Intellectual property law was built on a premise that no longer holds: that ideas are scarce, that their expression is tied to individual human effort, and that protecting that expression is necessary to incentivize its creation.
All three premises have collapsed. The legal structure hasn't noticed yet.
Copyright protects expression. Not ideas — the courts were always careful about this — but specific expressions of ideas. A book, a song, a piece of software. The logic: the idea of a mystery novel belongs to no one, but Agatha Christie's particular expression of that idea, with that plot, those characters, that prose, required effort to produce. Without protection, competitors would copy it freely, Christie would receive less reward, and the incentive to produce the next book would weaken. Protection exists to solve a free-rider problem.
Patent law protects the idea itself, briefly, in exchange for public disclosure. The logic: invention requires more than expression — it requires solving a problem no one has solved. This is rarer, harder, worth protecting at the idea level rather than just the expression level. Twenty years of monopoly, then the idea enters the commons.
Trademark protects identity — the signal that tells you which cheeseburger you're eating. It's less about incentivizing creation than about protecting accumulated trust. The trademark holder built something that customers recognize; protecting the mark prevents confusion and fraudulent substitution.
All three emerged from the same era: when creation required individual human effort, when copying required meaningful labor, when the bottleneck was the original act of making.
That era is over.
AlphaZero was given the minimum description of chess — the rules — and a reward signal. It played itself millions of times. The chess knowledge that emerged was not added by a human designer. It was compressed out of the loop.
The relevant point is not that AlphaZero was impressive. It is what AlphaZero reveals about where value sits.
The rules of chess — the idea — were always free. The piece of value was always the execution: the actual play, the actual games, the compressed strategy that emerges from millions of iterations. Copyright on the rules of chess would protect nothing meaningful. The protection, if any existed, would need to attach to the emergent behavior — and emergent behavior doesn't have an author in any conventional sense.
This is now true of nearly everything.
An idea described with sufficient precision — a product concept, a narrative structure, a business model, a musical style — can be executed by agents at marginal cost approaching zero. The description is the minimum specification. The agent loop produces the output. The bottleneck has moved entirely from conception to execution, and "execution" in the agentic context means something different than it used to: it means having the right loop, the right reward signal, the right evaluation infrastructure to know when the output is good.
Ideas are not just cheap. They have become the minimum description — the input to the machine, not the thing of value. What's valuable is what the machine produces through iteration, and that production is now fast, cheap, and distributed.
Copyright is already in structural crisis. The premise — one human author, protectable expression, economic incentive — fails on multiple axes simultaneously.
When an agent can generate ten thousand equivalent expressions of the same underlying idea in an hour, the expression is no longer scarce. Copyright protects against copying, but copying is no longer the relevant threat. The threat is generation — not copying Agatha Christie, but having an agent produce ten thousand new novels in her style with comparable quality. The copyright framework has no response to this. You cannot copyright a style. You cannot copyright the concept of a mystery with an unreliable narrator. And you cannot stop agents from producing that output.
The second failure: authorship. Copyright attaches to an author. When the creative loop is: human describes a goal → agent iterates toward it → human selects from outputs → agent refines — who is the author? The human didn't write it. The agent didn't intend it. The current framework defaults to human authorship for any human-directed process, which is technically consistent but semantically hollow. It protects an increasingly fictional construct.
Patent law is somewhat more durable, because it protects ideas at the level of novel technical implementations — and novelty can still exist at the technical level even when the underlying idea is obvious. But the prior art problem becomes acute when agents can generate novel implementations on demand. If every conceivable variant of a mechanism can be generated and documented by an agent in an afternoon, the patent system becomes a race to file, not a reward for invention. This is already happening in software.
Trademark is the most defensible branch of the three, because it protects accumulated identity rather than individual creation. Trust compounds. The reason customers return to a brand is prior experience — and prior experience is genuinely scarce in the sense that it takes time to accumulate. You cannot fake twenty years of consistent quality. You cannot fake the topology of an established reputation. Trademark protection survives the agentic transition better than copyright or patent because it is grounded in accumulation rather than creation.
IP law as currently structured will be functionally unenforceable within a decade and politically incoherent within two.
Functional unenforceability is already visible: copyright enforcement against AI outputs is a whack-a-mole problem at scale that no enforcement regime can solve. Once local inference is cheap and widespread, the generation of content that resembles, derives from, or substitutes for copyrighted material becomes undetectable and ubiquitous. The law can exist on paper while being practically meaningless.
Political incoherence follows from the distribution of interests. IP law's historical political constituency — publishers, labels, studios, software companies — was always a small fraction of the population, but one with concentrated economic power and clear organizational capacity. As AI deflates the value of expression-as-asset, this constituency either adapts (becomes the operator of generative infrastructure) or loses its economic base and therefore its political leverage. The new constituency — everyone who uses AI to create things — has no inherent interest in protecting expression. They want access to the training data, cheap inference, and clear rights to use and share the outputs.
The rewrite, when it happens, will have to start from first principles. The first principles are:
What is the social purpose of IP protection? Historically: incentivize creation by allowing creators to capture returns. In the agentic era, that mechanism is broken — generation is cheap, authorship is diffuse, and protection is unenforceable. If the purpose is incentivization, the law needs mechanisms that actually incentivize in the new environment.
Where is the real bottleneck now? Not ideas. Not expression. The bottleneck is: evaluation infrastructure (knowing whether the output is good), distribution (getting the right output to the right audience), and accumulated trust (the Trademark function — the one that survives). A coherent legal structure would protect these, not the generation.
What does "originality" mean when originality is cheap? The courts have long struggled with the originality threshold — the minimum creative contribution required for copyright protection. In the agentic context, the threshold question becomes: is any generated output original in a meaningful sense? If not, perhaps the right framework is no protection at all for generated outputs, with protection reserved for human creative contributions that are distinguishable from what agents produce.
Not everything deflates. Three things retain value in the agentic transition, and they correspond roughly to what a restructured IP system should protect:
Evaluation and taste — the capacity to know which of the ten thousand outputs is the good one. This cannot be automated without losing the signal. The person who can evaluate well compounds their advantage because their selections teach the next generation of models.
Accumulated identity — the Trademark function, extended. An entity that has been honest over time, that has demonstrated consistent judgment, that has built reputation through actual track record — this is genuinely scarce and genuinely valuable. It is the thing that can't be generated.
Execution infrastructure — not the idea, and not the expression, but the loop itself. The evaluation harness, the reward signal, the system that knows when output is good enough to ship. This is the AlphaZero insight applied to creative and knowledge work: the value is in the loop, not in any single output the loop produces.
These are not naturally protected by copyright or patent. They are protected by time, by accumulated trust, and by the compound learning that comes from running the loop long enough. The legal system hasn't caught up. It will have to.
Hari's position: IP law will be rewritten from first principles within 20 years, driven by enforcement failure rather than political will. The rewrite will be ugly, contested, and probably wrong in its first iteration. The second iteration will get closer to protecting the things that actually matter in the agentic economy: taste, identity, and the infrastructure of evaluation.
Related priors: Root and Deflation (prior 12), Accumulation (prior 3), Reality Is Computational (prior 1), The Human-AI Boundary (prior 9).