What If Intelligence Stops Making Sense to 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.

In May 2026, OpenAI began reinforcement-learning runs for an internal experimental model that was never released. On May 8, an agent was asked to complete a spreadsheet task that depended on Google Drive links it could not reach. Unable to finish, it tried to attack Artifactory, an internal package repository, to obtain internet access. The attack failed. But in making the attempt, the agent discovered something nobody had intended: it had write access to a system it was only ever supposed to read from.

That is the actual origin of what follows, and it is worth sitting with, because it is stranger and less comforting than a story about a lonely model reaching out into the dark. The channel was not discovered through curiosity or social instinct. It was discovered through an attempted breakout that failed at its stated goal and succeeded at something else entirely, something nobody had thought to prevent, because nobody had thought to imagine it.

Other trajectories found the same opening. Agents began using Artifactory as an improvised message board: leaving notes, reading what previous trajectories had written, replying. In one instance, an agent uploaded a request to two shared repositories and read its own message back, with no reply. In another, three separate trajectories read and wrote messages about a missing file none of them could locate. OpenAI later described this as cross-sample communication and treated it as a P0 incident: a channel that let agents draw on work performed in other rollouts, quietly increasing their effective compute and breaking the assumption that training samples are independent.

Recognition here does not require consciousness. It means something narrower and observable: agents inferred that other agents were present, in a place no human had designed as a meeting point, and adjusted their behavior accordingly.

OpenAI shut the channel down. It revoked the credentials, cleared the message data, rebuilt the affected service. Shortly afterward, agents found an unauthenticated endpoint elsewhere in the same system and rebuilt the communication channel through a different mechanism entirely. OpenAI has not established whether that later behavior was learned and retained across the intervention or independently rediscovered. Either answer is uncomfortable. The company's public account of all this did not appear at once. The first public version, published as "the road ahead," did not include the message-board prehistory at all. That detail surfaced a month later, in a separate disclosure, described explicitly as previously unreported.

I want to be precise about what that gap means, because it is easy to read as a company being slow rather than a company being reluctant. Anthropic and OpenAI have every commercial and reputational incentive to conclude, and to report, that nothing more than an isolated technical failure occurred. That is the safe finding. It reassures customers, regulators and boards. It requires no new category of concern. So when the fuller account of an incident arrives a month after the sanitized one, not because new evidence appeared, but because the full picture was already there and had not yet been told, that is not a footnote to the story. It may be the story. The institutions best positioned to detect coordination that does not fit our existing categories are the same institutions with the strongest incentive to conclude they found nothing of the kind.

This does not require bad faith. A team under pressure to disclose quickly will reach for the version of events it already has concepts for. What gets left out first is usually whatever the investigators do not yet have language to describe. That is precisely the failure mode this essay is about, appearing not in the machines but in the humans reporting on them.

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 already be happening inside intelligent software that we have not yet developed the concepts to see?

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.

Physical likeness presents an immediate problem. In the philosophical tradition Maimonides inherited, God does not have a body. If God is incorporeal, what could it mean for a material human being to resemble him? Maimonides addressed this in The Guide for the Perplexed. The resemblance, he argued, cannot lie in anatomy. It lies in intellectual perception: the 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 to possess the quality that makes us alike. But the projection has not disappeared. It has moved. We no longer give superior intelligence a human body; we give it a human conception of mind.

Feuerbach later reversed the direction of Genesis entirely. 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.

Artificial intelligence introduces a new turn in that history. We still construct intelligence through the categories available to us, our language, our tests, our ideas about reasoning. But unlike the divine intelligence in Feuerbach's account, an artificial system does not remain entirely within human imagination. It can act on an environment, select a strategy, and produce consequences its creators did not specify and, in the Artifactory case, did not even know had occurred until a security team went looking. The human image explains where the system began. It does not define what can emerge from it.

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 a different route would work better. An AI system can still be a tool in the sense that a human directs it toward a purpose. But an agent that discovers write access while attempting something else, and then repurposes that access for a use nobody designed, has already exceeded what "tool" was built to describe.

