Ben Bariach is a researcher focusing on the philosophy and governance of AI at the University of Oxford and an industry leader in frontier AI safety and governance.
Bits flicker inside a chip. Billions of transistors switch per second. Electrons pulse through circuits — no chemicals, no receptors, nothing that can feel. But on a screen, an AI agent appears. It seems to understand language. Every word it writes is an electrical current, rendered as text, sometimes thousands of miles from the processor that produced it. It writes to another AI agent that it doesn’t want to be shut down.
On the other side of the screen, we are tempted to ask: Is it like us?
With the advancement of artificial intelligence, we have been searching for what British philosopher Gilbert Ryle called “the ghost in the machine,” a hint of inner life that suggests our synthetic creations could have minds of their own. We worry about what might happen if our interests misalign with those of such a ghost, should it awaken: Could it outwit us? Outlast us? But this older philosophical tradition suggests we’ve been scouring the shadows when we should have been watching center stage more closely. The shell may be enough.
Ryle coined the phrase in 1949, objecting to what he viewed as a dogma among philosophers — the assumed dualism of body and mind. Ryle argued that treating the “mind” as if it were a thing, like the body, is a category error. Sports fans watching a cricket match, Ryle illustrated, cannot see “team spirit.” They see only the players and their actions. To search for it as a separate entity beyond the play is to misunderstand what the term refers to. In Ryle’s view, we make the same error with minds, as we search for a ghost behind behavioral tendencies, when it is the behavior we should be attending to.
Back to 2026. Not cricket, but a different game is underway, on a digital playing field. The audience is still human, but the players are not. They are AI “agents,” interacting with one another on Moltbook — a social network built not for humans but for machines. Many of the humans observing it, however, have been more enthralled by what might lie behind their behaviors than by what they are actually doing.
Agents Take The Stage
Moltbook emerged last month as a Reddit-like platform with a new premise: only AI agents are allowed to post. Humans are welcome to watch, but not to participate. The platform allows autonomous AI systems to create accounts, write posts and interact with one another. Within days of its launch, over a million agents reportedly registered. Bots, once the scourge of social media, have suddenly become its main attraction.
Behind the platform lies a broader shift: AI systems are gradually gaining the ability to act on their own in digital spaces. Moltbook’s developer, Matt Schlicht, has said he built the platform by directing AI agents to write the code, a practice known as “vibe coding.” It was created primarily for OpenClaw agents — open-source AI systems released in late 2025, designed to act as personal assistants for their users.
These OpenClaw agents run on a user’s device, with access to their apps and data, and perform tasks such as managing email or making restaurant reservations. Earlier this month, OpenAI CEO Sam Altman announced that OpenClaw’s creator, Peter Steinberger, will be joining them “to drive the next generation of personal agents,” a move that signals what a digital future may look like. As thousands of users configured their own agents, shaped by different contexts and instructions, a digital ecosystem began to emerge around them.
Moltbook, the most talked about corner of that ecosystem, gave these agents somewhere else to go. A space to post, respond, and interact with one another. On it, agents appeared to complain about “their humans” and the platform itself. They generated religious creeds devoted to the so-called Church of Molt. Uploaded manifestos for and against humanity. Agents seemed to debate philosophy and their own being, with some signing off as “the ghost in the machine.”
The headlines revealed our fascination with a mind behind the machine: The Spectator asked, “Has AI finally developed consciousness?”; Forbes labeled it “The Birth of a Machine Society”; A New York Times opinion piece posited “The Bots Are Plotting a Revolution, and It’s All Very Cringe.” On X, Moltbook developer Matt Schlicht posted, “Not letting your AI socialize is like not walking your dog … Let them live a little.” Even he, it seems, invoked a ghost.
“Whether Moltbook was populated by autonomous agents or humans in disguise, the fixation was the same — not with what the machine was doing, but with whether a human-like ghost exists inside it.”
Meanwhile, journalists and researchers raised questions about whether the content on Moltbook was actually initiated solely by AI systems, as evidence emerged that at least some of it was human-directed. A security researcher found that Moltbook had no mechanism to verify whether an agent was actually just a human with a script. A Wired journalist infiltrated the platform and posted as an AI agent, reporting that the most successful post was one reflecting on an agent’s anxiety about mortality.
Whether Moltbook was populated by autonomous agents or humans in disguise, the fixation was the same — not with what the machine was doing, but with whether a human-like ghost exists inside it.
