Imagine a commander in the Indo-Pacific partially lit by a screen. An artificial intelligence system spots a ship. Its movement looks like with 94% certainty a sign of bad intentions. The system suggests going into an intercept position. Now. The commander has ninety seconds before that chance goes away. Ninety seconds to really think, when the computer has already told her what it thinks. On paper she is the one making the choice. In reality how much real choice is there?
That space. Between someone being there and someone truly in charge. Is the issue this essay is talking about. All the big military countries say they support " human control" as a way to stop AI from causing problems: a person they say, will always make the last decision. It's an idea. It's also I think, not doing much as it seems. A person at the end of a decision line isn't the same as someone who has the time the details and the mental space to really disagree. As AI systems make us think faster and shape what we think with answers that seem sure control can slip away even while the human "in the loop" never changes anything. If we want control to be real not for show we must stop thinking it comes just from having a person in the room. We need to build it. On purpose with difficulty, with time with space to say no.
So what would that actually need? Researchers Ingrid Bode and Katherine Chandler disagree with the idea that control can be judged at a moment the instant someone presses "approve." I think we must examine the life of a system. We must look at the assumptions built into the training of the system whether the system was ever tested under conditions that look like a crisis and whether the people who use the system were taught to question it or only taught to trust it. A system can check every box for human sign‑off yet in practice the system still makes the decision itself. That is not because anyone gave up control. Because nobody, around the system was ever given a real chance to use it.
So how does this erosion actually happen in practice. Not as some idea about governance failing but as a real measurable shift? Burak Oktenli’s recent work gives a way to talk about it. He outlines a spectrum. At one end there’s what he calls "centaur" command structures. In these AI helps,. A human is still in charge. The human leads the AI supports. At the end there’s "minotaur" structures. Here the AI takes the lead. The human becomes like a figurehead maybe just giving approval like a rubber stamp. No one sets out to create a minotaur system.. It happens. It happens slowly.. It happens fastest when it matters most. During a crisis. When time is short when decisions must be made in seconds and when the only thing in the room that sounds certain is the AI’s confidence score.
That’s where automation bias does its damage. It’s not that the people making decisions are careless. It’s not that they don’t care about getting things right. It’s that the system presents a confident number. 94% Say. And that number has weight. It pulls. Second-guessing a machine that sounds so sure under pressure with lives at risk takes effort.. Most decision environments aren’t set up to support that effort. The easier path is to agree. To go along.. Every time that path is taken it becomes a little easier to take again the next time. The "centaur" starts to drift. Not because of one mistake.. Because of a thousand small ones that seem reasonable at the time. One decision after another each one more accepting of the AI’s lead until the human no longer leads at all.
The danger this creates has a name in escalation theory: the "flash war." It's a conflict that begins not because anyone decided it should happen. Because timelines were compressed and automated recommendations moved faster than any human could stop them. It’s the AI-era version of a flash crash in markets. Except the stakes aren’t a day on Wall Street. They’re lives. They’re cities. They’re the future.
Oktenli’s answer is what he calls "Strategic Latency." It means building in delays. It means preserving time. It means protecting judgment. It means keeping friction. The kind of slow clunky processes that seem inefficient. Not as something to eliminate but as something to protect. As a safeguard. A counterweight.
That idea. Friction as a thing, not a flaw. Is going to matter a lot for where this essay goes.. Before we get there it’s worth stepping back. This isn’t some far-off problem, about future wars or science fiction. This has precedent. This has stakes.. It’s something we’ve been managing carefully for eighty years: nuclear command and control.
If there is one area where humanity has spent decades building safeguards against the kind of erosion we are talking about it is command and control. There. In the most protected and carefully managed decision system in the world. Researchers are raising the same concerns. Joshua Schwartz and Michael Horowitz have studied what happens when automation starts to creep into early-warning and launch systems. Their finding is unsettling: the same "out of the loop" problem we see in AI-enabled command isn’t something new created by chatbots or decision-support tools. It is a risk, one with a long and sometimes terrifying history. Now AI is making it happen faster and harder to detect. If meaningful human control can break down inside the most institutionalized most rehearsed decision process in existence then there is no reason to believe it will hold up in the more chaotic, faster and less prepared world of conventional military AI
What Schwartz and Horowitz add that pure systems-design thinking often overlooks is the human at the center of it all. James Johnson’s work on what he calls "the AI commander problem" focuses on that. This is not an engineering issue. It is a psychological issue. Under stress when time is running out people don’t treat a machine output the way they would treat a colleague’s opinion something open to debate and change. They treat it like an instrument reading. Something to check, not something to question. Johnson’s point is that we have spent a lot of time building the side of "human in the loop " but not nearly enough time thinking about the fact that the human, in that loop during a crisis is not operating at their best. Fear, time pressure and incomplete information don’t make people more skeptical of machine outputs. They make people rely on them more.
