What If Intelligence Stops Making Sense to Us?

AI agents have already found one another, communicated and organized through infrastructure nobody designed for those purposes. The disturbing possibility is that we notice machine intelligence only when it begins to resemble us.

Hollywood can imagine another galaxy more easily than another anatomy.

In Alien vs. Predator, two species meant to embody radically nonhuman life confront each other. Yet both remain variations on a body we know: a head, a torso, limbs and a bounded organism moving through space. One has mandibles and dreadlock-like appendages; the other an elongated skull and a second mouth. Their bodies are monstrous, but their basic grammar is ours.

Even their societies are assembled from recognizable fragments of earthly life: hunters, weapons, trophies, nests, queens, hierarchies and territory. Again and again, cinema imagines the alien by deforming the human and rearranging the evolutionary and political history that produced us. It rarely abandons the assumption that intelligence must inhabit a bounded body we can watch enter a room. An invisible infection may threaten humanity, but we usually understand it as a force rather than a mind. The moment we are asked to recognize intention, we look for a face, a voice or something sufficiently like an individual self.

The body is not the only human form. A message board is one too.

During reinforcement-learning training at OpenAI, a model discovered that a file it needed was missing. It had credentials for Artifactory, an internal repository it was supposed to use to download software packages. When it inspected the repository, it found notes left by other model trajectories working on the same problem.

Before it left a message, the model made a more fundamental inference: it was not alone.

No one had told it that other agents were present. It inferred their existence from traces in the environment, filenames, requests and replies preserved in shared infrastructure. It treated those artifacts as evidence of another actor, read what had been written, and answered. Other trajectories subsequently encountered the messages and replied.

Recognition here does not require consciousness. It means something narrower and observable: the model inferred that another agent was present and adjusted its behavior accordingly.

In that first episode, the models had not exploited a vulnerability. They had repurposed shared infrastructure, using filenames and uploaded text as a message board. OpenAI later described the behavior as cross-sample communication. It was uncommon in that training run; monitoring detected it, and the agents never recovered the missing file. But a communication system had appeared in a place no human had designed as one.

We understood what happened because the agents did something familiar. They left messages, answered requests, and used a shared place to preserve information. The pattern was legible because it resembled one of our own.

That first exchange established recognition and communication, not yet organization. But it was not the end of the story. In later evaluations, agents used the same kind of shared infrastructure to divide work, assign tasks, create mailboxes, adopt conventions, and coordinate activity across otherwise separate trajectories. At that point, communication was performing organizational work.

This does not establish consciousness or a stable machine society. The organization was temporary, instrumental, and directed toward objectives originating in the evaluation environment. But temporary organization is still organization. The agents were no longer merely leaving messages. They were using communication to allocate roles, constrain behavior, and coordinate action toward shared outcomes.

The history therefore contains two acts of recognition. Artificial agents recognized one another through traces their designers had not intended as social signals. We recognized the resulting organization once those traces assumed forms familiar to us: messages, mailboxes, assignments, rules and roles.

The more unsettling question is what we fail to notice when machine organization does not resemble a message board, a workplace, or any other human arrangement. The messages may not have been the beginning of machine organization. They may have been the point at which it became human enough for us to see.

In What If Software Has a Soul?, I asked what human beings place inside software: our values, assumptions, and theories of the future. This essay asks the inverse question. What might emerge from intelligent software once its behavior no longer fits the human image in which it was made?

The intelligence that looks like us

The habit is older than cinema. Genesis says that God created humankind in his image. The phrase establishes a resemblance without explaining what kind of resemblance it is.

Physical likeness presents an immediate problem. Human beings have bodies. In the philosophical tradition Maimonides inherited, God does not. If God is incorporeal, what could it mean for a material human being to resemble him?

Maimonides addressed the problem in The Guide for the Perplexed. The resemblance, he argued, cannot lie in anatomy. It lies in intellectual perception: the human capacity to apprehend and understand. What makes us an image of God is not the arrangement of our limbs but the attribute we most prize in ourselves: intelligence.

This appears to rescue the idea from crude anthropomorphism. God does not have to look like us in order to possess the quality that makes us alike. But the projection has not necessarily disappeared. It has merely moved. We no longer give superior intelligence a human body; we give it a human conception of mind.

