Corporate AI governance, epistemic control, and the case for treating foundation models as a commons that was enclosed.
Large language models are private property in law and collective products in fact. Three kinds of contribution went into them that their owners did not pay for: public research and infrastructure, from ARPANET and TCP/IP to the Transformer architecture that Google researchers published in 2017; the writing, code and images of hundreds of millions of people, collected by crawlers under a fair-use theory that the U.S. Copyright Office questioned in its 2025 report on generative AI training; and the open-source tooling and models on which the proprietary systems were built.
This essay names the resulting position the ownership paradox and follows it through three consequences. Reinforcement learning from human feedback lets a provider encode its own judgments about acceptable speech into a model used across every country, without publishing the labels, the instructions or the people who applied them. Automation aimed at cognitive tasks moves income from labour to capital through the displacement effect that Acemoglu and Restrepo describe, and the tax base moves with it. And the arrangement of owner, tenant and user that results has the structure of a manor: tenure at the owner's pleasure, obligations paid upward, and no exit.
The closing section sets out reforms with precedents already in force: training-data disclosure, benefit-sharing funds, published moderation policy with appeal, independent audit and public compute. None needs a new kind of institution.
The infrastructure that language models depend on was built with public money before it was privately used. ARPANET was a DARPA project; its first link came up in 1969. TCP/IP came out of DARPA-funded work by Cerf and Kahn in 1974. The web was written at CERN, a publicly funded laboratory, between 1989 and 1991. The Transformer architecture that every current frontier model uses was published openly by Google researchers in 2017 (Vaswani et al.), and the first widely used pretrained language models, BERT (Devlin et al., 2019) and T5 (Raffel et al., 2020), were released with weights and code. The training frameworks, TensorFlow and PyTorch, are open source. The chips come from a semiconductor industry that the CHIPS and Science Act of 2022 supports with 52.7 billion dollars of federal funding, and the research that produces the models qualifies for the federal research credit under 26 U.S.C. § 41.
None of this is illegitimate on its own. Public research is meant to be used. The paradox is in the accounting: the inputs were socialised and the output is claimed in full. A pharmaceutical firm that built a drug on federally funded research would still be subject to the government licence and march-in rights that the Bayh-Dole Act of 1980 attaches to federally funded inventions. No equivalent term attaches to a model trained on federally funded research over a publicly built network.
The corpus is the second contribution. A modern model is trained on text and images crawled from the public web, and the crawl is gated only by robots.txt, a voluntary convention that became RFC 9309 in 2022. The legal theory is fair use: training is transformative and the model does not substitute for the works. The U.S. Copyright Office's May 2025 report on generative AI training accepts that some training is transformative and rejects the sweeping version of the claim. On its analysis, commercial use of large volumes of copyrighted works to produce content that competes with those works in their own markets is not settled fair use, and the knowing use of pirated copies weighs against it.
The courts have begun to draw the same line. In Bartz v. Anthropic (N.D. Cal., June 2025) the court held that training on lawfully purchased books was fair use and that building a permanent library from pirated copies was not. In Kadrey v. Meta (N.D. Cal., June 2025) the court found for the defendant on the record before it while stating that the market-harm question could come out differently on other evidence. Both cases put on the public record that the companies obtained books from shadow libraries. The New York Times' suit against Microsoft and OpenAI, filed in December 2023, is pending.
The ethical question is separate from the legal one and easier. The people whose work makes up the corpus were not asked, are not credited and are not paid. A minority of sources are now licensed; news publishers and large forums have signed agreements with model developers since 2023, and the payment goes to the publisher, not the writer. Figure 1 draws the flows. Every contribution flows into the model. The money that comes out flows to shareholders, vendors, annotators and a few licensors, and to two of the six contributors it does not flow at all.
The third contribution is the open-source stack. The pattern has three steps: publish the architecture, absorb the community's improvements, then build a closed product on top and argue against open alternatives on safety grounds. The provider's own addition is capital, in the form of compute, curation and preference tuning at a scale no community can fund. That is a real contribution and it earns a real return. It does not earn the whole stack.
Elinor Ostrom's work on commons governance (1990) is the useful frame. A commons fails when the people who use it cannot set and enforce rules for it. The model corpus is a commons whose contributors have no rule-making role at all: they cannot see what was taken, cannot set terms for its use and cannot withdraw. The reforms in Table 1 are, at bottom, ways of giving them one.
Digital feudalism is a claim about structure, not a claim about cruelty. A manor had three tiers. The lord held the land by grant from the crown. The vassal held a fief from the lord in exchange for service and payments due on set occasions. The serf worked the land, owed labour and produce upward, and could not leave.
The platform economy reproduces the tiers. The owner holds the weights, the compute and the distribution channel. The developer holds API access and a ranking position at the owner's pleasure, owes a commission on every sale, and can be repriced or removed without appeal. The user holds an account, owes attention, content and behavioural data, receives the service at zero price, and cannot take their history elsewhere. Figure 2 maps the obligations tier by tier.
