Ownership, Collective Knowledge, and the Raw Material of Intelligence
–An exploration of how generative AI transforms the accumulated knowledge of society into training data, raising urgent questions about ownership, compensation, and whether the raw material of intelligence should remain a commons or become enclosed within private platforms.
Artificial intelligence systems do not emerge from nothing. They are trained on immense archives of human expression: books, scientific papers, legal opinions, photographs, code repositories, news articles, forum discussions, and the endless stream of commentary that flows through the internet each day. The interface that now produces fluent answers is therefore not simply the product of algorithmic design. It is the product of accumulated human knowledge embedded within training data. In earlier essays, this series examined the political economy, governance structures, and institutional adaptations required by artificial intelligence. Beneath those layers lies a deeper question: who owns the raw material from which intelligence is built?
Historically, knowledge production operated within identifiable ownership frameworks. Authors owned manuscripts. Publishers licensed distribution. Scientists received recognition and institutional rewards for discovery. Even when knowledge circulated widely, it typically passed through structures that attempted to connect intellectual labor with attribution and compensation. Generative AI unsettles this architecture. The systems that power the interface ingest vast quantities of publicly accessible data and transform them into statistical models capable of producing new outputs. The boundary between individual contribution and collective resource becomes blurred. The knowledge that trains these systems appears increasingly communal in origin, yet increasingly concentrated in its technological application.
The Training Corpus
Large-scale generative models depend upon enormous training corpora composed of billions of documents and images. These datasets include public-domain texts, licensed materials, and content collected from across the open web. The scale of this information processing is unprecedented. Instead of storing discrete passages that can be retrieved verbatim, models analyze patterns within language and visual structures, learning statistical relationships that allow them to generate plausible responses to new prompts.
From a technical standpoint, the distinction between storing and learning patterns is crucial. The model does not contain a library in the conventional sense. Yet from a social perspective, the difference may appear less significant. The system’s capabilities still derive from the creative and intellectual labor of countless individuals whose work formed the training environment. Writers, artists, programmers, educators, journalists, and ordinary internet users all contributed pieces of the informational ecosystem that models analyze.
The result is an extraordinary compression of cultural production into computational form. The interface appears to speak with a single voice, but that voice reflects the distributed contributions of millions.
Is Data Labor?
Some economists and legal theorists have begun to frame this transformation in the language of labor. The argument is straightforward in outline, if radical in implication: if generative models derive capability from human-created material, then the individuals who produced that material have contributed value to the system. The model’s performance depends not only on computational power but also on the richness of the corpus it learns from.
Jaron Lanier articulated the problem early. In Who Owns the Future? (2013), he argued that digital economies systematically extract value from uncompensated user contributions, creating what he called “siren servers” that profit from an architecture of free exchange. E. Glen Weyl and Eric Posner expanded the analysis in Radical Markets (2018), proposing that data should be understood as a form of labor rather than as capital passively harvested by platforms. In a formal treatment for the American Economic Association, Lanier, Weyl, and collaborators argued that treating data as labor could restore a functioning market for user contributions and begin to address the distributional imbalance that the current system entrenches.
The analogy to earlier industrial transformations is instructive. In the nineteenth century, factories absorbed techniques that had once resided in skilled artisans. Over time those techniques were embedded into machinery and standardized production processes. Artificial intelligence performs a comparable transformation in the intellectual domain. The accumulated work of writers, artists, coders, and scholars becomes embedded in statistical representations rather than remaining attached to individual creators.
Whether this process should be interpreted as legitimate technological transformation or as a form of uncompensated extraction remains contested. The framing of data as labor reflects a growing effort to articulate that tension and to build institutional responses around it.
Copyright at the Frontier
Courts and regulators have begun to confront the legal implications of generative AI. Lawsuits brought by authors, artists, and media organizations argue that training models on copyrighted material without permission constitutes infringement. Plaintiffs claim that generative systems replicate stylistic patterns and informational structures derived from their work while bypassing traditional licensing arrangements. The range of claimants now spans the creative economy. The New York Times has sued OpenAI. The Authors Guild has filed on behalf of prominent novelists. Visual artists challenged Stability AI and Midjourney. Disney and major record labels have brought cases against image and music generators. Over seventy copyright infringement lawsuits are in various stages of litigation, and the number continues to grow.
Technology companies counter that training represents a transformative process rather than a form of copying. In this view, a model learns from texts in much the same way that human readers learn from books. Exposure to existing material informs future creation without reproducing the original works directly.