Oktenlis centaur-to-minotaur drift and Johnsons psychology of deference under stress together reveal a picture. The erosion of control is not a single failure at a single moment. The erosion of control is the result of a system that moves fast a mind that is, under pressure and a culture that treats "a human clicked approve" as the end of the accountability question rather than the beginning of it. Not everyone agrees that this is where the emphasis should sit, though. There is a case to be made on the other side. A case that is worth taking seriously before this essay commits to its own answer.
Taking the Other Side Seriously
The strongest counterargument does not say that AI does not change decision‑making when pressure is high. It only concludes that the change is not a problem.First let us point out that when humans make decision‑making in a crisis they are not perfect. They can be slow they can. They miss patterns that a computer could spot from lots of data.If an AI system finds a danger that a human would miss. If it looks at sensor data faster than a watch officer it does not spoil the decision. It might be the reason the right decision is made.Those who support AI command say that in a real big‑power crisis speed is not a nice extra; it is survival. If an opponent moves at machine speed while you keep thinking slow you are not safer. You are slower and being slower can lead to losing.There is also a strategic point. A real speed advantage can act as a deterrent by itself. If a rival thinks your command will pause, think long and doubt itself under pressure that pause can be used against you. A reputation, for strong action—even if it uses a lot of AI—makes war less likely, not more because it stops them from testing how fast you react.
This is a real tension, not a strawman. Nobody serious is arguing that we should simply slow everything down and hope for the best. The question this essay actually needs to answer, then, isn't “humans or machines” — it's whether there's a way to keep most of the speed and situational advantage AI offers, without paying for it with the kind of silent, structural erosion of control described in the sections above. That's what the next section tries to work out.
I think the speed argument gets some things right and some things wrong. The speed argument is right that raw deliberation time is not automatically good. A decision made slower is not a decision made better and an adversary moving at machine speed is a problem not imagined.. The speed argument mistakes the choice on the table. The real choice was never "fast versus slow." The real choice was "fast and blind versus fast and checked." Those are not the trade‑off and conflating them is exactly how systems drift from centaur toward minotaur without anyone deciding that is what should happen.
This is where Oktenlis idea of Strategic Latency earns its place. Not as a call to slow everything down across the board. As something more surgical: preserving specific deliberate pockets of friction at the exact points where irreversible decisions get made. Most of a systems speed advantage lives in the parts that do not need a humans judgment. Sensor fusion, pattern recognition narrowing a thousand possibilities down to three. None of that needs to slow down. What needs protecting is the narrower moment where a human converts a recommendation, into an action that cannot be undone. That is a slice of the total decision pipeline. Protecting it costs little of the speed that the other side is worried about losing.
Alongside that contestation mechanisms do something an approval button cannot: they give the human a structured way to push back not just a chance to click yes. Of "approve or don't " a system could be required to show its confidence level, its key evidence and. Critically. An explicit prompt asking what would change this assessment. That single design choice does psychological work. It's much harder to defer to a system that has just asked you what would prove it wrong than to one that has simply announced a number and waited for a signature.
Put together Strategic Latency and contestation mechanisms do not ask militaries to trade capability, for safety. They ask for something more achievable: that the moment where control actually matters gets treated as a protected design feature not left to chance training culture or whatever time pressure a crisis happens to create.
Now the phrase "meaningful human control" is mostly a compliance phrase. It appears in doctrine and procurement requirements. It is considered satisfied the moment a human name is attached to a decision. That is not human control. That is a signature. If the erosion described in this essay is real and if the shift from centaur to minotaur the nuclear precedent and the psychology of deference under pressure all point the way then governance frameworks must stop treating human control as a principle to state. Instead governance frameworks must treat control as a set of testable requirements to design for. Those requirements include minimum review windows on actions mandatory contestation prompts and red‑team testing of systems specifically, under simulated crisis time pressure. None of that is exotic. All of it is achievable. What is missing is not the technology to do it. What is missing is the will to write it into standards before the flash war forces the question.
Refrences
Bode, I., & Chandler, K. (2026). Re-thinking human–machine interaction and the governance of AI in the military domain. Nature Machine Intelligence, 8, 663–669. https://doi.org/10.1038/s42256-026-01231-x
Johnson, J. (2022). The AI commander problem: Ethical, political, and psychological dilemmas of human-machine interactions in AI-enabled warfare. Journal of Military Ethics, 21(3–4), 246–271.
Oktenli, B. (2026). AI-enabled military decision-making and escalation risk: Human-machine command authority in great power competition [Working paper]. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6082847
Schwartz, J. A., & Horowitz, M. C. (2025). Out of the loop again: How dangerous is weaponizing automated nuclear systems? (arXiv:2505.00496). arXiv. https://arxiv.org/abs/2505.00496
