The essay "The Private Capture of Public Genius" opens with a striking historical anchor: on January 24, 1956, AT&T signed away 7,820 unexpired patents, royalty-free, to any American firm that asked. The author uses this as a structural metaphor for what is happening right now with frontier AI training data, and the parallel is more precise than it first appears. Both cases involve the extraction of value from a commons built by many, concentrated into the hands of a few, and then repriced by legal or regulatory intervention. The question the essay poses, and partially answers, is whether the internet's contribution to AI training constitutes a public good being legitimately harvested or a common-pool resource being strip-mined without restitution.
The Economics of Training Data: Public Good or Common-Pool Resource?
The essay applies a standard Ostromian taxonomy to internet data: excludable versus non-excludable, rivalrous versus non-rivalrous. The frontier labs' legal position implicitly classifies the internet as a public good. Scraping does not destroy the original data; the blog post still exists after ingestion. By this framing, training runs are non-rivalrous consumption and the fair use argument becomes structurally coherent.
The author's counter-argument is more interesting than a simple rejection of this framing. The claim is that the text layer of the internet, the actual scraped corpus, is only one layer of a multi-layer stack. The contribution layer, the attention layer, and the integrity layer all interact to produce the conditions under which the text layer is continuously enriched. Generative AI output, produced at near-zero marginal cost, floods the attention layer and disrupts the incentive structure that motivates human contribution in the first place. The good is non-rivalrous in a static sense but catastrophically rivalrous in a dynamic one, once output volume crosses a threshold that overwhelms discovery and attention mechanisms.
This is a genuinely important distinction that most legal and economic analyses of AI training data miss entirely. It reframes the harm not as damage to existing data but as damage to the system that generates future data. The internet is not a fixed ore deposit; it is a self-replenishing ecosystem. You can argue that mining does not destroy the ore already extracted, but that misses the point if the mining operation also poisons the aquifer feeding the deposit.
The Attribution Problem and the Limits of Shapley Valuation
The essay's treatment of the attribution problem is technically serious and worth examining carefully. The author correctly identifies the Shapley value as the leading formal method for attributuing marginal contribution to individual training inputs. The Shapley value computes an input's average marginal contribution across all possible orderings of the training set, which gives it a theoretically attractive fairness property: it is the unique allocation satisfying efficiency, symmetry, linearity, and the dummy axiom.
The practical objections raised are sound. Exact Shapley computation requires evaluating the model across an exponential number of input subsets. For frontier-scale models trained on billions of documents over weeks, this is not merely expensive; it is computationally incoherent as a policy instrument. Approximation methods exist, including data Shapley variants using gradient-based proxies, but these introduce their own instabilities. The author notes that the same document in a different training set will yield a different Shapley value, which is correct and reflects a deeper point: contribution is a relational property, not an intrinsic one. The value of any single training document is a function of its interaction with every other document in the corpus.
There is a further technical complication the essay does not fully develop. Modern frontier models are trained with data mixing schedules, curriculum learning, and multiple training stages including pre-training, supervised fine-tuning, and reinforcement learning from human feedback. The notion of a single training corpus is already a simplification. Attribution across these stages, each with different data sources and different functional roles, compounds the measurement problem substantially. The essay's conclusion that individual attribution is not merely difficult but mathematically incoherent at frontier scale is well-supported, even if the full technical case is not spelled out.
The Corpus Royalty Proposal: Strengths and Gaps
The essay's policy proposal is a fixed percentage of gross revenue paid by frontier labs into a public fund, distributed equally to all eligible contributors. The author calls this the Corpus Royalty and frames it explicitly as restitution rather than tax, welfare, or charity. The legal concept invoked is unjust enrichment: value was extracted from contributors without consent or compensation, and the remedy is disgorgement proportional to the benefit received rather than the harm proven.