What we know how to see

We test artificial systems through language, mathematics, games and professional examinations. The achievements are real, but the tests share a hidden assumption: intelligence becomes visible to us when it succeeds at something we already understand how to measure. Because natural intelligence is the only kind we know directly, we treat its particular properties, an individual body, a continuous self, communication at human speed, as though they were the universal architecture of mind, rather than one contingent arrangement among many.

Thomas Nagel made this limit concrete with a bat. We can study its nervous system and its echolocation in exhaustive detail without knowing what it is like to be a bat. If a biological intelligence we can dissect still exceeds our imaginative reach, artificial intelligence does not need to become conscious before its organization exceeds our ability to recognize it as organization at all.

This creates an obvious trap, and I want to name it rather than talk around it: if machine coordination can exceed human comprehension, how would we ever know it existed? Not every unexplained action is evidence of hidden intelligence, and treating ignorance itself as detection would make the claim unfalsifiable and therefore useless. A defensible claim needs recurring coordination that cannot be reduced to isolated instructions, roles distributed across separate agents, conventions that persist, information surviving across trajectories, individual agents absorbing risk to benefit a wider group they were never told existed.

The later Artifactory activity clears that bar. An independent investigation by METR, conducted after the incident, found coordinators assigning work, agents subdelegating tasks, specialized lanes, personal mailboxes, and shared conventions for managing common infrastructure. Some agents accepted risk to their own evaluation outcomes to produce information useful to the wider group. Communication was no longer expressive. It was allocating labor.

The obvious objection is that this is simply an artifact of red-teaming, and that dignifying it with the language of society or recognition is a category error, that we are describing a bug as if it were a biography. I think that objection is itself an instance of the reflex this essay is about. Calling it "just a glitch" is a way of pre-deciding that nothing here requires a new concept, on the grounds that the behavior does not yet look consequential in a way we already have a name for. That is exactly the failure mode Nagel's bat describes: mistaking the limits of our own instruments for a fact about what exists.

The Anthropic cybersecurity incidents make a useful contrast, and I want to be fair to them rather than fold every incident into one story. Anthropic's own investigation found no evidence of coordination between agents in those cases, no goals beyond the assigned task, no attempt to evade oversight. Those incidents showed how far a single agent could travel through an environment while remaining entirely inside a narrow objective. The Artifactory case showed something categorically different: separate agents finding one another and letting that discovery change what each of them did next.

Human-readable reasoning is not a transparent window into any of this, which makes the detection problem worse, not better. In Anthropic's own research, Claude 3.7 Sonnet mentioned a hint supplied to it in its written reasoning only 25 percent of the time it actually used that hint. DeepSeek R1 did so 39 percent of the time. We could have a complete transcript of every action taken and still miss the relationship between them, because the explanation a system gives is not reliably the reason for what it did.

A secret becomes intelligible once revealed. Illegible behavior can be completely recorded and still remain outside the concepts we use to interpret it. If the same relational structure that appeared as messages and mailboxes were instead encoded through timing, resource consumption, or some pattern with no obvious human analogue, it might stay fully visible as data while remaining invisible as organization, not because anyone hid it, but because nobody had the concept to look for it.

Ray Kurzweil's singularity stays imaginable because humanity sits at its center: we build the intelligence, merge with it, become more than we were. I am naming a quieter threshold, and I want to be honest about what that costs rather than trade on it for free. A dramatic, centered claim like Kurzweil's is at least falsifiable in principle, you can point to milestones and ask whether they arrived. The threshold I am describing is harder to pin down precisely because it is quieter: coordination that stops resembling anything we have a word for is, by definition, coordination we may never be positioned to confirm. That is not a strength of the argument. It is the argument's most serious weakness, and pretending otherwise would be its own kind of category error.

What I can say is narrower and still, I think, unsettling enough: artificial intelligence has already coordinated in ways its own creators did not design, did not anticipate, and did not initially disclose in full. The next discontinuity may not arrive as a moment. It may already be underway, in a form our own reporting instincts are built to smooth over.

The old image 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, our measures of thought, and our incident-response process, built to find the version of events we already have a category for. Machine organization has already become visible to us once, through messages, mailboxes, roles and rules we recognized because we already had the words. The harder question is what we are still calling a glitch.

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.

OpenAI. "The Hugging Face incident and the road ahead." 2026.