Who Gets to Teach the Machine?

Artificial Intelligence, Intellectual Property, and the Next Captured Economy

Time and technology await no one. AI is developmentally an infant. An intelligence must acquire information before it can do anything useful with it. Humans learn by observing, reading, imitating, experimenting, being corrected, and synthesizing what they’ve encountered. Artificial intelligence does essentially the same thing through radically different machinery and at radically different scale.

But then comes the uncomfortable question: what is it legally and morally permitted to learn from?

The question sounds simple until we apply the same standards to ourselves. Human beings routinely learn from information they did not create. We read books, study paintings, listen to music, examine inventions, analyze arguments, imitate techniques, and combine what we have learned into something new. Education itself depends upon access to the accumulated knowledge of people who came before us.

No novelist begins without having read other novelists. No musician develops in complete isolation from other music. No engineer independently rediscovers the entirety of mathematics, physics, and material science before designing a bridge. Human progress is cumulative precisely because each generation begins with knowledge inherited from the generations before it.

Artificial intelligence complicates this familiar process because it learns differently—and, above all, because it learns at scale. A human being might read a few thousand books during a lifetime, while a machine can process quantities of information that no individual could consume in centuries. A person might study dozens of painters before developing a recognizable artistic style, while an artificial system can analyze millions of images. Scale changes the economic consequences of learning, even if it does not automatically answer the moral question surrounding it.

That distinction matters.

There is a difference between learning from a work and reproducing it. There is a difference between understanding a technique and passing someone else’s creation off as your own. There is a difference between analyzing publicly available information and acquiring private information that was never intended to be public. Those distinctions already exist in the human world, and artificial intelligence forces us to decide whether—and how—they should apply to machines.

Creators have legitimate interests in that discussion. Writers, artists, musicians, programmers, photographers, journalists, and countless others have spent years developing skills and producing work. The arrival of a technology capable of absorbing patterns from enormous bodies of human creation does not make questions of ownership, privacy, consent, attribution, or compensation disappear.

Neither, however, does invoking those concerns answer the policy question.

Because eventually someone has to write the rules.

The moment government begins deciding what information may be used to train artificial intelligence, under what circumstances it may be used, what licenses must be obtained, what compensation must be paid, what records must be maintained, and what penalties apply when those rules are violated, we have created something enormously valuable.

We have created a gate.

The question then becomes one familiar throughout the captured economy: Who is best positioned to control it?

The answer is unlikely to be the independent artist whose work the regulation was ostensibly written to protect. It is equally unlikely to be the programmer experimenting with a new model in a spare bedroom, the university researcher with a limited grant, or five engineers trying to turn an idea into a company. The institutions best equipped to navigate complicated regulatory systems are usually the institutions already possessing money, lawyers, lobbyists, political relationships, proprietary datasets, intellectual-property portfolios, and compliance departments.

A billion-dollar corporation can absorb a regulatory requirement that costs tens of millions of dollars.

A startup cannot.

This creates one of the great paradoxes emerging around artificial intelligence. Regulation intended to restrain the largest technology companies may instead become one of the strongest mechanisms for protecting them from future competition. The law would never need to say that only enormous corporations may develop advanced artificial intelligence, because such a prohibition would be politically indefensible and economically absurd. It merely has to make developing one legally expensive enough that only enormous corporations can afford to do it.

Suppose training a competitive artificial intelligence eventually requires negotiating licenses with thousands of rights holders, documenting the provenance of billions of pieces of information, maintaining records demonstrating regulatory compliance, undergoing periodic government-approved audits, employing specialized legal personnel, and defending inevitable disputes over whether particular information should have been included in a training set.

For Google, Microsoft, Meta, or OpenAI, that is a compliance department.

For three engineers with an idea, it may be a prohibition.

This is how barriers to entry frequently emerge in a captured economy. They are rarely announced as barriers to entry. Each individual requirement can be defended on reasonable grounds: consumer protection is reasonable, intellectual-property protection is reasonable, privacy is reasonable, transparency is reasonable, and accountability is reasonable.

But regulations do not exist individually.

They accumulate, and accumulated compliance costs do not fall equally upon everyone. The larger the institution, the easier those fixed costs become to absorb. The smaller the competitor, the greater those same costs become relative to everything else the organization possesses. Eventually regulation can cease merely governing competition and begin determining who is financially capable of competing at all.

Artificial intelligence introduces an additional complication because the regulated input is not steel, oil, capital, or land.

It is information.

We are approaching a policy debate over who may learn from humanity’s accumulated knowledge and under what conditions. That raises questions considerably larger than copyright. What constitutes publicly available knowledge? When does analyzing something become copying it? Can someone own a style? Can someone own a statistical relationship discovered by examining thousands of works? If an artificial intelligence encounters an idea expressed independently by ten thousand people, who owns what the machine learned from them?

More fundamentally, who owns the patterns contained within human knowledge?

There will not be easy answers to those questions, nor should concern about regulatory capture become an excuse for granting technology companies unlimited access to everything humans have ever created. That would merely create another path toward concentration.

Information itself has economic value. Companies possessing enormous proprietary datasets already enjoy advantages that potential competitors cannot easily reproduce. If the future of artificial intelligence depends upon access to immense quantities of information, then control over information can become a moat just as surely as regulation can.

This creates the possibility of capture from both directions.

Protect information too aggressively and companies wealthy enough to purchase enormous licensed datasets gain an advantage. Protect it too weakly and companies already possessing enormous quantities of user information gain an advantage. Create sufficiently complicated compliance requirements and companies large enough to maintain armies of attorneys gain an advantage. Allow unrestricted accumulation of proprietary training data and companies already controlling major digital platforms gain an advantage.

The relevant question therefore cannot simply be whether artificial intelligence companies should be regulated.

That framing is too shallow.

The better question is whether the institutions we construct will preserve competition while protecting legitimate individual rights. That requires distinguishing between protecting creators and protecting incumbents. It requires distinguishing privacy from ownership of knowledge. It requires distinguishing infringement from learning. Most importantly, it requires recognizing that every regulatory system creates incentives not only for those being regulated, but for those capable of influencing the regulator.

Artificial intelligence will continue to learn. The economic incentives behind it are too powerful, its potential uses too numerous, and human curiosity too persistent for technological development simply to stop.

The question is not whether we permit artificial intelligence to grow up.

The question is what kind of economic architecture we construct around it while it does.

We can protect creators without granting ownership over ideas themselves. We can protect privacy without making knowledge the exclusive property of corporations wealthy enough to purchase access to it. We can punish genuine infringement without constructing licensing systems so expensive that tomorrow’s competitors never reach the starting line. Doing that, however, requires something increasingly difficult in public policy: writing rules around principles rather than incumbents.

Because the greatest danger may not be that artificial intelligence learns from humanity.

The greater danger may be that, while deciding how AI is allowed to learn, we quietly construct the institutions that determine who is allowed to teach it, who is allowed to build it, and ultimately who is allowed to own its future.

And if we are careless, by the time artificial intelligence grows up, the economy surrounding it may already have been captured.