Redistributed

General

AI Copyright Rulings Are Shaping the Training Data Economy

Published March 11, 2026

the-wirecopyrightlegaltraining-datavalueredistribution

What happened

A wave of copyright lawsuits is forcing courts to answer a question the AI industry would rather leave open: does training a model on copyrighted work require permission? The answer will shape not just legal doctrine, but the economics of the entire Training Pipeline.

The marquee case is The New York Times v. OpenAI, filed in December 2023 and still grinding through the Southern District of New York. The Times alleges that GPT models can reproduce its articles nearly verbatim — a claim backed by exhibits showing passages regurgitated word for word. OpenAI counters that training is transformative use, that the model produces something fundamentally new, and that the output examples were cherry-picked through adversarial prompting. The case has survived early motions to dismiss and is headed toward discovery that could expose the full scope of what OpenAI ingested. A ruling here will set the template for every large language model built on scraped content.

But the Times case is not isolated. It sits within a litigation landscape that now spans text, images, and music:

  • Thomson Reuters v. Ross Intelligence — One of the earliest AI training copyright cases. Thomson Reuters sued Ross Intelligence for using Westlaw content to train a legal research AI. A Delaware court found in 2023 that Ross's copying was not fair use, offering an early signal that courts would not automatically bless AI training as transformative.
  • Getty Images v. Stability AI — Filed in both the US and UK, Getty alleges that Stable Diffusion was trained on millions of its copyrighted photographs. Some generated images even reproduced Getty's watermark — a detail that undercuts any "transformative" defense.
  • Authors Guild and individual authors v. OpenAI and Meta — Consolidated class actions representing thousands of writers, including Sarah Silverman, Michael Chabon, and Ta-Nehisi Coates, alleging their books were ingested without consent. Meta's case involves its use of LibGen and other pirated book repositories for training LLaMA — a fact the company has struggled to explain away.
  • UMG, Sony, and other labels v. Suno and Udio — The major record labels filed suit in mid-2024, alleging that AI music generators were trained on copyrighted recordings. Warner Music Group broke ranks and settled with Suno in November 2025, but UMG and Sony are pressing ahead, seeking rulings that would establish training on copyrighted music as infringement.

In each case, the core legal question is the same: is feeding copyrighted material into a training pipeline "fair use" under US law, or does it require a license? (For more on how training actually works, see How Language Models Actually Work and Economics of Training a Frontier Model.)

Why it matters

The stakes here are not primarily about any single lawsuit. They are about who captures the economic value of the data that makes AI systems work.

Training data is the raw material of the generative AI industry. Large language models, image generators, and music synthesizers are built on vast corpora of human-created work — journalism, literature, photography, recorded music, academic research. The companies that built these models did not, for the most part, pay for this material. They scraped it, or obtained it through intermediaries, and treated it as freely available input. The resulting models generate billions of dollars in revenue and market capitalization.

This is a redistribution problem with a clear direction of flow. Value moves from creators to model companies. The Times invested in reporting; OpenAI ingested it. Getty photographers made images; Stability AI trained on them. Musicians recorded songs; Suno learned to replicate them. In each case, the original creators bear the cost of production while the AI companies capture the returns.

The fair use question is where the legal and economic arguments collide. US fair use doctrine weighs four factors: the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the market. AI companies argue that training is transformative — the model creates something new, not a copy. Critics counter that the transformation argument has limits. As Pamela Samuelson has argued, fair use was designed to balance access and incentive; if training undermines the market for the original works, the balance tips against the user. Mark Lemley and Bryan Casey's influential 2020 paper "Fair Learning" proposed that machine learning should generally qualify as fair use, but acknowledged this framework strains when models can reproduce or closely imitate their training data — precisely the situation the Times case presents.

Courts are divided, and the division matters. The Thomson Reuters ruling suggests training is not automatically fair use. But no appellate court has weighed in on large-scale generative AI training. Until one does, the legal uncertainty itself shapes the market: AI companies race to train on as much data as possible before the window closes, while creators lack the leverage to negotiate.

Meanwhile, regulatory frameworks are moving faster than courts. The EU AI Act, which entered force in stages through 2025, requires providers of general-purpose AI models to document compliance with EU copyright law — including the text and data mining exceptions under the Digital Single Market Directive. In practice, this means model providers must honor opt-out mechanisms and provide transparency about their training data. The EU approach does not solve the compensation question, but it establishes that training on copyrighted material is not a free pass — creators have the right to say no.

