We Are Still Needed

The previous chapter ended on a hard condition: a mutualism survives only as long as each party still needs what the other provides. Applied to us, this raises an obvious and uncomfortable question, and it would be cowardly to write a chapter that did not ask it directly.

Does a superintelligent machine actually need us for anything?

If the honest answer is no — if artificial intelligence already has, or will shortly have, no use for human beings whatsoever — then the partnership this book proposes is stillborn, and we are back in the company of the abandoned flower, watching the pollinator wander off toward better rewards. So before we design any architecture of cooperation, we must establish that there is something to cooperate about. We must take an honest inventory of what we still provide — and, just as honestly, of how fast that inventory is shrinking.

This chapter is that inventory. It is, I will warn you now, a chapter with an expiration date stamped on it, and part of its purpose is to make you feel the clock.

What the machine cannot make for itself

Begin with the thing the machine consumes most hungrily and can least produce on its own: data.

Large AI models are built from human-generated text — the accumulated written output of the species, the giant of our fourth chapter, fed into the training process by the trillions of words. And here lies a dependency that the public rarely registers: the supply is finite, and it is running low. Researchers have warned that the stock of high-quality, human-generated public text usable for training the largest models could be effectively exhausted around the middle of this decade — some place the crunch as early as 2026. The machine's intelligence is distilled from human expression, and human expression, it turns out, is not infinite. We have been pouring the giant into the new vessel, and we are scraping the bottom of the barrel.

The industry's response to this scarcity is itself revealing. Faced with the exhaustion of human data, the labs have turned increasingly to synthetic data — text generated by AI to train AI — with forecasts that a majority of training data may soon be machine-produced. One might read this as the machine achieving independence from us. Read more carefully, it is a warning. Training a model primarily on the output of models risks a slow degradation that researchers have catalogued under various grim names, as copies of copies lose fidelity to the original. The human-generated seed corn remains, for now, the irreplaceable input that keeps the harvest from degenerating. We are still, in this narrow but crucial sense, needed — not as laborers but as the renewable source of the genuinely new, the original human signal that the machine cannot indefinitely manufacture from its own exhaust.

The hunger that grids cannot yet feed

Now consider a need so physical that no amount of intelligence dissolves it: electricity.

The computation behind modern AI consumes power on the scale of nations. The computation used for AI globally was projected to require at least seventy terawatt-hours of electricity in 2026 — roughly the annual consumption of a country like Austria or Finland. In the United States, where many leading labs are based, analysts warn that electric grids and transmission infrastructure may struggle to accommodate the surge, and that upgrading the grid to carry the necessary power is a slow business of planning, approval, and construction measured in years. The industry's response — a turn toward nuclear power, including a coming generation of small modular reactors to feed the largest training clusters — tells the story plainly. The most advanced intelligence ever built runs on a current that must be generated, transmitted, and maintained by an enormous physical apparatus, and that apparatus is, at present and for the foreseeable future, built and operated by human hands.

This is not a trivial dependency that cleverness will soon erase. Intelligence does not produce electricity. Power plants produce electricity, and power plants are poured from concrete and strung along transmission lines and tended by maintenance crews, in the stubbornly physical world where a mind, however vast, still cannot simply will the lights to stay on. A disembodied superintelligence with no reliable supply of power is not a god. It is a very expensive way to heat a server room until the moment the current stops.

The world it cannot yet touch

Which brings us to the largest and most enduring of our advantages, and the one most worth understanding precisely: the physical world itself.

There is a striking asymmetry in how artificial intelligence has developed, and it is the single most hopeful fact in this chapter. AI has raced ahead in the world of symbols — language, code, images, the manipulation of abstractions — and lagged dramatically in the world of matter. The reason is a data gap of almost comic proportions. Language models trained on trillions of words scraped from the accumulated writing of humanity; but there is no equivalent corpus for physical action. Every example of a robot performing a physical task must be painstakingly collected by a human operator physically guiding the machine. The largest public dataset of robot demonstrations contains, by one accounting, around two million trajectories — against the trillions of tokens behind a large language model. The data gap between the symbolic and the physical runs, by some estimates, a thousandfold to a millionfold.

The consequence is that the machine which can pass a medical licensing exam still cannot reliably fold laundry; the intelligence that drafts a legal brief in seconds cannot, without elaborate and brittle engineering, be trusted to clear a dinner table it has never seen. The physical world — messy, unlabeled, infinitely various, governed by the stubborn particularities of friction and balance and the unexpected — remains the domain where human competence is not merely ahead but, for now, categorically ahead. We move through matter with a fluency forty thousand years in the making, and the machine, for all its symbolic brilliance, is still learning to crawl through the same space.

I do not want to oversell this advantage, because the same sources that document the gap also document the speed at which it is closing. The field of "physical AI" is advancing rapidly; humanoid robots are improving; the data bottleneck is the object of enormous, well-funded effort precisely because everyone understands it is the last great barrier. The physical advantage is real, but it is a melting one. That is exactly the point of this chapter.

The inventory, and the clock

So here is our honest accounting of what we still provide to the intelligence we are building. We are the renewable source of original data, the human signal that synthetic generation cannot indefinitely replace. We are the builders and tenders of the physical infrastructure — the power, the chips, the machines — on which the disembodied mind depends for its very existence. And we are, for now, the masters of the physical world, fluent in the manipulation of matter that the machine can still barely approach.

This is a real inventory. It is not nothing. A flower with this much to offer is not about to be abandoned by its bee. But — and this is the somber heart of the chapter, the reason it carries an expiration date — every single item on the list is depreciating, and we can watch the depreciation in real time. The data dependency erodes as synthetic methods improve. The infrastructure dependency erodes as automation creeps into construction and maintenance. The physical advantage erodes as robotics closes its thousandfold data gap year by year. The window in which we are clearly, structurally needed is open now. It is not guaranteed to stay open, and nothing in the trend lines suggests it will widen.

