Redistributed

Public data trains private models. The internet's collective output — Wikipedia articles, Reddit comments, news archives, academic papers, books, images, code — becomes the raw material for systems worth hundreds of billions.

The creators of that data see nothing. The companies that train on it capture everything.

This is enclosure: the fencing of commons into private property. It happened with land in 18th-century England. It's happening with data now.

The Extraction

Large language models are trained on datasets scraped from the open web:

Common Crawl — A nonprofit that archives the web. 250 billion pages. OpenAI, Google, Meta, and Anthropic all train on it. The content includes copyrighted news articles, books, personal blogs, forum posts — anything publicly accessible.

The Pile — EleutherAI's 800GB training set (Gao et al., 2020). Includes Books3 (196,640 books, many copyrighted), PubMed Central (academic papers), Stack Exchange (code Q&A), and more. Documented, but not consented.

WebText / OpenWebText — Reddit-sourced datasets. Every upvoted post becomes training data.

LAION-5B — 5.85 billion image-text pairs scraped from the web (Schuhmann et al., 2022). Includes copyrighted photos, art, and personal images — faces, children, medical images.

The scale of extraction:

The New York Times estimates that OpenAI trained on millions of its articles — decades of journalism, created at enormous expense, ingested without payment or permission (Grynbaum & Mac, 2023).

Artists found their work in LAION, used to train image generators that now compete with them. Getty Images found 12 million of its watermarked photos in Stable Diffusion's training data.

Programmers discovered their GitHub code in Copilot's outputs — sometimes with their names still in the comments.

The Legal Battles

NY Times v. OpenAI (2023-): The Times sued OpenAI and Microsoft, alleging that ChatGPT can reproduce Times articles nearly verbatim. The complaint argues this isn't fair use — it's commercial exploitation that undermines the Times' business model.

OpenAI's defense: training is transformative use; the model creates new outputs, not copies. The case will define whether AI training on copyrighted material is legal.

Getty v. Stability AI (2023-): Getty Images sued Stability AI for using 12 million images — including watermarked photos — to train Stable Diffusion. Getty argues this is mass infringement at scale.

Authors Guild v. OpenAI (2023-): Seventeen authors, including John Grisham and Jonathan Franzen, sued over the use of their books in training. They allege OpenAI acquired pirated copies (from shadow libraries like Library Genesis) to train GPT-4.

Sarah Silverman et al. v. Meta (2023-): Authors sued Meta over LLaMA training on Books3, which contained their copyrighted works. The court dismissed some claims but allowed others to proceed.

The pattern: Creators produce value. Tech companies extract it. Legal battles determine whether this continues.

The Fair Use Question

US copyright law allows "fair use" — limited use of copyrighted material without permission for purposes like criticism, education, or transformation.

The AI companies' argument: Training is transformative. The model doesn't copy; it learns patterns and generates new outputs. This is like a student reading books — you don't pay royalties for learning.

Lemley and Casey (2020) articulate this in "Fair Learning": machine learning is a transformative use that creates new expressive outputs, not copies.

The creators' argument: This isn't learning — it's industrial-scale copying for commercial purposes. The outputs directly compete with the inputs. Stable Diffusion creates images "in the style of" living artists. ChatGPT writes news summaries that compete with news organizations.

Henderson et al. (2023) argue that current fair use doctrine wasn't designed for this — billions of works, used commercially, by the world's richest companies.

The outcomes so far: Courts haven't issued definitive rulings. Early decisions are mixed — some claims survive, others are dismissed. The question may ultimately reach the Supreme Court.

Commons Frameworks

What if we thought about training data not as property to extract, but as a commons to govern?

Elinor Ostrom won the Nobel Prize for demonstrating that common pool resources — fisheries, forests, irrigation systems — can be managed sustainably without privatization or state control (Ostrom, 1990).

