If AI cannot be trusted in a classroom, why should it be trusted in orbit?

On September 9, 2026, Evan Hubinger, Anthropic’s alignment science lead, said on social media that he personally believed there was a greater than 10% chance that artificial intelligence could kill all of humanity within the next decade. He was responding to Jacob Coxon, a researcher who had just resigned from Anthropic after working on pretraining at both that company and OpenAI, and who wrote in his own resignation thread that “the people building AI earnestly believe that it could kill us all by the end of the decade.” Hubinger did not walk the claim back. He confirmed it, while noting that the danger he had in mind was not today’s models but a future superintelligence for which the industry does not yet have a solved alignment plan.

Coming from the safety lead at one of the two or three companies most responsible for the technology’s current trajectory, this was not a fringe warning. It was, in effect, an admission from inside the room.

The exchange was framed almost entirely around AI in the abstract: chatbots, agents, the eventual arrival of self-improving systems. But the domain where this risk is least abstract and most immediate is not a hypothetical future superintelligence. It is the set of systems already being integrated into nuclear early warning satellites, missile-tracking constellations and the command and control architecture that decides, in minutes, whether a detected launch is real. If a company’s own safety researchers are willing to say on the record that current AI development carries a double-digit chance of civilizational catastrophe, the question that follows is not academic: What happens when systems built by that same industry, with that same acknowledged uncertainty, are the ones watching for nuclear launches from orbit and feeding what they see directly into the decision loop of nuclear-armed states?

The nuclear-space nexus is where AI safety debate stops being abstract

This is the nuclear space nexus, and it is where the abstract AI safety debate stops being abstract.

The mechanism runs through space architecture itself. Satellite constellations, orbital processing units and ground relay stations are the physical layer that decides what a nuclear armed state believes is happening in the world. The AI systems currently doing that work are narrow, frozen weight models, convolutional networks for image classification, sensor fusion algorithms for combining radar and satellite feeds, nothing resembling the self improving superintelligence Coxon and Hubinger were warning about. That distinction is precisely what makes the space layer worth examining rather than dismissing. 

For decades, that layer worked on a simple “bent pipe” model: satellites captured raw infrared and radar data and beamed it down to Earth for human analysts to interpret. That model is being replaced. Modern space architecture is shifting toward onboard edge AI, with computer vision models embedded directly into satellite payloads deciding, in orbit, what counts as a missile plume before a human ever sees the raw image. At the same time, the shift from a handful of large geostationary satellites to proliferated low Earth orbit constellations of thousands of smaller satellites means no human team can parse the feeds directly. Data fusion platforms now do that work, aggregating radar, imagery and signals intelligence into a single real time track. Even a narrow, well tested model compresses the time available for human judgment, since the entire argument for automating this layer is that orbital and hypersonic threats move faster than people can react to. 

And while today’s systems are not the unaligned frontier models Coxon feared, they still introduce a structural risk: the same edge processing and data fusion architecture being built now is the substrate that a future self improving system, if one ever gained access to military networks, would inherit and exploit. A false positive in this setting is not a chatbot hallucination. It is a satellite misreading a solar flare or a glint of debris as a missile plume, and passing that reading up a chain of command that may have only minutes to decide whether to believe it.

A useful, if unlikely, point of comparison arrived days apart from the Hubinger exchange. On September 2, New York City Mayor Zohran Mamdani announced a one-year moratorium on generative AI for public school students through eighth grade, affecting nearly 600,000 children and disabling AI functions in dozens of previously approved classroom programs. The justification was not that the technology does not work. It was that the city could not yet be confident it belonged in a setting where the cost of getting it wrong, a child’s developing capacity to think for herself, was too important to risk on a technology still being evaluated. 

“Children need teachers and human connection in order to learn and in order to grow,” Mamdani said. “We hold an obligation to do the same” when it comes to AI’s role in that development.

It is worth sitting with the asymmetry this produces. New York City possesses centralized, enforceable authority over what software runs on its own school-issued devices, and used that authority to pause AI in a setting where the downside of error is developmental, not existential. Nuclear-armed states, by contrast, are integrating comparable AI systems into early warning and command infrastructure where the downside of error is civilizational, and there is no equivalent central authority capable of imposing a comparable pause. If the mere possibility of AI eroding a child’s critical thinking was sufficient grounds for New York City to disable the technology outright pending further study, the far larger and more immediate possibility of AI misreading a satellite feed and compressing a nuclear decision to seconds deserves at least the same precautionary instinct, applied to the domain where a mistake cannot be undone.