Early judicial decisions have begun to sketch the boundaries of this debate. In June 2025, Judge William Alsup ruled in Bartz v. Anthropic that using copyrighted books to train large language models constitutes fair use, describing generative AI as “quintessentially transformative.” Two days later, Judge Chhabria reached a similar conclusion in Kadrey v. Meta, while cautioning that plaintiffs with stronger evidence of market harm would “often win.” These rulings are narrow and fact-specific.
The legal system must still determine whether the transformative-use analogy will hold broadly. The outcome will shape the structure of the AI industry. If training requires licensing, development will likely shift toward negotiated agreements with publishers, artists, and data providers. If courts affirm broad fair-use interpretations, generative models may continue to expand using publicly accessible information as raw input. The litigation is young. The precedent is not yet set.
The Emergence of Data Dividends
Between unrestricted access and strict licensing lies a third possibility: data dividends. Under this approach, individuals or communities whose data contributes to training datasets would receive a share of the economic value generated by AI systems.
The idea draws inspiration from natural resource economics. When corporations extract oil, minerals, or timber from land, governments often require royalty payments that return part of the value to the public. In 2019, California Governor Gavin Newsom proposed a “data dividend” in his State of the State address, arguing that because the state’s data-driven economy does not exist without its data-generating public, that public should receive a fair share of the benefits. Security researcher Bruce Schneier and computer scientist Barath Raghavan subsequently proposed an “AI Dividend” modeled on Alaska’s Permanent Fund, in which large technology companies would pay per-output licensing fees into a fund distributed equally to all residents.
The California Data Dividends Working Group, an ad-hoc team of scholars and practitioners, developed a pragmatic implementation plan rooted in a critical insight: the economic value of data comes primarily from aggregation across large groups rather than from any single individual’s contribution. This reframing matters. It shifts the conversation from “my data” to “our data,” and from individual property claims to collective governance.
Implementing such a system presents formidable challenges. Attribution becomes complex when millions of contributions shape a model’s capabilities indirectly. Measuring the relative importance of individual inputs may prove impossible. Yet the proposal reflects a broader concern about distribution. The debate over data dividends is less about precise accounting than about ensuring that the wealth generated by AI does not accrue exclusively to the firms that control its infrastructure.
The Commons Tradition
Questions about shared resources are not new. Long before artificial intelligence, societies struggled to manage forests, fisheries, grazing lands, and irrigation systems that multiple communities depended upon simultaneously. These resources were often governed as commons, managed collectively rather than owned by a single actor.
Economist Elinor Ostrom demonstrated that commons can function sustainably when communities establish clear rules governing access, monitoring, and benefit sharing. Her 1990 book Governing the Commons presented empirical evidence from settings as diverse as Swiss alpine meadows, Japanese forests, Spanish irrigation systems, and Philippine fisheries, identifying eight design principles that characterized successful self-governing institutions. Her research challenged the widely held assumption that shared resources inevitably collapse through overuse, earning her the Nobel Memorial Prize in Economic Sciences in 2009.
The informational environment that feeds AI resembles a digital commons. It is produced collectively through millions of small contributions rather than by a single coordinated authority. Researchers at the Mozilla Foundation and Indiana University’s Ostrom Workshop have begun translating Ostrom’s design principles to the specific challenges of data commons governance, addressing questions of ownership, stewardship, and benefit-sharing in digital contexts. The challenge lies in designing governance structures capable of preserving openness while sustaining incentives for creative participation.
If the commons becomes fully privatized, knowledge production may narrow into proprietary channels controlled by a handful of platforms. If it remains entirely unregulated, creators may lose incentives to produce original work. The balance between openness and recognition will determine the long-term health of the knowledge ecosystem.
Global Inequalities in Data
The composition of training data also reflects existing global inequalities. Much of the content used to train generative systems originates from regions with extensive digital infrastructure, particularly English-speaking countries and technologically advanced economies. Languages with smaller digital footprints appear less frequently in training datasets, and local cultural contexts may be underrepresented.
Researchers have documented these dynamics in detail. Timnit Gebru, Emily Bender, and their collaborators argued in their influential 2021 paper “On the Dangers of Stochastic Parrots” that large training datasets assembled from internet text overrepresent hegemonic viewpoints and encode biases potentially damaging to marginalized populations. Safiya Umoja Noble’s Algorithms of Oppression (2018) demonstrated how search engine algorithms perpetuate harmful racial and gender stereotypes through the data they prioritize. A broader body of scholarship, from Joy Buolamwini’s empirical work on facial recognition bias to Kate Crawford’s mapping of AI’s material infrastructure and Virginia Eubanks’s documentation of algorithmic harm in public welfare systems, reinforces the central concern: the data that trains the machine is not neutral, and its omissions are not accidental.