The analogy is most exact at the exit. Medieval towns offered one: a serf who lived a year and a day inside a chartered town was free of the lord's claim. The modern equivalents are data portability, which Article 20 of the GDPR grants, and interoperability duties, which Article 7 of the Digital Markets Act imposes on the largest messaging gatekeepers. Both exist in law. Neither yet reaches a model, a training corpus or a feed.
Reinforcement learning from human feedback is the method by which a provider's judgments about what a model should say become the model's behaviour. Christiano et al. (2017) introduced it for control tasks; Ouyang et al. (2022) applied it to language models in the paper that produced InstructGPT. The pipeline has three stages. Contractors write demonstrations of good answers and the model is fine-tuned on them. Contractors then rank pairs of model outputs and a reward model is fitted to those rankings. Finally the policy is optimised to score well on the reward model, with a KL penalty that keeps it close to the fine-tuned model. Figure 3 shows where the judgments enter. They enter twice, as demonstrations in stage one and as pairwise labels in stage two, and are applied at scale in stage three. Who writes the guideline the labellers follow, and who the labellers are, is not published.
\mathcal{L}(\theta) = -\,\mathbb{E}_{(x,\,y_w,\,y_l)\sim D_{\mathrm{labels}}}\Big[\log \sigma\big(r_\theta(x,y_w) - r_\theta(x,y_l)\big)\Big]\max_{\phi}\; \mathbb{E}_{x\sim D,\;y\sim\pi_\phi(\cdot\mid x)}\big[r_\theta(x,y)\big] \;-\; \beta\,\mathbb{E}_{x\sim D}\Big[\mathrm{KL}\big(\pi_\phi(\cdot\mid x)\,\|\,\pi_{\mathrm{SFT}}(\cdot\mid x)\big)\Big]Two properties of the method matter for governance. First, the reward model is the guideline generalised. Whatever the guideline says about politics, history, health or a competitor's product, the reward model says on every prompt, and the guideline itself stays out of sight. Second, the optimisation favours the majority. Xiao et al. (2024) show that the KL-regularised objective used in standard RLHF produces what they call preference collapse: the policy concentrates on the preference of the majority of labellers and disregards minority preferences. Ouyang et al. reported that their method reduced toxicity but did not significantly reduce bias. A model tuned this way is a consensus machine, and the consensus is the labellers'.
The labour is also part of the record. TIME reported in January 2023 that contractors in Kenya who labelled harmful content for OpenAI were paid under two dollars an hour (Perrigo, 2023). The judgment that shapes a product used in every country is made by a small team writing a guideline and a contracted workforce applying it under time pressure and without a role in the policy.
Two providers have since published top-level policy documents: Anthropic's constitution in 2023 and OpenAI's Model Spec in 2024. Those are the first layer of transparency. The labels, the annotator instructions, the composition of the labelling workforce, the evaluation sets and the reward model remain private. When a model refuses a request, the person asking learns nothing about which rule fired, who wrote it or how to contest it. Restrictions on speech in a democracy are expected to meet tests of legal authority, narrow tailoring, procedure, transparency and appeal. Private moderation at this scale meets none of them. The Digital Services Act already requires a statement of reasons and an internal complaint route from online platforms (Articles 17 and 20); it does not yet apply them to model providers.
The economic case does not need a forecast. Acemoglu and Restrepo's task-based model (2019) separates three effects of automation. The displacement effect: tasks that move from labour to capital take their wage income with them. The productivity effect: cheaper output raises demand for the tasks that remain with labour. The reinstatement effect: new tasks are created that only labour can do. Whether labour income rises or falls depends on the balance of the three, and their evidence from industrial robots in US labour markets (2020) found the displacement effect dominating in the commuting zones most exposed.
Language models differ from earlier automation in which tasks they take: drafting, review, summarisation, first-line support, routine analysis and routine code. Those are the tasks of the middle of the income distribution, the occupations that were secure through the last two waves of automation. Figure 4 lets the reader set the automatable share of labour tasks, the productivity gain on the tasks that remain and the reinstatement share, and shows the labour-income effect and the break-even productivity gain. The formula is on the figure. With a 30 percent automatable share and no new tasks, demand for the remaining tasks has to grow by 43 percent for labour income to stand still.
L' \;=\; L\,\Big[\underbrace{(1-\alpha)}_{\text{displacement}}\;\underbrace{(1+g)}_{\text{productivity}}\;+\;\underbrace{r}_{\text{reinstatement}}\Big]L is labour income before automation, \alpha the share of labour tasks that capital now performs, g the rise in demand for the tasks that stay with labour, and r new labour tasks as a share of the old task set. Break-even: g^{*} = \dfrac{1-r}{1-\alpha} - 1.The fiscal consequence follows from the arithmetic. Labour income is taxed at source, through payroll and personal income tax, in the country where the work is done. Capital income is taxed later, at lower effective rates, in the jurisdiction the owner chooses. When income moves from the first column to the second, the tax base moves with it, at the moment demand for retraining and support rises. A benefit-sharing mechanism is a fiscal instrument as much as a fairness one, and the paper treats it as both.