The honest complication is that a school district and the international nuclear order are not remotely comparable in their capacity to act on that instinct. A ban is enforceable when one authority controls the devices in question; it is close to unenforceable when the object being restricted is software that can be updated remotely over an encrypted satellite uplink, and when no state trusts another enough to grant the kind of intrusive access that verifying compliance would require. This is precisely the gap that existing international efforts are now trying, with limited success, to close.

International bodies efforts to establish global norms regarding AI and nuclear command

The most direct attempt sits at the United Nations. On December 1, 2025, the General Assembly adopted Resolution A/RES/80/23 on the risks of integrating artificial intelligence into nuclear command, control, and communications, passing 118 to 9 with 44 abstentions. The resolution calls on states to adopt and publish national policies affirming that AI-enabled NC3 systems will remain under human control and will not be capable of autonomously initiating a nuclear launch decision. That it passed with broad support is meaningful. That nine states voted against it, and that France, a nuclear power that itself insists human control must be preserved at every critical stage of a launch decision, was reportedly among them, says something sharper about how far declaratory consensus is from binding practice.

Alongside the UN process, the Summit on Responsible AI in the Military Domain, known as REAIM, has become the closest thing to a running multilateral forum on this question. Its third summit was held in A Coruña, Spain, on February 4 and 5, following earlier summits in The Hague in 2023 and Seoul in 2024. It was explicitly framed as an attempt to move from declared principles to “concrete, practical and realistic steps.” That framing was itself a quiet admission that the first two summits had not managed the transition. The 2026 outcome document was endorsed by only 39 states, down from more than 60 in 2024, a decline several analysts have linked to deteriorating relations among the major AI and nuclear powers rather than to any narrowing of the underlying risk.

What both processes share is the same structural weakness: They are norm-setting exercises operating in a domain where the thing being regulated, software, cannot be counted, inspected or verified the way a missile silo can. A state can comply with the letter of a resolution on human control while still deploying an AI system that filters, prioritizes and frames the sensor data a human ultimately sees, shaping the decision long before a human formally makes it. The requirement that a person retain final authority does very little if that person’s picture of reality has already been constructed by a model no outside party can audit.

How independent research bodies evaluate these UN blueprints

Independent research bodies evaluating these blueprints have converged on a similar diagnosis. In an assessment of the UNGA resolution published shortly after its adoption, the Observer Research Foundation concluded that the measure signals a genuine and broadly shared anxiety among states, but that it remains constrained by exactly the fault lines one would expect: the divide between nuclear and non-nuclear states, the reluctance of major nuclear powers to accept binding restrictions on capabilities they consider central to deterrence and the fact that the resolution’s core language, human control and oversight, was never defined with enough technical precision to be verified rather than merely declared. The same critique applies to REAIM’s outcome documents, which rely on states to describe their own compliance rather than submitting to any external check. Both processes are, in the assessment of the researchers who study them, aspirational rather than enforceable, useful for establishing that a problem is taken seriously, but not yet capable of constraining the behavior of the states whose behavior actually matters.

This is, in effect, the diplomatic community’s answer to the technical impossibility of a blanket prohibition. Recognizing this, the UN General Assembly and allied processes have instead pursued a narrower, more realistic strategy: isolating AI from specific, high-consequence use cases, nuclear launch authority chief among them, rather than attempting to regulate military AI as a single undifferentiated category. Resolution A/RES/80/23 is itself an example of this approach in practice. It does not ban AI from space or defence systems generally; it targets one function, the initiation of a nuclear launch decision and insists that function alone remain exclusively human. The strategy is sound in principle. Its weakness, as the evaluations above make clear, is that use-case segregation only works if the boundary between an isolated function and the surrounding system can actually be verified. 

None of this means the underlying comparison to Mamdani’s classroom ban should be read too literally. What it usefully exposes is the gap between the precautionary standard a single city was willing to apply to a developmental risk and the precautionary standard the international system has so far been able to apply to an existential one. The nuclear order, by contrast, continues to integrate AI into its most sensitive systems while the very researchers building the underlying technology are, in public, on the record, unable to rule out a double-digit chance of civilizational catastrophe. The core policy question this raises is not whether AI can be kept out of nuclear command systems entirely.

That ship has largely sailed. 

It is whether the international community can move from resolutions that describe the danger, and use-case boundaries that sound precise on paper, to verification mechanisms that actually constrain it, before the gap between declared principle and deployed practice is tested by a false alarm instead of a vote count.