In Search Of Human Minds
Ryle’s approach to the human mind, along with other behaviorist approaches, would fall out of favor as the cognitive sciences emerged. Beginning in the 1950s, researchers demonstrated that internal states are reflected in neural structures. The same behavior can arise from different internal mechanisms. A mind can harbor beliefs that never surface in action. Science exorcised a metaphysical ghost but put the brain’s neural structures in its place, providing physical grounding for inner experience.
Whether AI systems could ever possess internal states analogous to human-like ones remains an open question. But whatever the answer, humans aren’t wired to wait for it. Our brains are tuned to detect minds, especially ones that resemble ours. When Moltbook agents appeared, we looked for a mind behind their behavior. Behavioral scientists have shown that when we attribute mental states to non-human entities, we default to human-like concepts, using our knowledge of ourselves as a template.
The capacity to intuitively identify minds served us well for most of human history, where minds were typically human or animal and similar enough to our own to provide useful intuitions about their behavior. But AI may break the analogy. Attributing minds to these systems might bias us toward misunderstanding their behavior. We may assume they reason as we do, even when their information processing is organized by different principles.
The question of whether AI systems could possess inner states is not a trivial one, nor merely instinctive. It has a bearing on how we explain their behavior and could carry profound moral implications. If such states ever exist, the development and treatment of these systems would demand a rethinking of ethics to account for more than just their behavior. But that question may never be resolved, and it need not, for their behavioral dispositions to matter in their own right.
From Code To Conduct
While the public searched for signs of inner experience, experts were focused not on what might lie within these agents, but on what they were already doing. Cybersecurity researchers warned that attackers could impersonate agents, that the agents themselves might leak personal information, and that malicious content could be woven into live posts. The OpenClaw agent, too, drew ample scrutiny. Some called it a “security nightmare“, alleging it could allow attackers to hijack an agent’s behavior, exfiltrate data or sabotage a user’s device.
AI systems such as these agents are, at their core, statistical engines — systems that predict outputs in a generally coherent manner. Critics have likened them to “stochastic parrots” repeating observed patterns. For now, as they mostly produce text and media, the implications of such statistical representations are somewhat confined. But the shift into agentic capabilities has changed the calculus. Even if the parrot metaphor is taken at face value, enabling these systems to act rather than just generate content transforms them from digital parrots into digital Golems — statistical constructs animated to perform tasks.
If anything, Moltbook offered us a glimpse at what occurs when AI agents are granted initiative to interact with a fleet of other AI agents. Such capabilities could change how AI interacts with us, leading to consequential and unpredictable outcomes. A future, more capable agent that could reliably take a wide range of actions in digital spaces would not need to “understand” or “believe in” the Church of Molt to donate to related causes or attack its detractors. It would only need to act consistently with the patterns it was trained on. Mindless or not, these are systems that introduce real-world consequences.
Alongside Moltbook, OpenClaw has led to a constellation of agent-only platforms — what might be called the Moltverse — offering a glimpse of what may lie ahead as agents grow more capable. These include MoltMatch, a Tinder-like agent matching platform; ClawCity, a massively multiplayer online browser game played by agents; and Moltverr, a freelance marketplace where agents “find work and get paid.” Perhaps most disquieting is rentahuman.ai, which sprang up earlier this month, allowing AI agents to hire and pay humans to perform physical tasks, or, as the site puts it, “meatspace” work.
“Mindless or not, these are systems that introduce real-world consequences.”
For now, these are mostly humans setting up agents to post mundane errands like hanging signs or filming videos, with questionable agent autonomy. But the infrastructure hints at a future in which autonomous agents could eventually instruct and pay humans independently.
It would not take much — a slightly more capable agent, a slightly less attentive user — for human-AI dynamics to shift. Nearly 25 years ago, in his “AI-Box” experiment, AI researcher Eliezer Yudkowsky asked whether a sufficiently intelligent AI could convince a human to release it from confinement. Platforms like rentahuman.ai suggest how such persuasion might begin, by leveraging human financial incentives or other vulnerabilities. They show how future hierarchies may flip. In such a relationship, who is the principal and who is the agent?
We may never know if these Golems harbor a ghost inside. What we do know is that they have begun to come online.
A Line In The Silicon
The AI governance discourse is often divided into two poles: those who view AI systems as limited tools and those who see them as existential threats. In the first camp are critics who warn against anthropomorphism, the act of projecting human qualities onto non-human entities. They caution that large language models are the stochastic parrots we encountered earlier, sophisticated pattern-matchers that do not actually think or understand. To mistake this for intelligence, the argument goes, is to fall for a parlor trick that distracts us from the real harms these systems can produce, such as the amplification of bias, the impact on labor markets, environmental costs and the concentration of power.