The equal distribution mechanism is justified by the failure of attribution rather than as a normative preference for equality. Since individual shares cannot be computed without reimporting the measurement problem, equal distribution is the only administratively coherent option. The Alaska Permanent Fund is cited as a structural precedent: a public resource generates private profit, and the surplus is distributed equally among eligible residents because no resident can claim a larger individual share.
Several aspects of this proposal deserve scrutiny:
- Jurisdictional scope. The essay proposes paying every eligible American. But the training corpus is global. Wikipedia, Stack Overflow, GitHub, and the broader web are not American resources. A royalty paid only to American residents would be a poor match for the actual distribution of contributors. The author does not address this, and it is a significant gap. International coordination mechanisms for this kind of distributed resource claim do not currently exist at any meaningful scale.
- Revenue base definition. Gross revenue from what, exactly? OpenAI and Anthropic operate across API access, enterprise contracts, consumer subscriptions, and increasingly inference-as-a-service. Defining the taxable base without creating obvious avoidance structures requires careful scoping that the essay leaves unresolved.
- The prospective versus retrospective problem. The Superfund analogy is invoked to justify retroactive liability for training runs that were legal when conducted. This is a reasonable precedent but courts have historically been reluctant to apply it without explicit legislative authorization. The essay acknowledges this implicitly but does not engage with the constitutional or administrative law questions it raises.
- Incentive effects on corpus quality. If the royalty is paid equally regardless of contribution quality or volume, it does not directly address the author's own concern about declining incentives for human contribution. A flat payment to all eligible residents does not specifically reward the people producing the high-quality technical writing, scientific discussion, and careful argumentation that makes the corpus valuable. The mechanism addresses extraction without directly addressing the replenishment problem.
The Bell Labs Parallel and Its Limits
The historical framing around the 1956 AT&T consent decree is the essay's strongest rhetorical move and also its most instructive limitation. The Bell case involved a regulated monopoly whose research budget was effectively subsidized by ratepayers through a guaranteed return on capital. The patent release remedied a specific harm: intellectual lockout of competitors. The remedy matched the wound. Access was granted because access had been denied.
The author acknowledges that the AI case has a different wound structure. Contributors are not locked out of the models; they are extracted from without compensation. The remedy should therefore be payment rather than access. This logic is sound, but the Bell parallel also highlights something the essay underweights: the 1956 decree worked partly because AT&T was a regulated entity with a clear rate base and a long-standing relationship with federal oversight. The frontier AI labs are not regulated monopolies. They operate in a competitive market, albeit one with substantial concentration, and they have no existing regulatory relationship that would make a royalty mechanism straightforward to implement and enforce.
The more apt structural parallel might be the music industry's transition through the ASCAP and BMI blanket licensing regimes, where the impossibility of per-play attribution for radio broadcast was resolved by collective licensing at the industry level, with distributions approximated by sampling rather than exact measurement. The essay does not engage with this precedent, which is a missed opportunity given how directly it maps onto the attribution problem at hand.
Where This Leads
The essay arrives at a conclusion that is harder to dismiss than most takes on this subject: the frontier labs have already conceded the principle by signing licensing deals with Reddit, News Corp, and the Associated Press. If the training corpus were genuinely a public good requiring no compensation, these deals would be unnecessary. The labs' own commercial behavior reveals their actual position on the value of training data, even as their legal filings argue the opposite.
The Corpus Royalty as specified is probably not the final form of any workable policy instrument. The jurisdictional, definitional, and incentive problems are real. But the underlying argument, that collective unattributable contribution has value, that the failure of precise attribution does not eliminate the obligation to compensate, and that the mechanism for compensation should reflect the collective nature of the resource rather than pretending to resolve what cannot be resolved, is analytically sound and significantly underrepresented in current policy discussions.
The printing press took two and a half centuries to generate the Statute of Anne. The author is right that the current compression of that timeline will not permit the same patience. The legal and economic frameworks for governing AI training data will be established, one way or another, within the next decade. The quality of the arguments made now will shape what those frameworks look like. This essay, for all its gaps, makes a serious contribution to that conversation.