What we may owe the future if machines become better at science than we are.

There was once a profession called computer. Computers were people. Banks, observatories, militaries and research institutions employed rooms full of them to perform calculations by hand — skilled workers doing intellectually demanding, economically valuable work.

Then we built machines that were vastly better at calculation. The profession didn’t adapt. It disappeared.

Nobody considers this a tragedy. We didn’t preserve thousands of human computers so that long multiplication would remain a living human skill. The machine was better, cheaper and faster, and civilization moved on.

Now imagine that the thing becoming cheaper and faster isn’t calculation. It’s discovery.

Not today’s AI, which helps a scientist write Python, search the literature or clean a dataset. Imagine a mature successor: a system that reads the literature, generates hypotheses, builds mathematical models, designs experiments, operates robotic laboratories, interprets results and decides what to run next.

Give it a problem — find a room-temperature superconductor that can be manufactured for less than $20 per kilogram — and three days later it returns a process.

It works.

At that point, what exactly is a department of condensed-matter physicists for?

That question leads somewhere uncomfortable, and the usual reassurance doesn’t survive contact with it.

Chess survived the machine. Science may not.

When Deep Blue beat Kasparov in 1997, it seemed possible that something had been taken from chess. Instead, humans kept playing. Engines are now incomparably stronger than any grandmaster, and millions of people still play, watch and care. Nobody responds to a brilliant world-championship move by noting that Stockfish had something better.

Finding the objectively best move was never the whole point. Chess is partly about watching one mind strain against another. Running survived the motorcycle; weightlifting survived the forklift.

Science has a different economic structure.

A pharmaceutical company doesn’t primarily want to witness the beautiful human struggle that eventually produces a drug. It wants the drug. An aerospace company wants the engine. Society wants the cure, the battery, the theorem, the forecast.

If machines become overwhelmingly better at producing those things, the human scientist faces a problem the chess player never did. The machine isn’t merely better at the activity. It removes the economic reason for humans to become exceptionally good at it.

We already don’t understand what we use

There’s a reason to think this transition may be gentler than it sounds: we already tolerate enormous intellectual dependence.

No individual understands a modern technological civilization. Consider the device you’re reading this on. Underneath it sit quantum mechanics, semiconductor physics, photolithography, processor architecture, firmware, operating systems, cryptography, networking protocols, compilers and tens of millions of lines of software. You understand some slice of that stack. So does every engineer who helped build it. Civilization works because understanding is distributed across millions of people.

AI introduces a stranger possibility: that some of that understanding stops being distributed among people at all.

This isn’t purely hypothetical. In 1976, Appel and Haken proved the four-color theorem by reducing it to nearly two thousand cases and handing them to a computer. The result was accepted, and it made a generation of mathematicians uneasy, because no human had checked the argument end to end. When Thomas Hales proved the Kepler conjecture in 1998, his referees eventually reported that they were confident but couldn’t fully verify it; settling the matter took a further decade-long project to make the proof machine-checkable. AlphaFold has produced predicted structures for hundreds of millions of proteins, and biologists build on them daily without being able to derive a single one.

Now scale that up. Imagine a machine-generated proof occupying terabytes. Independent checkers verify every formal step. The theorem is correct, and other mathematics built on it behaves. But no human understands the proof from beginning to end.

Has it been proven? Probably.

Now imagine the same thing in physics: a theory that makes extraordinarily accurate predictions, confirmed repeatedly, used by engineers to build devices that work — while the conceptual machinery required to derive those predictions has grown so large and alien that no physicist holds it whole.

Would we refuse to use it? History suggests we wouldn’t. We would build with it.

And a strange division of labor might emerge, between discovering knowledge and making knowledge comprehensible. Some scientists would no longer work at the frontier, because humans could no longer occupy the frontier. Their work would be to travel as far toward it as human cognition allows: compressing machine results into approximations, intuitions, shorter proofs and explanations. The scientist becomes, in part, a translator between two intellectual worlds.

That is the philosophical version of the problem. The practical one is harder.

What happens when nobody learns the hard way

Suppose AI becomes ten times as productive as a good physicist. Then a hundred. Then a thousand.

Companies need fewer physicists. Institutes need fewer. Salaries and opportunities respond.

Now consider a talented eighteen-year-old deciding what to do with their life. Why spend fifteen years becoming a world-class theorist when the machines are already better, and will be dramatically better by the time you finish?

So fewer people begin. A decade later there are fewer postdocs. Then fewer professors. Then fewer people capable of training the next generation at the highest level.

Nothing dramatic happens. There is no announcement that humanity has abandoned physics. The field simply becomes more machine-mediated with every cohort. And the process feeds itself:

• Better AI reduces demand for human expertise.
• Less demand produces fewer human experts.
• Fewer experts make society more dependent on AI.
• Greater dependence makes maintaining human expertise even less attractive.

The equations will still be there. The textbooks will still be there. Every paper humanity ever wrote may be perfectly preserved.

That’s not the same as the capability still being there.

A blueprint is not a capability

People sometimes say humanity “lost the technology” to build the Saturn V. That’s not quite right. Newton’s laws weren’t forgotten and the documentation didn’t vanish.

What disappeared was the living industrial organism around the rocket. Production lines closed. Specialized tooling was scrapped. Suppliers moved on. Components stopped being made. Engineers retired, and manufacturing knowledge that had lived partly in documents and partly inside thousands of experienced people dispersed with them. When NASA engineers returned to the F-1 engine decades later, they had drawings — and still found it more practical to pull surviving hardware out of storage, scan it, and relearn the thing by doing.

