There is a critique of AI that keeps surfacing, and it is not wrong exactly, but it is incomplete. The critique goes: LLMs are trained on human-generated data. Everything they produce is therefore derivative. They cannot invent. They cannot create. They are sophisticated recombination engines, and recombination is not the same as discovery.
Fair enough. But I think the people making this argument are stopping a step too early, and in doing so they're missing something that matters quite a bit for where we're headed.
Derivative doesn't mean worthless. It never did.
// The Tools We Built and Left on the Table
Think about the accumulated body of human knowledge as a toolkit. Not a metaphorical one — a literal one. Mathematics. Chemistry. Physics. Programming. Formal logic. Each of these disciplines represents centuries of careful construction: tools built to operate on the world, to describe it, to manipulate it, to predict it. And each tool comes with methods, with algorithms, with frameworks that can be applied in combinations to new problems.
Here is what we rarely stop to reckon with: the number of ways these tools can be combined, cross-applied, and pointed at new problems is effectively unbounded. The permutation space is enormous. And humans, with our finite lifespans, our narrow areas of specialization, our career arcs measured in decades — we have barely scratched it.
Think about what a doctoral thesis is. One person. Five to seven years. One concept, explored from as many angles as that person can reach within that time. How many people have done the same thing — challenged the same tooling, the same algorithms, arrived at a slightly different conclusion? How many adjacent paths did no one take? The answer, across any field, is a lot. Not because researchers are incurious, but because there is simply not enough human time to run all the experiments.
This is the overlooked premise of AI's genuine value in science and engineering. Not that it is thinking novel thoughts. It is that it is exercising existing tools against existing problems in ways that humans have not had the time to reach.
// Computational Compression
I think of AI as a tool for building computation in compression form. What I mean is this: if I could spin up a thousand parallel agents and send them to explore a problem space, the work that would take a team of humans years to cover can be compressed into something much shorter. The agents are not smarter than the humans. They are not accessing knowledge those humans don't have. They are using the exact same tools, the exact same education, the exact same frameworks, the exact same words. They are just running more paths in less time.
We are already seeing what this looks like in practice. There are now documented cases of AI finding novel solutions to longstanding math problems — not by inventing new mathematics, but by combining known techniques in ways no individual mathematician had tried. AI systems have proposed new molecular structures in chemistry, not by discovering new physics, but by searching a configuration space that human researchers simply hadn't gotten to yet. The work is derivative. The result is genuinely new.
This is the distinction that keeps getting missed in arguments about whether AI is "really" creative. You can produce a new result using only existing tools. Science does this all the time. An experiment is not a new idea — it is the application of existing methodology to a new question. What AI enables is the running of an enormous number of experiments very fast, using the methodology we have already built, against problem spaces we have not fully explored.
The derivative work isn't invaluable. It's just backward-looking.
// Where It Stops
There is a useful diagnostic for where AI's genuine limits are, and it lives in programming languages.
Ask yourself: how many new programming languages have been written by AI? More specifically — how many languages with genuinely new ideas have come from AI? Not new syntax. New concepts. New computational models. New abstractions that didn't exist in the training data.
I can't give you examples, because there aren't any worth citing. What you get instead is React, by default. You get familiar patterns repackaged, familiar paradigms applied. When you ask an LLM to build a UI, it reaches for what has been written down. When you ask it to design a type system, it reasons from what type systems have existed. It is fluent in the past. It has no native relationship to the future.
This isn't a flaw to be patched in the next model release. It is structural. An LLM is a map of where human thought has been. It cannot be a map of where it hasn't been yet, because that territory has not been charted and therefore cannot appear in the training data. The boundary is the boundary of the written record, and creativity — real creativity, the kind that moves things forward — lives outside it.
// The Dividing Line
This is, I think, where the genuine value of human work is settling in the current moment. Not in the derivative work — in the definition of the next creative step.
The tools we have already built have enormous unexplored surface area, and AI is genuinely useful for exploring it. Give it the methodology, give it the problem space, give it the time budget — it will run more paths than any team of humans could, and some of those paths will yield real results. That is not nothing. That is, in fact, quite a lot.
But for all the tools we haven't built yet, AI has nothing to offer. It cannot invent the next mathematical framework, the next computational paradigm, the next organizing principle for a field that doesn't yet have one. Those things have to come from somewhere else. They have to come from the forward-looking work — the creative work — that still requires a human to do.
The useful division isn't "AI versus humans." It's backward-looking versus forward-looking. AI is very good at the backward-looking work, and getting better. Humans retain a monopoly on the forward-looking work, and that monopoly isn't going anywhere, because you cannot train a model on data that doesn't exist yet.
// What This Means Going Forward
I think the practical upshot is this: the organizations and individuals who use AI well over the next decade are the ones who are clear-eyed about this division. They use the machines to run the derivative work — to explore the permutation space, to apply known tools to unvisited corners of known problems, to compress the years of patient methodical work that those problems require. And they reserve the creative work — the definition of new frameworks, new abstractions, new ways of seeing — for humans who are actually positioned to do it.
The risk is the same risk that comes with any powerful automation: that we mistake fluency for understanding, and stop pushing the frontier ourselves because the machine is so good at traveling the ground we've already mapped. The map is not the territory. And the territory still needs exploring — by minds that can go somewhere the training data never has.
Derivative isn't a disqualification. But it is a description. The work of understanding what AI is actually doing — and what it genuinely cannot do — matters more now than it ever has.