Her principles:

Clear boundaries — who can access the commons?
Rules match local conditions — one-size-fits-all doesn't work
Collective-choice arrangements — affected parties participate in rule-making
Monitoring — users and external parties can observe use
Graduated sanctions — violations have consequences
Conflict resolution — accessible, low-cost dispute mechanisms
Self-governance recognized — higher authorities respect local rules
Nested enterprises — multiple levels of governance for large commons

Applied to training data:

What would an Ostrom-style data commons look like?

Boundaries: Training data as a defined resource with usage rights
Participation: Creators have voice in how their work is used
Monitoring: Provenance tracking — which data trained which models
Sanctions: Compensation for use, exclusion for violations
Governance: Independent bodies, not just company policies

Yochai Benkler (2006) showed how commons-based peer production — Wikipedia, open source software — can produce massive value outside market logic. The question is whether AI can follow this model, or whether it will enclose the commons entirely.

Technical Approaches

Data provenance: Some researchers are building systems to track what data trained which models. This would enable:

Audit trails (did this model train on my work?)
Attribution (who contributed to this output?)
Compensation (pay proportional to contribution)

Opt-out mechanisms: Robots.txt allows websites to exclude scrapers. But AI companies haven't consistently honored it. The Spawning AI project lets artists opt out of specific training datasets — but relies on company compliance.

Licensed datasets: Some companies are shifting to licensed data. Adobe's Firefly trains only on stock images Adobe has rights to. This is a genuine attempt to build AI ethically — compensating creators and respecting rights. It avoids legal risk but concentrates training data among those who can afford to license at scale. The tension is real: the most ethical approach to data may also be the most exclusionary, favouring well-resourced companies over open-source projects and researchers.

Consent infrastructure: The W3C is exploring standardized machine-readable consent for AI training. The Do Not Train specification would let creators mark their work as off-limits.

What Redistribution Would Look Like

The minimalist approach: Opt-out that actually works. If creators can exclude their work, the balance shifts.

The licensing approach: Collective licensing (like ASCAP for music). Creators join a pool; AI companies pay into the pool; revenue is distributed by usage. This already exists for some uses — extending it to AI training is technically feasible.

The compensation approach: Require payment for training data at scale. This could be:

Per-work: $X per document/image used
Revenue share: percentage of AI company profits
Compute-weighted: proportional to how much a work influenced the model

The commons approach: Public training data governed by public institutions. Publicly funded datasets, trained models, and inference — returning public value from public data.

The transparency approach: Mandatory disclosure of training data composition. If companies had to reveal what they trained on, market and legal pressure would follow.

The Stakes

Training data is the oil of AI — the raw material that makes everything else possible. Who owns it, who controls it, and who profits from it will shape the industry for decades.

Right now, the answer is: tech companies extract it, creators produce it, and the public bears the externalities.

That's not inevitable. It's a policy choice — one being contested in courts, legislatures, and standards bodies right now. And there is precedent for getting these transitions right. The music industry's battles over digital distribution in the 2000s eventually produced workable licensing frameworks (ASCAP, BMI, SoundExchange) that, while imperfect, route billions in royalties to creators annually. A similar transition for AI training data is possible — if the political will exists to build it.

Sources

Benkler, Yochai. The Wealth of Networks. Yale University Press, 2006. Publisher
Gao, Leo et al. "The Pile: An 800GB Dataset of Diverse Text for Language Modeling." arXiv, 2020. arXiv
Grynbaum, Michael M. and Ryan Mac. "The Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted Work." New York Times, 2023. Link
Henderson, Peter et al. "Foundation Models and Fair Use." arXiv, 2023. arXiv
Lemley, Mark A. and Bryan Casey. "Fair Learning." Texas Law Review, 2020. SSRN
Levendowski, Amanda. "How Copyright Law Can Fix Artificial Intelligence's Implicit Bias Problem." Washington Law Review, 2018. Link
Ostrom, Elinor. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press, 1990. Publisher
Samuelson, Pamela. "Legally Speaking: The Dead Souls of the Google Book Search Settlement." Communications of the ACM, 2011. ACM
Schuhmann, Christoph et al. "LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models." NeurIPS, 2022. arXiv

General

Who Owns the Training Data?