You cannot fix that by finding a drawing and pressing Print. You have the information. You no longer have the capability.

The same thing happens in less glamorous places. An organization holds the source code for a critical system written decades ago. The code is backed up. Documentation exists. The binaries run. Then something breaks, and the people who actually understood the system retired years ago. The organization discovers, expensively, the difference between possessing an artifact of knowledge and possessing people who can work with it.

The source code was never the whole system. Neither is the textbook.

That distinction matters enormously if AI becomes the repository not merely of information, but of competence.

The GPS problem, scaled up

We already know how to respond to this in narrow domains.

Satellite navigation is easier and generally more accurate than a sextant and a star table. Yet the US Naval Academy, having largely dropped celestial navigation in the age of GPS, put it back into the curriculum.

Why teach an inferior method? Because GPS can fail. It can be jammed, spoofed or denied. The old skill acquires a value that doesn’t show up in an ordinary efficiency calculation. Its value isn’t how often is this the best way to navigate? It is: what happens on the day the better system isn’t there?

Now replace GPS with machine intelligence.

Advanced AI isn’t an ethereal substance floating above civilization. It runs on physical machines, and it requires semiconductor fabrication, accelerators, memory, networking, electrical generation, cooling, data centers, software infrastructure and highly specialized supply chains.

Weights can be copied, and open models make access far less centralized. That is real resilience. But the most capable systems may still depend on industrial resources that few organizations, and few countries, can independently reproduce.

That changes what scientific independence means. A country might have excellent universities, tens of thousands of engineers and generations of accumulated knowledge, and still find itself severely constrained if it loses access to advanced semiconductors, large-scale compute, or the models its research ecosystem has quietly come to depend on. Sanctions could do that. So could war, supply-chain collapse, or a sufficiently effective cyberattack.

The frightening scenario isn’t that every computer disappears. It’s more mundane: the machines keep working, but your machines don’t. And after a few generations of scientific automation, your ability to compensate depends on how much independent human competence you bothered to preserve while it still looked unnecessary.

We already pay people to be inefficient

This sounds absurd until you notice how much of it we already do.

Military reserves are expensive capabilities maintained precisely because ordinary demand cannot justify keeping them ready. Strategic industries are subsidized when importing would be cheaper. Airline pilots train, repeatedly, for failures that automation almost always prevents. Seed banks preserve genetic material that nobody currently has a commercial reason to plant.

In every case, efficiency points one way and resilience points another.

Perhaps advanced scientific expertise eventually belongs in that category — a cognitive strategic reserve.

Concretely, that could mean less than people fear and more than a gesture. A government might fund a deliberately small number of research groups on the explicit condition that they can operate under restricted-AI conditions, and audit that claim the way you audit a fire drill. Universities might keep a track — not the default one — in which students periodically solve hard problems unaided, and are assessed on it. Institutions might run clean-room replications: take a machine-derived result, reconstruct a human-legible path to it, and publish the compression as a first-class output rather than as pedagogy. Some scientific bodies might adopt a second mission alongside producing new knowledge: keeping knowledge humanly alive.

The point wouldn’t be to compete with AI, any more than celestial navigation competes economically with GPS. The point is that if the machine goes silent, somebody still knows where the stars are.

The three best objections

This argument has real weaknesses, and they are worth stating plainly.

Automation has historically created more work than it destroyed. Spreadsheets didn’t end accountancy; they changed what accountants do, and there are more of them now than before. The honest response is that this turns on whether human judgment retains a niche the machine doesn’t fill. If it does, scientific employment reorganizes rather than shrinks, and the whole worry dissolves. The scenario described here is the one where it doesn’t — and the fact that every previous wave of automation left a niche is evidence, not proof.

If AI is that capable, it can rebuild whatever we lose. True, and it defuses the mild versions of this concern. But a reserve is insurance against precisely the case where AI is unavailable to you. Insurance is worthless in the world where you didn’t need it. That is what insurance is.

A reserve of physicists cannot rebuild a fab. This is the strongest objection. If compute becomes unavailable, a thousand well-trained theorists will not conjure a semiconductor supply chain out of chalk. Human expertise isn’t a substitute for industrial capacity, and a reserve that pretends otherwise is a cargo cult. What survives is the narrower claim: a country that retains independent scientific competence degrades gracefully rather than catastrophically. It can triage, adapt, keep critical systems running and rebuild in the right order. That is a smaller benefit than carrying on as before. It isn’t nothing, and it’s cheap.

The uncomfortable question

For most of history, knowledge was scarce and human intelligence was the machinery that produced it. We may be approaching a reversal, in which machine intelligence makes scientific problem-solving abundant while independent human understanding becomes slow, expensive and increasingly unnecessary.

That would be an extraordinary achievement. It could also create a kind of fragility we have never had to manage.

The danger isn’t that AI turns hostile, or even that it makes a catastrophic mistake. The quieter possibility is that it works extremely well — so well that becoming a world-class human physicist stops making economic sense. Then one generation doesn’t. Then another. And eventually we find that humanity possesses more scientific knowledge than at any point in its history, while fewer and fewer humans are able to independently create it.

Somewhere there is a gifted eighteen-year-old wondering whether it’s worth spending fifteen years mastering a discipline in which a machine surpassed humanity before they were born.

From the perspective of efficiency, the answer may be obvious.

From the perspective of civilization, I’m not sure it is. Because the most important reason to understand something may turn out to be the one reason our economic systems are worst at pricing:

so that someone still understands it when understanding it is no longer necessary.