Quick take: The legal status of training AI on copyrighted content is actively contested in courts globally. AI companies argue training is transformative fair use; creators argue it constitutes infringement at massive scale without compensation. The outcomes will determine what data future AI can train on, who gets paid for what AI learned, and whether the most valuable content datasets remain accessible for AI development.
Every large language model, image generator, and code completion AI was trained on data created by people who didn’t consent to that use and received no compensation for it. The books, articles, artworks, and code that trained these systems were either scraped from the web or downloaded from digital archives — available online, but created by authors and artists who retain copyright in their work. The legal reckoning for this approach is now underway in courts in multiple countries.
What AI Companies Did and the Legal Theory They Rely On
Training AI models involves ingesting copyrighted content, processing it to extract statistical patterns, and encoding those patterns as numerical weights. The resulting model doesn’t store or reproduce the original content — it encodes patterns learned from it. AI companies argue this is transformative fair use: the training process produces something categorically different from the input, similar to how a student reading thousands of books and developing their own writing style isn’t infringing copyright on the books they read.
The fair use argument has four statutory factors under US law: purpose and character (is the use transformative?), nature of the work (factual vs. creative), amount used, and effect on market. AI companies argue training is highly transformative in purpose, and that statistical pattern extraction doesn’t reproduce the expression of the original works in a way that harms the market for those works. This is a plausible legal argument — fair use has protected some transformative uses of copyrighted content — but it hasn’t been tested in court for AI training at scale.
Major active lawsuits include the New York Times v. OpenAI (the Times claims ChatGPT can reproduce Times articles nearly verbatim and competes directly with the Times’ digital subscription business), Authors Guild lawsuits against major AI companies on behalf of authors whose books trained language models, Getty Images v. Stability AI (claiming the image generator was trained on Getty’s licensed images without authorization), and the Andersen v. Stability AI case on behalf of visual artists.
The Creator Argument
The creator-side argument focuses on two distinct harms. First, the technical harm: AI image generators and language models can produce outputs that directly substitute for the original creators’ work, reducing demand for human-created content. If AI trained on an artist’s style can produce similar work, that substitutes for hiring that artist. If AI trained on journalism can answer questions the newspaper would otherwise answer, that competes with the newspaper’s subscription model.
Second, the consent and compensation argument: creators never agreed to have their work used as training data, received no payment for it, and were not informed. This is distinct from normal fair use scenarios where a limited amount of a work is quoted or sampled — AI training involves ingesting entire works and extracting maximum value from them. The creators’ position is that whatever the legal outcome, the ethical baseline should involve consent and compensation.
The New York Times case is particularly consequential because the Times documented instances of ChatGPT reproducing substantial portions of Times articles nearly verbatim — challenging the “transformative” framing by demonstrating that models can reproduce original expression. If OpenAI cannot explain how the model can reproduce specific articles without memorizing them, the fair use argument weakens substantially. The case may turn on technical questions about memorization vs. generalization in large language models.
How These Cases Will Shape AI Development
The outcomes matter enormously for AI development trajectories. If training on copyrighted content requires licensing, the cost and complexity of training large models increases substantially — advantaging large companies that can negotiate licenses over smaller players, potentially requiring AI companies to maintain records of what they trained on, and creating ongoing royalty obligations. It could also create a valuable licensing market for content creators.
If training is ruled fair use, AI companies can continue scraping internet content freely — but this may accelerate the collapse of incentives to produce high-quality content online if the economic benefit flows to AI companies rather than content creators. The question of who funds high-quality journalism, literature, and art in a world where AI companies extract value from that content without payment is a legitimate economic concern independent of the legal question.
The International Dimension
Copyright law varies significantly across jurisdictions. Japan explicitly permits text and data mining including for commercial AI training — one reason AI companies increasingly use Japanese data. The EU’s AI Act requires disclosure of training data, and the EU Copyright Directive has provisions that could apply to AI training. The UK has proposed a text and data mining exception for AI, then retreated. Different national approaches create regulatory arbitrage opportunities and complicate global AI deployment.
International coordination on AI training data rights doesn’t currently exist and is unlikely in the near term given competing national interests. Countries that want to develop domestic AI industries have incentives to permit broad training data use; countries with strong creator economies have incentives to require licensing. This creates an ongoing tension without a clear resolution path.
Content creators concerned about AI training can take practical steps now: use robots.txt directives to block known AI crawlers (many major AI companies have published crawler identifiers they instruct to respect robots.txt), add explicit “no AI training” notices to published content, and participate in opt-out registries that some AI companies have established. None of these is fully effective — historical scrapes already occurred — but they address future training data collection.
- AI companies argue training on copyrighted content is fair use because it’s transformative — this is contested in multiple major lawsuits.
- Creators argue two harms: market substitution (AI competes with their work) and lack of consent or compensation.
- The New York Times case is pivotal because it documented verbatim reproduction — challenging the “transformative” argument directly.
- If licensing required: training costs rise, advantages large companies, creates creator revenue but complicates AI development.
- If fair use: free training continues but may erode incentives to create the high-quality content that trained the best models.
- Japan permits AI training; EU requires disclosure; US unresolved — jurisdictional variation creates regulatory arbitrage.
Frequently Asked Questions
Did AI companies have the right to train on my work?
Legally unresolved, currently contested in court. The existing copyright framework didn’t anticipate AI training and courts are now determining how it applies. As a moral matter, many creators believe they had a right to consent and compensation that was not sought. As a legal matter, both sides have arguments and the outcome depends on how courts interpret fair use doctrine for this specific use case.
What is the New York Times vs. OpenAI lawsuit about?
The Times claims OpenAI and Microsoft trained ChatGPT on Times articles without authorization, that ChatGPT can reproduce those articles in ways that substitute for reading them on nytimes.com, and that this causes direct harm to the Times’ subscription and licensing business. The Times documented specific cases of near-verbatim reproduction as evidence that training isn’t purely transformative. The case is proceeding in US federal court.
Will creators be compensated for past AI training?
Uncertain. If lawsuits succeed, damages could include past-use compensation. If licensing frameworks develop, they typically don’t include retroactive compensation for historical training. The more likely outcome, if creators prevail, is ongoing licensing requirements for future training rather than retrospective payment for past training — though class action suits seek both damages and injunctive relief.
AI training data copyright, New York Times OpenAI lawsuit, AI fair use doctrine, AI training data rights, artists AI copyright, training data licensing, AI copyright law, creator rights AI