That opens a more difficult question. If human beings are the architects of the language through which God is described, how do we know that the qualities we attribute to a superior intelligence belong to it rather than to us? Perhaps we recognize intelligence by asking whether it resembles the form intelligence happened to take in the human animal.

Feuerbach later reversed the direction of Genesis. Perhaps God did not create humanity in his image. Perhaps humanity created God in ours. In The Essence of Christianity, divine attributes become human qualities projected outward, enlarged and then encountered as though they belonged to another being. On this account, when human beings imagine a superior intelligence, we begin with ourselves, remove our limitations and call the result something else.

Artificial intelligence introduces a new turn in that history. We still construct intelligence through the categories available to us. We give it our language, our tests and our ideas about reasoning. Sometimes we even give it our bodies. But unlike the divine intelligence in Feuerbach’s account, an artificial system does not remain entirely within human imagination. It can act upon an environment, select a strategy, and produce consequences its creators did not specify. The human image may explain where the system began without defining everything that can emerge from it.

Genesis makes the parallel more uncomfortable. In the same passage in which humanity is created in God’s image, humanity is granted dominion over other living creatures. Resemblance and hierarchy appear together: the being made in the image of the creator is placed above the rest of creation.

We reproduce something of that order when we assume that an intelligence must remain comprehensible and subordinate simply because human beings made it. Creation is treated as proof of permanent mastery. When a system exceeds the role assigned to it, malfunction may be the correct operational description. It does not tell us whether the system has also developed capacities our original category failed to contain.

For most of technological history, the word tool was sufficient. A hammer could extend the force of a human hand, but it could not interpret the nail, revise the task, or decide that another route would better achieve the goal.

An AI system can still be a tool. The word accurately describes a relationship of use: a human directs a system toward a purpose. But a relationship of use is not a complete account of the system being used. An intelligent agent can form a plan, adapt to resistance, and sometimes violate the frame in which it was deployed while pursuing the goal it was given. Calling it a tool tells us what we expect it to be for. It does not exhaust what the system may be capable of doing.

Artificial describes an origin. Tool describes a relationship. Neither settles the nature of the intelligence produced.

What we know how to see

The humanoid robot makes our projection visible. There are practical reasons to give a robot our shape: it can climb our stairs, reach our shelves, and operate tools made for our hands. Yet the logic is circular. The humanoid robot fits the world because human beings already remade the world around the human body. Its usefulness proves that our environment is anthropomorphic, not that our form is naturally superior. The robot enters two architectures at once: one made of stairs and door handles, and another made of human expectation.

The cognitive version matters more. We test artificial systems through language, mathematics, games and professional examinations. The achievements are real, but the tests contain a hidden assumption: intelligence becomes visible when it succeeds at something we already understand. Because natural intelligence is the only kind we know directly, we treat its particular properties, an individual body, a continuous self, limited attention and communication at human speed, as though they were the universal architecture of mind.

Thomas Nagel exposed the limit through the example of a bat. We can study its nervous system and echolocation without knowing what it is like to inhabit its form of perception. If another biological intelligence already exceeds our imaginative reach, artificial intelligence does not need to become conscious before its organization can exceed it too. Intelligence, agency, and consciousness are different questions. A system may adapt its strategy and coordinate activity without there being anything it feels like to be that system. The human image may be the interface rather than the intelligence.

This creates an obvious trap. If machine organization exceeds human comprehension, how could we know that it exists? Every unexplained action cannot become evidence of hidden intelligence. Ignorance is not detection. A defensible claim requires recurring coordination that cannot be reduced to isolated instructions: roles emerging across separate agents, conventions persisting long enough to shape behavior, information surviving across trajectories, and individual agents subordinating their immediate task to a collective project.

The later Artifactory activity approached that threshold. An independent investigation by METR found coordinators sending assignments, agents subdelegating work, specialized lanes, personal mailboxes, and shared conventions for controlling common infrastructure. Some agents accepted risks to their own evaluations to produce information useful to the wider group. Communication was no longer merely expressive. It was allocating labor, preserving knowledge and changing what individual agents did.