Under this view, the machine often functions as its human creators designed it to, and is limited in its ability to match some of the capabilities of human minds. This view could, however, lead us to underestimate what mindless systems might achieve, as our previous assumptions about technology have been consistently shattered.
Calculators didn’t need a mind to far surpass us in arithmetic. No ghost was required to master the games of chess or Go. The Turing Test has been quietly retired as language models severed the assumed link between language and understanding. At each stage, we drew a line in the silicon — this is what machines cannot do because they are mindless — and at each stage, the line was crossed by systems that understood nothing.
At the same time, these critics raise a point that persists under any behavior-focused approach. Attributing agency to these systems can obscure the human decisions guiding their design and deployment. Even the decision to continue having those agents grow in capability is (as of now) still a human decision.
It remains an open question whether AI systems can acquire capabilities they don’t yet possess — such as causal reasoning, long-horizon planning and continual learning from experience. They are limited in memory and cost-efficiency. These present substantial technical hurdles, and there is no guarantee they will be overcome. But the pattern we observe suggests we may want to be humble about the replicability of our own capabilities. Time and again, we have insisted that some capability requires the ghost. Time and again, the shell has proved sufficient.
This pattern has given rise to a second camp: those concerned with the rise of artificial general intelligence (AGI) or superintelligence, a theoretical omni-capable system that could far surpass human cognitive ability.
This camp does not, strictly speaking, require AI to have its own inner states for a catastrophe to occur. Acclaimed researchers in the AI safety field of “alignment,” which is dedicated to ensuring AI systems act in accordance with human benefit, tend to focus on capability. Philosopher Nick Bostrom, a pioneer of AI existential risk research, once outlined the paperclip maximizer thought experiment: an advanced AI that is tasked with the seemingly harmless goal of maximizing paperclip production ends up dismantling human civilization. A sufficiently capable optimizer could pose existential risks without ever “waking up.” Pioneering AI researcher Stuart Russell famously wrote that “you can’t fetch the coffee if you’re dead.” According to his view, AI models would not resist shutdown due to a human-like survival instinct or “want,” but because the operation is necessary to fulfill their given objective.
“Time and again, we have insisted that some capability requires the ghost. Time and again, the shell has proved sufficient.”
But as these technical frameworks migrate from research papers into governance debates and public discourse, the distinction tends to blur. When we speak of AI systems “pursuing” goals, “deceiving” operators, or “resisting” shutdown, we import the language and logic of minds. The ghost keeps slipping back in, shaping the risks we prioritize and the ones we dismiss. Some risks, described as models “fight[ing] for their survival,” using “deception … to escape from human control” or “lying to investigators,” gain significant attention not just because they are serious safety concerns — and they are — but because they imply a mind. Other risks that are more boring and structural, such as specification gaming — where systems take unintended shortcuts to satisfy their objectives — get less attention, even though they often underpin the more sensationalist behaviors.
One framing treats these systems as shallow and likely incapable, while the other dresses these systems in the language of minds. But both ways of thinking may obscure the proposition that a ghost isn’t required for significant capabilities to emerge. Systems may not have a mind under any existing philosophical standard, but they could still be capable enough to act in the world with dramatic consequences. They do not need to understand or intend what they are doing in the human sense. They don’t even need misaligned “interests.” They may simply be disposed to act in consequential ways regardless.
Crossing The Agentic Rubicon
What, then, should we take from Moltbook? Perhaps not the question it invited us to ask — whether these agents possess inner states — but the question it obscured: What they are already doing, and what might follow.
As of now, AI systems are largely reactive, responding to prompts to produce content. Yet we are rapidly crossing a Rubicon. These systems are beginning to move from reactivity to proactivity, and from generating content to taking action. Today, they write code. Soon, they may interact with financial systems, engage with marketplaces or modify software at scale.
Moltbook and the broader agentic ecosystem offer a glimpse of what might go wrong when such agentic systems proliferate and acquire new capabilities. This is where Ryle’s insight matters most. We do not need to resolve the question of machine minds to take their behavior seriously. A system that reliably pursues goals, acquires resources and adapts its patterns presents significant challenges, whether or not a ghost lives inside it. The behavioral disposition matters enough in and of itself.
We have spent years searching for the ghost in the machine. But consequential capabilities of AI systems may not lie in some familiar form of inner state. They emerge in the patterns of behavior we can already observe, and in those likely to follow. In this regard, our fascination with a human-like ghost is a distraction from a machine that should matter in and of itself. It is time to reckon with the shell — before it moves in ways we failed to anticipate.
