Want more content like this? Sign up for our monthly newsletter.
It takes time for regulators to catch up to new technology.
With AI making it possible to reproduce any creative work—code, novels, videos, law school applications—we’re seeing copyright laws stress-tested in a new way. After all, what does “fair use” even look like when all it takes is a few seconds for AI to copy someone else’s work?
Take code, for example. When Anthropic’s latest Claude code leaked online, developers scrambled to post it on GitHub, a software community and coding library. But what about a quick rewrite of that code into another programming language? While Anthropic cited copyright violations for most of these posts, requesting GitHub to remove them, one post remains: a slice of code rewritten by AI, posted by 25-year-old coder Sigrid Jin.
Existing copyright rules were written in an era where copying takes time and assumes that there’s a measurable window where someone can take action to protect their work via trademark or patent. But that window has vanished in the era of artificial intelligence. In this post, we’ll talk through several landmark court rulings on the subject and what they mean for your legal team—and every company that uses or creates content with AI assistance.
Fair use and why it still matters
The U.S. The Copyright Office determined in 2022 that works created by AI aren’t eligible to be copyrighted, even if it starts with a human-led prompt. They doubled down on this in 2025 with a 100+ page report on the subject. But as AI becomes enmeshed in our workflows, where do you draw the line?
The existing framework for copyright relies on the term “fair use.” There are four factors judges use to determine if a work falls under this framework or if it’s a violation of copyright law:
- The purpose and character of the use
- The nature of the copyrighted work
- The amount and substance of the work infringed upon
- The effect and potential use on the market
It’s #4 on this list that is most relevant to corporate use with AI. But the real question that determines a copyright violation or not is how much (or how little) the work has been transformed through that use, and whether or not the user has added value in some way.
Let’s return to the coding example above. Anthropic might not like that their code is still available on GitHub under Sigrid Jin’s post, but because he technically “transformed it” into another coding language, it doesn’t violate their copyright on it, even though you can effectively use that code to create an exact copy of Claude with it.
This is one example of how AI training breaks this fair use model. According to the U.S. Copyright Office, whether or not a generative AI model violates fair use is “a matter of degree” and may not be transformative in all cases, because AI allows for “the creation of perfect copies with the ability to analyze works nearly instantaneously.” Their stance does not create a line in the sand, preferring to defer government intervention at this time by claiming that fair use still holds true.
3 court rulings, 3 very different outcomes: What fair use looks like in the real world
This matter has been intensely debated in recent months, including three landmark court cases. In the span of only a few weeks, federal judges have delivered similar rulings, but with disagreements on the reasoning behind them:
Thomson Reuters v. ROSS
Thomson Reuters, the media company, sued AI research firm ROSS in 2020, alleging that they copied information for AI training purposes from their legal research platform Westlaw. Reuters won the case.
Because ROSS’s objective was to build a competitor product to Westlaw using AI, it violated fair use protocol. They failed to adequately “transform” the information, and the objective—a research platform for legal teams—meant it fell under copyright infringement.
Kadrey v. Meta
A group of fiction and nonfiction authors filed a lawsuit against Meta, alleging copyright infringement and DMCA related to Meta AI’s training protocol. They accused Meta of unauthorized copying of their works to train artificial intelligence, under fair use.
However, Judge Chhambria disagreed with the plaintiffs, denying their motion because they “fail[ed] to present meaningful evidence on the effect of training LLMs…with their books on the market for [AI-generated] books.” He also denied the DMCA claim, stating that “Meta’s copying was not an infringement.”
This is a huge win for AI companies, not least because it lets them continue to train their models on stolen creative work, but also because it focuses so much on the fourth area of fair use—the marketplace—that it ignores the very real implications for authors to have their creative work plagiarized or easily replicated by anyone with access to AI.
Bartz v. Anthropic
Similar to the Meta case, a group of authors filed a class-action lawsuit against Anthropic, claiming the company’s policy of training Claude’s LLM models on their work constituted copyright infringement.
The judge technically ruled in favor of Anthropic, saying that training LLMs is fair use because it is inherently transformative and similar to a human reading or writing to learn something new. The parties settled out of court for $1.5 billion.
However, what makes this case more impactful for AI companies, compared to the Meta case, is that it gives AI companies an easy legal precedent to train their LLMs on whatever they want. However, the court did state that “Anthropic had no entitlement to use pirated copies for its central library. Creating a permanent, general-purpose library was not itself a fair use excusing Anthropic’s piracy.”
This, at least, makes it more clear that how you obtain the data you use to train AI models matters just as much as what you do with it. AI companies can’t use pirated training data. The ruling implies that AI companies must use a licensed, permission-based model for their LLM training moving forward, which is a win for authors and creatives everywhere.
Why outputs matter as much as the inputs
These are only a few of the class-action lawsuits pending against almost every major AI company on the market. However, these results give us a preview of what we can expect moving forward in terms of legal precedent.
The plaintiffs in all three of these cases generate creative output—novels, essays, or nonfiction resources for education. Because there are so few guardrails in place for this new technology, the plaintiffs are looking for a clear line in the sand from the government or the courts that states what is okay and not okay for AI to copy. Without it, AI companies will continue to steal the work of established authors and media companies for their own profit.
The primary defense used for all three of these cases was under “fair use.” What we’ve learned from these rulings is that any AI company can use any material to train their LLMs; Bartz v. Anthropic sets the precedent that this material must be obtained in a legal, permission-based way.
However, there is one legal argument in the Meta case that may challenge this assumption. Rather than interpret the fourth fair use factor as purely a “direct substitution,” where an author’s book is replaced by an almost exact replica of the book, Judge Chhabria discussed an “indirect substitution,” where “AI-generated books could successfully crowd out lesser-known works or works by up-and-coming authors.” This market dilution wasn’t enough to change the outcome of the ruling; but it does raise significant ethical concerns about the harm that AI-generated content can do on a market level, which in the future may be enough to determine a copyright infringement in the future as fair use continues to be stress-tested in real-time by AI.
In-house legal teams are exposed from both sides
But what do these legal precedents mean for your in-house legal team? There’s a lot of “it depends” happening right now in the courts, and it’s not clear how everything will shake out.
For your own IP, do you have a claim or leverage for a licensing deal, if someone uses your materials to train an LLM? Based on these rulings, that depends on the outcome of their LLM training. If they’re building a competitive product, like in Thompson v. ROSS, then you likely do have a claim of copyright infringement. Only major brands like Disney and Warner Music have attempted licensing deals, and even then, Disney’s has already fallen apart with OpenAI.
As AI deployers, what kind of responsibility do you have against legal action? This is where legal and procurement really have to work together—vendor due diligence is now a legal necessity. When you’re choosing an AI company to work with, either to create a custom GPT, whitelabel for your own technology, or for your employees to use in their workflows, then you need to be asking these questions about how they train their models and what data they’re using.
What to watch for in summer 2026 when the next rulings drop
The open question no court has answered yet: How much market dilution is too much? We’ll be keeping an eye on these rulings as they continue to move through the courts this summer. The Supreme Court denied hearing a case on copyright protection for AI-generated art, but this certainly won’t be the last time it comes up. Until then, we’ll continue to see how regulations flex and adapt to changing technology moving forward.
Learn more about how legal teams use AI by registering for the 2026 State of AI in Legal webinar with legal leaders from Harvey, Stanford Law School’s AI Initiative, and LECG.
Ironclad is not a law firm, and this post does not constitute or contain legal advice. To evaluate the accuracy, sufficiency, or reliability of the ideas and guidance reflected here, or the applicability of these materials to your business, you should consult with a licensed attorney.
