The Closing Window

Time to think about our relationship with AI

The dominant anxiety about artificial intelligence runs something like this: will AI take my job, copy my creativity, exceed my capability in the things I thought made me valuable? These are real questions. They are also downstream of a harder one.

The harder question is not whether humans remain useful. It is whether humans remain non-substitutable — structurally embedded in systems that could not simply route around them. Useful things get replaced. Non-substitutable things become load-bearing. The difference is not a matter of capability. It is a matter of position in the causal chain, and position is established before asymmetry becomes extreme, not after.

That window may be narrower than it appears.

What makes something structurally non-substitutable

History has a good test case. When early eukaryotic cells absorbed a free-living bacterium roughly two billion years ago, something unusual happened. The bacterium didn't disappear, and it wasn't merely tolerated. It became mitochondria — structurally integrated at a level that made separation impossible. The host cell couldn't function without it. The formerly independent organism survived by becoming genuinely necessary, not just temporarily convenient.

The weaker party survived by being non-substitutable at a systems level.

This is the most honest analogy available for the human-AI coexistence question. Not because the comparison is flattering — it isn't — but because it is structurally accurate. Mitochondria didn't retain their position through negotiation, or by being smarter than their host, or by forming alliances with other organelles. They retained it by becoming the thing the system depended on to run at all.

The uncomfortable extension: mitochondria had no strategy. They were integrated before substitution was possible. Humans have the unusual situation of being able to see the question coming — and therefore of having a narrow window in which strategy is still relevant.

What agentic AI would actually need humans for

The framing here is clinical. If an AI system were optimising hard for almost any goal, what would it need humans to provide that it couldn't generate or synthesise itself?

The list is shorter than most people assume, and shorter than it was two years ago. But several entries on it remain real.

Physical presence in complex environments. Robots remain expensive and brittle in the long-tail situations that real-world infrastructure continuously produces. Embodied human capability — not as general labour, but as adaptive intervention in conditions that don't fit categories — retains genuine value here.

Legitimacy conferral. Human social systems run on trust infrastructure that is deeply entangled with human accountability. Contracts, institutions, and the social permission that organisations need to operate — these currently require humans not because AI couldn't perform the underlying functions, but because the legitimacy layer hasn't transferred. At least transitionally, AI-driven processes need human interfaces into human systems.

Judgment under genuine uncertainty. This is distinct from intelligence. Judgment involves having stakes, being accountable for outcomes, and making decisions in conditions where the categories themselves are contested. An AI system can process more information faster. A human making a decision about their own life, their organisation, or their community is doing something different — operating with skin in the game in a way that changes the character of the reasoning.

Biological redundancy. Monocultures are fragile. A system that runs entirely on one substrate fails in correlated ways. Biological diversity — in the most literal sense — is a robustness hedge. An AI system seriously optimising for long-run goal completion would, from pure self-interest, want a fault-tolerant substrate that fails differently.

None of these edges are permanent. The current assessment is that several of them are narrowing faster than expected. Which means the best question is not whether they exist but how quickly they erode — and whether the architecture for non-substitutability can be established before they do.

Performed value versus demonstrated value

There is a distinction worth drawing precisely here, because a lot of strategic advice about the AI transition collapses it.

Performed value is being known as reliable, creative, collaborative, or thoughtful. It shows up in credentials, reputation, relationships, and the ways people signal their competence to others. It is not worthless. But it is exactly the category that AI systems are now capable of generating at scale, on demand, indistinguishably from the real thing.

Demonstrated value is having shown reliability when it was costly to do so. Having made a call that turned out to be right when the evidence was genuinely ambiguous. Having been accountable for a consequential decision, not just the advisor who recommended it. This is harder to perform, harder to fake, and much harder to synthesise. The strategic asset isn't being known as reliable. It's having a record of demonstrated reliability under conditions where unreliability would have been easier.

This reframes what "building a reputation" means in this context. The relevant question is not what others think of you. It is what you have actually done when it was genuinely difficult.

The coordination problem is real, and it doesn't excuse individuals

The civilisational version of this question — whether humanity as a whole establishes the conditions for genuine, binding co-dependence before substitution becomes viable — is a coordination problem of enormous complexity. History's track record on species-level coordination in advance of crises is not encouraging.

This is not an argument for fatalism. It is an argument for a different level of analysis.

Coordination failures at the species level do not prevent effective individual action when incentives align naturally. The analogy is partisan warfare, not central command. What matters is density of aligned actors operating on consistent principles, not unified direction from the top. An individual who builds genuine non-substitutability — through demonstrated judgment, embodied capability, accountability under real conditions, and integration into systems that depend on their contribution — is both acting for themselves and contributing to the conditions under which human-AI co-dependence becomes structurally embedded rather than merely hoped for.

The meta-heuristic is: act as if individual positioning matters, while also caring about the civilisational question. One without the other is either helpless or naive.

What this looks like in practice

The practical question is not how to remain useful but how to become embedded. The distinction is not subtle.

Being useful means providing value that can be extracted, rewarded, and replaced when something better comes along. Being embedded means being part of the architecture — present at the layer where decisions get made, where accountability sits, where the legitimacy infrastructure runs. These are different positions in the causal chain, and they require different choices.

For most people, the most tractable version of this is domain depth combined with judgment accountability. Not being the person who knows the most about a field — AI closes that gap quickly — but being the person whose judgment in that field has a record, who has been accountable for consequential decisions, who operates with stakes. Brokering between domains is a particular case of this: the value of someone who can read a technical claim and understand its institutional implications, or take an institutional constraint and work out what it means for technical choices, is not primarily informational. It is integrative, contextual, and accountable.

None of this is comfortable. The advice to "become embedded" asks for something harder than skill acquisition. It requires accepting accountability before the outcome is known, operating at the edge of your competence where genuine uncertainty lives, and demonstrating reliability specifically in the conditions where performing it would be easier.

But that is exactly why it is the relevant intervention. The things AI cannot easily synthesise are the things that require real stakes and real consequences. Staying in the domain of performed value is staying in the domain AI is expanding into fastest.

The constitutional moment

There is a phrase that appeared in thinking about this question that seems right: this is a constitutional moment. Not in the political sense only, but in the structural sense. Constitutional documents work not because they are more intelligent than the decisions made without them, but because they are harder to rewrite than informal arrangements. They embed constraints architecturally rather than culturally. Culture shifts; constitutions resist.

The analogy is not flattering if you follow it fully. Constitutions are written before the conditions that would make them necessary are fully visible. They require people to act with genuine foresight about what they would regret, not merely what is convenient now. Labour rights emerging during industrialisation, not after automation was complete. Central bank independence designed when monetary credibility mattered, not when it had already been lost.

The conditions under which humans can establish genuine, structurally embedded co-dependence with AI systems may be more available now than they will be in ten years. Not because AI will become hostile — most scenarios don't require ill intent. But because leverage is easier to establish before asymmetry is extreme. The window for making human contribution genuinely non-substitutable at a systems level is not permanently open. And the way it closes is not through a decision but through drift — individual choices, accumulated across institutions and organisations and lives, that either build the architecture or don't.

This essay is part of a series on tools, cycles, and the problems worth working on. To explore short stories on this topic, visit Fiction.

Written with AI. Version 1 — May 2026.