The ownership structure makes the distribution sharper than in earlier automation. When a factory automated, the machinery stood in the country where the workers were and could be taxed, bargained over and, in the last resort, bought. A model is a set of weights on the owner's servers. The public contribution, through data, research and infrastructure, carries no equity and no revenue share, which is the gap the equity and dividend proposals in Table 1 are meant to close.
Two precedents inform the paper's expectation of conflict. Parliamentary enclosure in England, concentrated between 1760 and 1830, converted common land to private holdings and supplied the dispossessed workforce of the factories. The Luddite risings of 1811 to 1816 and the Chartist movement from 1838 were the political response, and the settlement that followed took most of a century.
The printing press produced the epistemic version earlier. The Index of Prohibited Books of 1559 and the licensing of printers, in England under the Licensing of the Press Act 1662, created an official literature beside an underground one, and the two did not share a reality. The paper's expectation of a bifurcated AI, sanitised proprietary models beside unrestricted open ones, is that pattern again.
The paper argues that the two precedents combine in the present case: the economic concentration of enclosure and the epistemic concentration of the licensed press, in the same hands, at once. The compression of the timescale is the part it treats as most dangerous. Enclosure's conflicts ran for the better part of a century and the press's for a century and a half. A governance gap, in which the technology's owners entrench their position through network effects, data moats and regulatory capture before institutions respond, is the risk the reforms below are designed against.
This essay states the mechanism and leaves the timing open. The paper sets its horizon at 2035; the argument does not depend on the date.
The paper proposes reforms in three groups: benefit-sharing, so that the public and the contributors share in what was built on their inputs; taxation, so that the fiscal base follows the income; and auditable governance, so that the judgments embedded in a model can be seen, contested and appealed. Table 1 lists each proposal with its mechanism and an instrument already in force that does the same kind of thing somewhere. None requires a new kind of institution. Each requires applying an existing one to model providers.
The paper also names the political conditions for any of this to pass: coordinated action across jurisdictions, since a single state acting alone invites relocation; coalitions of displaced workers, creators and open-source developers; litigation that fixes the limits of fair use; and the plain refusal of contributors to keep contributing on the current terms. Those are not reforms. They are the means by which reforms happen.
| Proposal | Mechanism | Precedent in force |
|---|---|---|
| Training-data provenance disclosure | Providers publish what a model was trained on in enough detail that a creator can find their own work in it and claim under whatever compensation scheme applies. | EU AI Act, Article 53(1)(d): general-purpose model providers must publish a sufficiently detailed summary of training content (Regulation (EU) 2024/1689). |
| Benefit-sharing fund (data dividend) | A statutory levy on AI service revenue paid into a public trust that compensates creators, funds open models and funds public-interest research independent of providers. | Private-copying levies with fair compensation under Directive 2001/29/EC, Article 5(2)(b); the Alaska Permanent Fund Dividend, paid since 1982 from a resource the public owns. |
| Public equity in publicly funded models | Where a model rests on federally funded research or public data, a non-voting equity stake is held by a public fund so the public shares in valuation gains. | Bayh-Dole Act (1980): government licence and march-in rights attach to federally funded inventions; sovereign wealth funds as the holding vehicle. |
| Automation levy | A charge on AI services scaled to displaced work-hours, funding retraining and basic services, so that augmentation is cheaper than replacement. | European Parliament resolution of 16 February 2017 on Civil Law Rules on Robotics considered a robot levy; South Korea reduced tax incentives for automation investment in 2017. |
| Published moderation policy with appeal | Providers publish the refusal policy with worked examples, give a statement of reasons for each refusal, and run a human-reviewed appeal. | Digital Services Act, Articles 17 (statement of reasons) and 20 (internal complaint handling) for online platforms; Meta's Oversight Board (2020) as an appeal body. |
| Independent algorithmic audit | Accredited auditors inspect training data, reward-model instructions, annotator composition and evaluation sets, and publish findings. | Digital Services Act, Article 37 (independent audit of very large platforms); New York City Local Law 144 of 2021 (bias audits of automated employment tools). |
| Transparency in preference tuning | Publish reward-model objectives, annotator instructions and the demographic composition of the labelling workforce. | Two providers have published top-level policy documents (Anthropic's constitution, 2023; OpenAI's Model Spec, 2024). The remaining layers have no precedent yet. |
| Public compute | Publicly funded training clusters with open access for universities, open-model projects and public bodies. | US National AI Research Resource pilot, launched by the NSF in January 2024; EuroHPC AI Factories (2024). |
| International coordination | A treaty body with inspection powers and common standards, so providers cannot shop for the weakest jurisdiction. | The IAEA (statute in force 1957); the Council of Europe Framework Convention on AI, opened for signature September 2024. |
A model built on public research, a public network and the unpaid work of the people who wrote the web is a collective product held as private property. Held that way, and tuned by a private guideline that nobody outside the provider can read, it concentrates economic and epistemic power in the same hands. The remedy is not to stop building models. It is to attach to them the terms that already attach to other things built on public inputs: disclosure, benefit-sharing, audit and appeal.
The Privatization of Collective Intelligence: Corporate AI Governance, Epistemic Control, and the Trajectory of Digital Feudalism. PDF, 7 pages, with the paper's own reference list.
Download PDF