AI Training on Copyrighted Content: What Authors, Creators, and Businesses Need to Know

The legal fight over AI and copyright is no longer a hypothetical debate among academics. It is playing out in federal courtrooms right now, with authors, news organizations, and visual artists all suing AI companies over how their work was used to train large language models. If you are building with AI, advising clients who do, or simply trying to understand your exposure, you need to know what is actually at stake.

Let's start with the core tension: AI systems learn by ingesting enormous amounts of text. That text was written by people. Most of those people never gave permission for their work to be used this way. Whether that matters legally, and how much, is exactly what courts are being asked to decide.

How AI Models Actually Use Your Content

To understand the copyright question, it helps to understand what "training" means in practice. A content-aware AI model, such as a large language model (LLM), reads through billions of documents. It does not store those documents like a filing cabinet. Instead, it adjusts billions of internal numerical weights based on patterns in the text. When the training is done, the original documents are not sitting inside the model in any readable form.

This distinction matters enormously to the legal analysis. AI companies argue that because the model does not "contain" the training data, using that data to train is transformative and does not infringe copyright. Authors and publishers argue the opposite: the training process creates a commercial product that derives its entire value from their work, and that should require a license.

Neither position is crazy. Both are genuinely contested.

The Four-Factor Fair Use Test

In the United States, the primary legal shield AI companies invoke is fair use. Fair use is not a blanket permission to use anyone's work. It is a balancing test courts apply case by case, weighing four factors:

1. The purpose and character of the use (commercial vs. educational, transformative vs. reproductive)
2. The nature of the copyrighted work (creative fiction gets more protection than factual reporting)
3. The amount and substantiality of the portion used (did you take the whole thing, or just a piece)
4. The effect on the market for the original work (does the new use substitute for the original)

The most important factor in modern case law is the first one, specifically whether the use is "transformative." The Supreme Court's Campbell v. Acuff-Rose decision (1994) established that transformative use weighs heavily in favor of fair use. AI companies lean hard on this: they argue that training a model to generate new text is fundamentally different from reproducing the source text for readers to consume.

Courts have accepted transformative use arguments in analogous cases before. Google's book-scanning project, which digitized millions of books without author permission, was found to be fair use in Authors Guild v. Google (2d Cir. 2015). The court emphasized that Google's search index let users find snippets, not read entire books, and that the project served a different purpose than the originals.

The AI training cases are pushing that logic much further.

Where Verbatim Output Changes the Picture

Here is where the analysis gets more concrete, and more dangerous for AI developers.

If an AI model, during inference (the part where it generates a response), reproduces a substantial portion of a copyrighted work verbatim, that is a very different legal situation from the training question. Reproduction of the actual text is the classic definition of copyright infringement. The training-is-transformative argument does not help you if your chatbot is spitting out three paragraphs of a Stephen King novel word for word.

The New York Times v. OpenAI and Microsoft lawsuit, filed in late 2023, hinges largely on this. The Times demonstrated that certain prompts could cause GPT-4 to reproduce lengthy, near-verbatim excerpts from Times articles. The complaint includes side-by-side comparisons of the original articles and the model outputs. This is not a theoretical harm; it is documented reproduction of specific protected text.

Derived content sits in a murkier space. If a model produces a summary, a paraphrase, or a response clearly inspired by a source but not copying it directly, the copyright claim is much harder to win. Paraphrasing alone does not infringe copyright; only expression is protected, not ideas or facts. A summary of a news article written in different words is generally not infringement. A fictional story with the same plot as a novel, written in different words, is also generally not infringement (though it might raise other claims).

The practical upshot: verbatim or near-verbatim reproduction is where the legal risk is clearest. Derived content, especially when genuinely transformed, is harder to attack under copyright law alone.

What the Ongoing Cases Are Telling Us

Beyond the Times case, the landscape includes:

• Silverman v. OpenAI: Authors Sarah Silverman, Richard Kadrey, and others sued over their books being used in training data. A federal judge dismissed most claims in 2023 but allowed some to proceed, including one focused on whether outputs are "infringing derivative works."
• Getty Images v. Stability AI: Getty sued over image generation, arguing that Stability AI's model was trained on Getty's watermarked images. Courts are still working through this, but it highlights that the copyright issues extend well beyond text.
• Concord Music Group v. Anthropic: Music publishers sued over AI reproducing song lyrics, which are among the most tightly controlled copyrights in existence.

None of these cases have produced final, binding rulings yet. We are in the early innings. But the fact that courts are allowing some claims to survive dismissal signals that AI companies do not have a clean legal safe harbor, at least not yet.

The Opt-Out Problem and What Robots.txt Cannot Fix

Many AI companies now offer "opt-out" mechanisms, where publishers can signal they do not want their content used for training. Some use the robots.txt standard; others have proprietary systems. OpenAI, for example, launched a media manager tool that lets publishers block future crawling.

The problem is that these opt-outs are prospective. They do not address the training data that was already collected, often before any opt-out system existed. An author whose 2018 novel was scraped in 2021 and used to train a model released in 2022 gets nothing from an opt-out tool launched in 2024.

There is also a deeper issue: opting out of training is not the same as licensing your work. Copyright law does not require rights holders to opt out of infringement. The default position in copyright is that you need permission to use someone's work, not that the creator has to affirmatively refuse. Opt-out frameworks implicitly flip this default, and many authors find that unacceptable.

Practical Takeaways for Businesses Using AI

If you are integrating AI into your products or workflows, here is what to take away from all of this:

• Verbatim reproduction is your clearest risk. If you are building a system that retrieves or generates text, set up guardrails to prevent extended reproduction of third-party copyrighted content. Retrieval-augmented generation (RAG) systems that surface source text should be designed carefully.
• Fair use is not a guarantee. It is a defense, not a permission slip. You can invoke it after you get sued; you cannot rely on it to prevent a lawsuit from happening.
• Licensing deals are emerging. The Associated Press, Axel Springer, and others have struck licensing agreements with AI companies. If your AI product is heavily reliant on a specific body of content, a licensing conversation is worth having now rather than after a court order.
• Watch the cases. The New York Times case in particular could produce significant precedent on both the training question and the verbatim output question. A ruling could shift the entire industry's exposure overnight.
• Jurisdiction matters. The EU's AI Act and the EU Copyright Directive have different frameworks than U.S. law. If you operate internationally, you may face stricter rules on training data transparency and opt-out rights.

Where This Is Heading

My honest read: courts are likely to carve out at least some protection for training as a transformative use, especially for models trained on broad internet corpora rather than targeted content from a single publisher. The Google Books precedent is real and relevant.

But the verbatim output cases are harder for AI companies to win. And the broader political and regulatory pressure to compensate creators is not going away. Expect some form of licensing framework or statutory compensation system to emerge over the next several years, similar to how compulsory licenses work in music.

The authors and creators raising these concerns are not anti-technology. They are asking a reasonable question: if your product is worth billions of dollars and it was built on our work, why do we get nothing? That question does not have a clean legal answer yet, but it has a moral weight that courts, regulators, and the public are all starting to feel.

If you are navigating AI adoption and want to understand your legal and strategic exposure, Thought Spark AI works with businesses to build AI systems that are both effective and defensible. Reach out to start a practical conversation about what AI integration looks like for your specific situation.