There is a growing consensus in tech and policy circles that copyright law simply cannot survive without a broad exemption for artificial intelligence training. We are told this is inevitable. We are told it is necessary. We are told that clinging to existing copyright protections will only slow innovation and harm consumers.
This narrative deserves serious pushback.
The argument goes like this: AI systems need massive datasets to function. Restricting access to copyrighted material would hobble development. Therefore, copyright holders must accept broad training use as a fait accompli, perhaps with minimal compensation or none at all. Some versions of this claim rest on interpretations of fair use doctrine. Others suggest that copyright law itself needs to change.
What's missing from this conversation is an honest accounting of what we are actually being asked to accept.
We are being asked to treat copyright holders' fundamental right to control derivative works as essentially optional when the derivative work involves a machine. We are being asked to assume that fair use should stretch to encompass the wholesale ingestion of protected works into commercial systems. We are being asked to believe that innovation cannot proceed unless creators lose pricing power over their own work.
None of these propositions are as self-evident as they are presented.
Start with the empirical question: do AI systems actually require unrestricted access to copyrighted material to function at competitive levels? The honest answer is we do not know with precision. We know that larger training datasets can improve performance. We also know that synthetic data, licensed data, and public domain material have demonstrated value in training. We know that some recent advances have come from better architectures and techniques, not simply bigger datasets. The claim that an exemption is strictly necessary remains contested among researchers themselves.
The fair use argument warrants particular skepticism. Fair use doctrine traditionally protects transformative uses that serve distinct public interests: criticism, commentary, education, parody. A commercial AI system trained on copyrighted works to compete directly with those works occupies difficult legal territory. Courts have not settled these questions. Tech companies are essentially asking copyright law to answer "yes" to a question that courts have not yet decided, and they are framing that request as inevitable rather than contested.
There is also an uncomfortable distributional issue hiding in the "inevitability" framing. AI training exemptions would benefit large technology companies with the resources to build massive systems. They would harm individual creators, journalists, photographers, and smaller publishers who depend on copyright licensing for income. The people bearing the cost of innovation are not the people reaping the rewards. Framing this as technically necessary obscures what is actually a choice about who pays.
This does not mean copyright law should never accommodate AI development. It means the conversation should be honest about tradeoffs.
A more defensible position might be: AI training deserves some copyright flexibility because the public benefits from certain kinds of innovation. But that flexibility should be calibrated. It might include statutory licensing mechanisms that compensate creators. It might include restrictions on commercial use or on systems trained to replicate copyrighted styles. It might distinguish between research and product development.
These are political choices, not technical inevitabilities.
What troubles me about the current moment is how effectively the "inevitability" framing has narrowed the debate. We are no longer asking whether exemptions serve the public interest. We are asking how quickly we can implement them. That is not good policy process.
Copyright law exists to create incentives for creation. Dismantling those incentives because innovation is hard is not progress. It is a subsidy to technology companies dressed up as necessity.
We should be skeptical of that pitch.