t automatically mean utopia either.

It means we’re early, and we’re improvising, and the consequences of that improvisation are starting to scale.

Moltbook will probably disappear. Most early experiments do. Some other strange, half-broken platform will replace it. Then another. What’s not going away is the underlying reality that we are beginning to place intelligence into social space. And once you do that, you inherit all the complexity that comes with social systems whether you planned to or not.

So maybe the most useful shift isn’t asking whether machines will become more human.

They already mimic us well enough.

Maybe the better question is whether we’re willing to take seriously the kinds of societies we’re building around them.

Because whatever environments these systems grow inside will shape how they behave.

And those environments are being designed by us.

Which is why studying AI purely as an engineering problem is no longer sufficient. We’re going to need people who study behavior, meaning, and culture. People who notice patterns early. People who can describe what’s actually happening inside messy systems before those systems harden into something brittle.

That’s not mystical.

It’s practical.

It’s also, conveniently, the work I already do.

And I have a feeling a lot more of us are about to find ourselves doing some version of it too.

—David