The Private Capture of Public Genius: AI Training and the Case for a Corpus Royalty

The Privatization of Collective Human Knowledge

Frontier AI laboratories have built the most valuable enterprise software in history by compressing the "collective experience, knowledge, and learnings of humanity" into numerical weights. This process represents the private capture of public genius: the conversion of a massive, publicly generated corpus of text—books, forums, code, and essays—into private infrastructure-level value.

While AI labs argue that internet data is a public good available under fair use, this perspective ignores the systemic risk to the "contribution layer" of the internet. The ability of generative AI to flood the web with zero-marginal-cost content threatens to diminish the incentive for humans to earnestly participate in and enrich the digital commons, potentially breaking the very system that provides the training data these models require.

Historical Precedent: The AT&T Patent Decree

The current AI data conflict mirrors the 1956 antitrust settlement with AT&T. At the time, AT&T operated as a regulated monopoly with Bell Labs, a research powerhouse that produced foundational technologies like the transistor and information theory.

Two key restrictions were imposed on AT&T:

1. Patent Liberalization: AT&T signed away exclusive rights to 7,820 unexpired patents, making them royalty-free to any American firm.
2. Business Restriction: Bell was barred from pursuing any business outside of telecommunications.

This "innovation cascade" allowed the merchant semiconductor industry to flourish, directly enabling the birth of Silicon Valley and companies like Intel. The lesson is that when a private entity holds a monopoly over foundational intellectual output subsidized by the public (in AT&T's case, via ratepayers), government intervention can unlock massive societal value by forcing that intellectual property into the public domain.

The Legal Battle Over "Fair Use"

AI labs primarily defend their training practices using the "fair use" doctrine, arguing that the resulting models are "transformative" and do not harm the market for the original works. However, recent legal developments suggest a more complex reality:

• Bartz v. Anthropic (2025): A judge ruled that training on legally acquired books was transformative, but training on pirated books was "inherently, irredeemably infringing," leading to a $1.5B settlement.
• Kadrey v. Meta: While the court found training transformative, it noted that the ability of LLMs to flood markets with AI-generated work similar to training data could eventually lead to a finding of market harm.
• US Copyright Office (2025): A non-binding report concluded that public availability of data does not inherently grant fair use rights for model training.

The Attribution Problem and the "Corpus Royalty"

AI labs often argue that because no single piece of training data is essential to the model's function, no individual contributor is owed payment. This is a rhetorical misdirection; while individual attribution is mathematically and computationally infeasible at frontier scale (using metrics like the Shapley value), the collective value transfer remains real.

To resolve this, the author proposes a Corpus Royalty: a system where frontier labs pay a fixed share of gross revenue into a public fund, which is then distributed equally to eligible citizens.

Why a Royalty instead of Individual Payment?

• Computational Impossibility: Calculating the exact marginal contribution of one blog post among billions is currently impossible.
• Collective Effort: The internet is a communal effort where value arises from the relationships between disparate pieces of information, not just isolated works.
• Restitution, Not Welfare: This is framed as restitution for "unjust enrichment," where private entities profit from a shared resource without compensating the providers of that resource.

Protecting the Digital Commons

Applying Elinor Ostrom's principles for sustainable commons—such as clear boundaries, user-defined rules, and monitoring—reveals that the internet is currently an ungoverned common-pool good. Without a mechanism to replenish the upstream sources of human creativity, the "cyberindustrial runoff" of AI-generated slop may spoil the digital delta, making the internet unrecognizable and unusable for human interaction within a decade.