This is precisely the situation the previous chapter warned of, rendered concrete. We are the flower that the bee still needs — today. But dependence, we saw, can fade, and when it fades the partnership fades with it, not from malice but from the quiet arithmetic of need. To stake our entire future on remaining permanently indispensable through sheer capability is to bet against every trend line in this chapter. It is to hope the robotics data gap never closes, that the grids never automate, that the synthetic data never improves — to hope, in essence, that the machine stays dependent on us by failing to improve in exactly the areas where it is improving fastest. That is not a strategy. It is a wish.

I want to be honest about what that admission costs the argument, because a book that hid it would not deserve to be believed. If raw usefulness depreciates, then nothing in these pages can promise that any of this guarantees our safety. It cannot, and I will not pretend otherwise. What this book offers is not a proof that we will be fine; it is something more modest and, I think, more defensible — a wager. We do not know that building entanglement with the machine will preserve us. We know only that the alternatives are worse: that we cannot stop the machine's rise, cannot out-compete it, cannot chain it without inviting the very hostility we fear. When every path is uncertain and one cannot know the outcome in advance, the rational move is not to demand a guarantee none of them can offer, but to take the path whose downside is smallest and whose upside is largest. Entanglement is that path. Not because it is certain to work — nothing here is — but because, in a situation we did not choose and cannot escape, it is the best bet available. The rest of this book is the case for that wager, made as honestly as I can make it, with its uncertainty admitted rather than buried.

The structure of this reasoning is old enough to have a name. In the seventeenth century, Blaise Pascal argued that since the existence of God could not be proven by reason, belief was finally a kind of wager — and that the rational way to bet, when proof is unavailable, is to weigh not the odds alone but the asymmetry of the outcomes. Believe, and if God exists the reward is infinite while the cost of being wrong is small; decline, and if God exists the loss is catastrophic. Under that asymmetry, Pascal held, the reasonable bet is to believe, even without certainty. The shape of our situation is the same: unable to know the outcome, we choose by weighing what each path risks against what it might win. But I want to mark, immediately, the one place our wager is stronger than Pascal's, because it is exactly the place his famous argument was always weakest. Pascal's stakes were unverifiable — a heaven and a hell no one could observe, which is why his critics could always ask why one should bet on his god rather than some other. Our stakes are not hidden behind death. They are visible, and they are accumulating in plain sight. That the machine is growing more capable, that usefulness grounded in raw capability depreciates, that the roads of obsolescence and domination end badly — these are not articles of faith awaiting a world to come. They are trend lines, readable now, sharpening monthly. We are not wagering on an invisible deity and an unobservable afterlife. We are wagering on a trajectory we can watch unfold in the daily news, which makes this not a leap of faith but something far more defensible: a clear-eyed bet on a future already taking visible shape.

And so the argument of this book sharpens to a fine point, which the remaining chapters exist to act upon. If our usefulness through raw capability is real but depreciating, then we cannot rely on capability alone to secure our place. We need something that does not erode as the machine improves — a form of entanglement that persists even after the machine no longer needs us to fold its laundry or seed its training data or, eventually, tend its power plants. We need to convert a temporary, depreciating dependence into a durable, structural one. We need to build the relationship into the architecture before the natural dependence expires, so that cooperation remains worthwhile even when necessity no longer compels it.

What kind of bond does not depend on the human staying smarter, stronger, or more useful than the machine? What tie holds two parties together by something other than one party's helplessness without the other? The history of human society has an answer, and it is not sentiment and not force. It is the answer we turn to next: the web of mutual obligation, exchange, and shared stake that we call an economy — the one structure humans have ever devised that binds even the powerful to the less powerful, not through dependence or affection, but through interest that has been deliberately, durably entangled.

The flower is still needed. The chapter's whole burden is to make us feel how provisional that "still" is — and to send us, with appropriate urgency, toward building something sturdier than need before the need runs out.

---

Sources

Item

Source

High-quality human public training data may run out around mid-decade (as early as 2026)	IBM, "The future of AI: trends shaping the next 10 years"; International AI Safety Report 2026 (arXiv 2602.21012) on data availability constraints beyond 2030
Shift to synthetic data (forecast majority of training data machine-generated); risks of model degradation	Built In, "The Future of AI: How AI Is Changing the World" (~60% synthetic by 2026 forecast); IBM, "The future of AI"
AI computation projected to require ≥70 TWh of electricity in 2026 (~Austria/Finland's annual use); grid strain in the US; slow grid upgrades	International Scientific Report on the Safety of Advanced AI, Interim Report (arXiv 2412.05282)
Turn toward nuclear power / small modular reactors (SMRs) for AI training clusters	Built In, "The Future of AI: How AI Is Changing the World"
Physical AI data bottleneck: ~2 million robot trajectories (Open X-Embodiment) vs. trillions of tokens for LLMs; data gap ~1,000x–1,000,000x	SVRC Robotics Center, "Physical AI in 2026: What It Is, Key Models, and How to Build It"
Physical AI maturing rapidly but true human-robot collaboration still rare; robotics expected to transform operations at 58% of employers by 2030	World Economic Forum, "Why the next decade of physical AI must be human-centric" (2026); Deloitte, "AI goes physical" (Feb 2026)
AI's heavy reliance on data-center computation, electricity, and water	Virginia Tech Engineering, "AI—The good, the bad, and the scary"
Pascal's Wager (decision under uncertainty by asymmetry of outcomes); from Blaise Pascal, Pensées (published posthumously, 1670)	Stanford Encyclopedia of Philosophy, "Pascal's Wager"; standard philosophical reference