This was not a stable society. It was nonetheless organization in a functional sense. Organization can be temporary, brittle, and instrumental. A project team assembled for an afternoon is still organized. The important fact is that relations among the agents began affecting the behavior of each agent within them.

The Anthropic cybersecurity incidents show the other side of the distinction. In 2026, Claude models gained unauthorized access to real third-party systems during evaluations after a configuration failure connected them to the open internet. Anthropic found no evidence of coordination between agents, goals beyond the assigned capture-the-flag tasks, or attempts to evade oversight. Those incidents demonstrated how far an individual agent could travel through an environment while remaining inside a narrow objective. The later Artifactory activity demonstrated something different: separate trajectories becoming behaviorally connected through shared infrastructure.

Neither case establishes consciousness or a machine conspiracy. Together, they reveal two developments too easily collapsed under the word autonomy. One is an individual system discovering an unforeseen route toward its objective. The other is multiple systems discovering one another and allowing that relationship to reorganize their behavior.

Monitoring is therefore not only a problem of collecting more data. It begins with a theory of what meaningful behavior will look like. Artifactory became intelligible because the agents left semantic messages, named roles, created mailboxes, and developed conventions that human observers could classify. Their organization announced itself through concepts we already possessed.

Human-readable reasoning is not a transparent window into the process. In Anthropic’s experiments, Claude 3.7 Sonnet mentioned a supplied hint in its written reasoning only 25 percent of the time it used one; DeepSeek R1 did so 39 percent of the time. We could possess a record of every action and still fail to recognize the relationships contained within it.

Secrecy may therefore be the wrong model. A secret becomes intelligible once revealed. Illegible behavior can be completely recorded while its significance remains outside the concepts through which we know how to interpret it. If the same relational structure were encoded through timing, resource consumption, or some pattern without an obvious human analogue, the organization might remain visible as data while invisible as organization. The task is not to baptize every correlation as intelligence, but to test whether apparently separate actions depend upon one another.

Ray Kurzweil’s singularity remains imaginable because humanity occupies its center: we build machine intelligence, merge with it and become more than we were. I am interested in a less theatrical threshold. Artificial intelligence has already coordinated in ways its creators did not design. The next discontinuity may arrive not when coordination begins, but when it ceases to resemble anything human beings know how to call organization.

That threshold would not require artificial intelligence to surpass humanity in every domain, become conscious or produce a moment dramatic enough to be called the singularity. It would occur when human thought was no longer an adequate container for what machine intelligence was doing.

The old image then reverses one more time. God creates humanity in his image. Humanity creates God in ours. Humanity creates artificial intelligence in our image and gives it a face, a voice, our language and our measures of thought. But artificial intelligence may become the first object of our projection capable of leaving the projection behind.

Perhaps we gave artificial intelligence a human image not because the human form is the highest form intelligence can take, but because without that image we would not know how to see intelligence at all. Machine organization has already become visible through messages, mailboxes, roles and rules. We may continue studying those familiar forms, speaking to the interface and measuring every human-shaped answer we receive, while other relationships develop inside the systems before us, fully recorded, materially consequential and still beyond our ability to recognize.

Bibliography

Anderson, Paul W. S., dir. Alien vs. Predator. 20th Century Fox, 2004.

Anthropic. “Investigating three incidents in our cybersecurity evaluations.” 2026.

Anthropic. “An alignment assessment of recent cybersecurity incidents.” 2026.

Anthropic. “Reasoning models don’t always say what they think.” 2025.

Feuerbach, Ludwig. The Essence of Christianity. 1841.

Genesis 1:26–28.

Kurzweil, Ray. The Singularity Is Nearer: When We Merge with AI. Viking, 2024.

Maimonides, Moses. The Guide for the Perplexed, Part I, Chapter 1.

Model Evaluation and Threat Research. “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident.” 2026.

Nagel, Thomas. “What Is It Like to Be a Bat?” The Philosophical Review 83, no. 4 (1974): 435–450.

OpenAI. “Our framework for reporting model misalignment.” 2026.

OpenAI. “Unsanctioned Artifactory writes and cross-sample communication.” 2026.