Every executive whose company produces content — publishing, media, design, code, research — is watching the AI copyright battle with a mix of frustration and uncertainty. On one side: the scrapers. On the other: the lawsuits. Two years in, nobody has a framework that works.

A new paper from MIT Sloan and USC Marshall proves why both sides are economically broken — and proposes a market mechanism that solves the problem without destroying either creators or AI progress.

The finding that matters most: strong copyright protection creates something the authors call an “originality penalty.” The more original and valuable your work, the more the AI learns from you, the more its output resembles yours, and the less your future work is worth. The creators who should benefit most from copyright are the ones hurt most by it.

Three Market Failures Copyright Cannot Fix

The AI copyright debate has been stuck on a single binary question: is training on public data fair use or is it infringement? Every lawsuit, every licensing deal, every legislative hearing comes back to that yes-or-no. This paper proves the binary itself is the wrong frame.

The authors model the AI training data economy as a market with three structural failures that copyright law cannot fix, regardless of how the courts rule. These three failures form a triangle. Copyright treats them as separate legal problems. The paper proves they are a single market design problem.

The Proposed Solution: A Market, Not a Law

The authors propose a data intermediary — think of it as an exchange for AI training data, managed by a neutral market maker. Creators submit their content. AI companies bid for access. The intermediary selects an optimal set of training data that maximizes the value of the model while fairly compensating the creators whose work contributed.

The key innovation is that compensation is not based on individual content sales — which would collapse under the substitution problem — but on a market mechanism that prices the marginal contribution of each creator’s work to the final model’s performance. Creators whose work shifts the model’s capabilities get paid more. Creators whose work the model already absorbed get paid less. The market discovers the price, rather than a court or a regulator.

This mechanism — which the authors call optimal data inclusion — simultaneously solves all three market failures. It eliminates the originality penalty because compensation reflects contribution rather than exclusivity. It overcomes the curse of precision because the market prices differentiation. It resolves the correlated content externality because the intermediary aggregates decisions rather than leaving them to individual creators.

What This Means in Practice

For content-producing organizations — publishers, studios, design firms, research institutions — the originality penalty means that blocking AI crawlers and asserting copyright may be strategically self-defeating. If your content is valuable enough that models learn from it, the copyright protection that gives you short-term leverage may accelerate the long-term substitution effect that reduces your value. The better strategy is to participate in market-based mechanisms that price your contribution, not exclude it.

For AI companies, the paper provides a rigorous economic argument for why voluntary data licensing deals are structurally fragile. A one-time payment to a publisher for access to their archive does not solve the correlated content externality — the next publisher’s deal changes the economics of the first. The solution is not bilateral licensing but multilateral market mechanisms that aggregate supply and demand.

For regulators and policy makers, the paper offers a clear alternative to the binary copyright framework. Instead of deciding whether training is fair use, design a market where the price of training data is discovered through competition rather than litigation.

Key Takeaways for Strategy and Policy Leaders

Thanks to the Researchers