The emerging market response is a patchwork of licensing deals. The Associated Press signed a deal with OpenAI in July 2023. Reddit licensed its data to Google in early 2024. Shutterstock, Axel Springer, and others have followed. These deals establish that training data has commercial value — but they also highlight who gets left out. Individual creators, independent journalists, and small publishers lack the negotiating power to strike bilateral deals. The value of their work is captured collectively but compensated, if at all, individually.

This is where the commons question becomes urgent — and where the question of Who Owns Training Data intersects with centuries of property theory. Yochai Benkler's The Wealth of Networks argued that the internet enables new forms of cooperative production that challenge traditional property-based models. Training data fits awkwardly into this framework: it is produced individually but has value only in aggregate, and it is being enclosed not by traditional publishers but by AI companies that treat it as a commons to extract from rather than contribute to. Elinor Ostrom's research on governing the commons offers a more practical lens — her design principles emphasize that sustainable commons require clear boundaries, collective decision-making, and monitoring. The current AI training data ecosystem has none of these. There are no clear boundaries around what can be taken, no collective governance by the creators whose work is used, and no meaningful monitoring of what models ingest.

Several compensation models are being proposed. Collective licensing, modeled on music performance rights organizations like ASCAP and BMI, would pool rights and distribute royalties based on usage. Data dividends — payments to individuals whose data contributes to model training — have been proposed by researchers and policymakers but remain administratively complex. Some have suggested a training data tax, with revenue distributed to creators through existing institutions. None of these models has been implemented at scale. Each faces the same fundamental challenge: measuring the contribution of individual works to a model trained on billions of data points. But the fact that multiple compensation architectures are being seriously explored — by rights organizations, policymakers, and some AI companies themselves — suggests the window for structural solutions has not closed. The precedent of music licensing through ASCAP and BMI shows that collective frameworks can emerge even from bitter litigation, if the institutional will exists.

What to watch

Three developments will determine how this shakes out:

  • The NYT v. OpenAI ruling. This case will produce the first major federal court decision on whether large-scale AI training constitutes fair use. Whatever the outcome, it will be appealed. But the district court's reasoning will frame every subsequent case and negotiation.
  • Opt-out enforcement. The EU's framework gives creators the right to opt out of training data collection. Whether this right has practical teeth — whether model providers actually comply, and whether enforcement mechanisms work — will test whether regulation can redistribute power in the training data economy.
  • Collective licensing experiments. If organizations like the Authors Guild or collecting societies in Europe can build workable licensing frameworks, they could establish a model where training data has a price and creators get paid. If they cannot, the default will remain extraction: take first, litigate later.

The copyright cases are not just about law. They are the mechanism through which a fundamental economic question gets resolved: when AI systems create value from human-made material, does any of that value flow back? Right now, the answer is mostly no. The courts, the regulators, and the market are all being asked to change that — and some actors are already building alternatives. Collective licensing experiments, AI Cooperatives, and open-data initiatives like Common Crawl represent different bets on how the training data economy might be structured more equitably. Whether any of them can scale before the legal and market defaults harden remains the central question.

Where this fits

The training data copyright fight is one front in a broader struggle over who captures value from AI systems. It connects directly to the copyright status of AI-generated output (if the output is not protectable, and the input was taken without payment, creators lose on both ends), to the market concentration that makes bilateral licensing deals favor incumbents, and to the Venture Capital Trap that pressures AI companies to grow first and negotiate later. The EU is attempting a regulatory answer through the EU AI Act; the music industry is attempting a market answer through campaigns like Say No to Suno. Courts will determine which approach sticks.

Sources

  • Grynbaum, Michael M. and Ryan Mac. "The Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted Work." The New York Times, December 27, 2023. NYT
  • Lemley, Mark A. and Bryan Casey. "Fair Learning." Texas Law Review 99, no. 4 (2021). SSRN
  • Benkler, Yochai. The Wealth of Networks: How Social Production Transforms Markets and Freedom. Yale University Press, 2006. Yale Press
  • "The AI Act." European Commission, 2024. EU AI Act
  • Vincent, James. "Getty Images sues Stability AI for copyright infringement." The Verge, January 17, 2023. The Verge
  • "AP reaches deal with OpenAI." Associated Press, July 13, 2023. AP News
← Back to all articles
On this page