As generative AI becomes a primary interface for information retrieval, these imbalances risk reproducing global hierarchies of knowledge. Communities whose languages and experiences are sparsely represented may find their perspectives filtered through models that lack contextual nuance.
Efforts to diversify datasets therefore represent more than technical improvements. They are political decisions about whose knowledge becomes part of the world’s cognitive infrastructure.
The Platform Question
Ownership of training data intersects with ownership of platforms. The organizations building frontier AI systems control both the models and the interfaces through which users interact with them. This dual control concentrates influence over how knowledge is produced, distributed, and monetized.
Such concentration resembles earlier infrastructure industries. Railroads, electrical utilities, and telecommunications networks eventually became subject to regulatory oversight because their monopoly characteristics made them essential public services. Lina Khan’s influential 2017 analysis in the Yale Law Journal reframed this problem for the digital age, arguing that platform power should be understood in terms of infrastructure control rather than conventional price-based metrics. Shoshana Zuboff’s The Age of Surveillance Capitalism (2019) extended the critique further, documenting how platforms commodify human behavioral data into prediction products traded in behavioral futures markets. Cognitive infrastructure may follow a similar regulatory trajectory if access to generative systems becomes fundamental to economic and social participation.
The political question is therefore not only who owns the data but also who governs the systems that transform that data into usable intelligence.
The Creative Ecosystem
Artists, writers, and musicians occupy a particularly sensitive position in the data commons debate. Generative models capable of producing images, stories, and compositions rely heavily on existing creative work for training. For some creators, these tools expand artistic experimentation by allowing rapid exploration of stylistic possibilities. For others, they threaten economic livelihoods by producing outputs that mimic established styles.
Early institutional responses have begun to emerge. In late 2025, Universal Music Group and Warner Music Group each settled copyright infringement lawsuits against AI music generator Udio and entered into licensing agreements structured around artist opt-in consent, giving copyright holders control over whether their works are used to train generative models. These arrangements represent a nascent attempt to balance technological capability with creative labor protections.
The tension illustrates a broader paradox. Artificial intelligence may expand the creative landscape while simultaneously disrupting the economic structures that sustain creative professions. Resolving that paradox requires institutional mechanisms capable of recognizing creative labor without freezing technological experimentation.
Glass Half Full
The emergence of a data commons for artificial intelligence forces societies to reconsider longstanding assumptions about ownership, authorship, and collective knowledge. It exposes the extent to which modern intelligence is already distributed across networks of individuals whose contributions accumulate into shared cultural infrastructure.
The interface is built upon that shared archive.
If societies develop governance frameworks that acknowledge both the collective nature of knowledge and the individual labor that produces it, the data commons may become a foundation for equitable technological development. Mechanisms such as licensing agreements, collective bargaining structures, or public data trusts could allow innovation to proceed while preserving incentives for participation. Sylvie Delacroix and Neil D. Lawrence have proposed a formal “bottom-up data trust” framework in which data subjects collectively govern the terms under which their information is used, and organizations such as the Open Data Institute have begun piloting trust structures in practice. These models suggest that governance of the data commons need not choose between centralized regulation and unregulated extraction.
Artificial intelligence has revealed the commons that has always existed beneath human culture. The challenge now is not discovering it, but deciding how it will be governed.
Further Reading
Arrieta-Ibarra, Imanol, Leonard Goff, Diego Jiménez-Hernández, Jaron Lanier, and E. Glen Weyl. “Should We Treat Data as Labor? Moving Beyond ‘Free.’” AEA Papers and Proceedings, Vol. 108 (May 2018).
Buolamwini, Joy, and Timnit Gebru. “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” Proceedings of Machine Learning Research, Vol. 81 (2018).
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.
Delacroix, Sylvie, and Neil D. Lawrence. “Bottom-up Data Trusts: Disturbing the ‘One Size Fits All’ Approach to Data Governance.” International Data Privacy Law, Vol. 9, No. 4 (2019).
Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press, 2018.
Gebru, Timnit, Emily M. Bender, Angelina McMillan-Major, and Margaret Mitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (2021).
Khan, Lina. “Amazon’s Antitrust Paradox.” Yale Law Journal, Vol. 126 (2017).
Lanier, Jaron. Who Owns the Future? Simon & Schuster, 2013.
Noble, Safiya Umoja. Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press, 2018.
Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990.
Posner, Eric A., and E. Glen Weyl. Radical Markets: Uprooting Capitalism and Democracy for a Just Society. Princeton University Press, 2018.
Schneier, Bruce, and Barath Raghavan. “AI Could Pay Dividends to Americans.” Politico, June 29, 2023.