Trump Administration and Fair Use: How Did We Get Here, and What Is Fair Use?

Fair use is a term we are hearing increasingly in the debate around AI and copyright. In the US, AI companies are relying heavily on the doctrine to argue that they do not need licences to use copyrighted works for AI training.

But what is fair use, and why does it exist?

The fair use doctrine, codified under §107 of the US Copyright Act, exists in part to ensure that copyright protection does not become so absolute that it prevents freedom of expression, research, criticism, education and further creativity.

Think about writing an article like this one. What if copyright law meant that you could never quote from a book when criticising it, reproduce part of a work for academic research, or reference an existing creative work when producing something new?

Copyright is intended to protect and incentivise creativity, but taken to its absolute extreme, it could also inhibit the very creative and developmental cycles it is intended to encourage. Fair use provides a mechanism for balancing those interests.

And that brings us to AI. Was a doctrine developed to enable criticism, scholarship, research and further creativity also intended to allow commercial AI companies to ingest potentially millions of songs, lyrics, compositions, books, images and other copyrighted works to build technology capable of generating content that may ultimately compete with the works on which those systems were trained?

That is one of the fundamental copyright questions now facing the US courts.

Before we get into Donald Trump, AI and how we arrived at this point, it is worth understanding how fair use actually works.

 There is no rule that says “AI training is fair use.” Nor does describing something as “transformative” automatically make it fair use.

 Under §107, courts consider four factors:

• Purpose and character of the use

• Nature of the copyrighted work

• Amount and substantiality of the portion used

• Effect of the use on the potential market for, or value of, the copyrighted work

No single factor automatically determines the outcome. They are weighed together, based on the circumstances of the particular use. And when those four factors are applied to generative AI, things become considerably more complicated.

So how does the courts look at this, if we look at cases such as Authors Guild v. Google (Google Books) / Bartz v. Anthropic / Kadrey v. Meta, the courts ruled that fair use was an acceptable defence. If we look at Kadrey v. Meta, Meta prevailed on fair use for Llama training, but the judge expressly warned that the ruling did not mean AI training is generally fair use; the authors' market-harm evidence was inadequate.

So lets look at the opposite scenario, where fair use was dismissed Dr. Seuss Enterprises v. ComicMix.  In this case the court look specifically at the harm done to Dr. Seuss Enterprises and the book industry. A Star Trek/Dr. Seuss mash-up copied Seuss's style, composition and structure. The court found it non-transformative and commercially targeted a market Seuss could exploit/licence. All four factors weighed against fair use.

 How does this translate into music and generative AI?

Put simply, a generative AI company needs data to train its model. In the case of a music-generation model, that may mean ingesting and analysing enormous quantities of existing music so that the system can learn patterns and relationships within melody, harmony, rhythm, lyrics, instrumentation, structure and other musical characteristics.

The larger and more diverse the training dataset, the more material the model potentially has from which to learn.

This is where the licensing debate begins.

 If copyrighted music has already been copied and used to train a model without permission, obtaining a licence afterwards does not undo that original act. And once knowledge has been incorporated into a trained model, the commercial dynamics of licensing become considerably more complicated. Rights holders are therefore increasingly demanding permission should have been obtained before their works became training material in the first place. So let's imagine a model has ingested millions of songs and can now generate 100 new songs.

 How does a US court decide whether the use of those copyrighted works for training was fair use?

We go back to the four factors:

• Purpose and character of the use – Why were the works copied and what was the purpose of using them? Was the use commercial? Was it genuinely different from the purpose of the original works?

• Nature of the copyrighted work – What kind of works were used? Songs, lyrics and compositions sit at the highly creative end of the copyright spectrum, which generally affords them stronger protection than predominantly factual works.

• Amount and substantiality of the portion used – How much of each copyrighted work was copied? Was the entire recording or composition used, and was taking that amount reasonably necessary for the claimed purpose?

• Effect on the potential market for, or value of, the copyrighted work – Does the use substitute for the original works, interfere with existing or reasonably likely licensing markets, or enable products that compete economically with the works that were used?

This is also where we frequently hear another word: “transformative.”

But transformative does not simply mean: Does the AI-generated song sound different from the original?

 In fair-use law, transformation is principally concerned with the purpose and character of the use. AI companies therefore argue that when a model ingests a song, it isn't using that song for its original purpose. The machine isn't “listening” to music for entertainment. It is computationally analysing the work, learning statistical relationships and using that information to develop a generative model. That, they argue, is a fundamentally different and therefore transformative use.

 Rights holders see another side to that argument. If copyrighted songs are copied without permission to build a commercial system whose purpose is ultimately to generate new songs, and those songs then compete for the same listeners, streams, sync opportunities, advertising budgets and licensing markets as the music used to train it, how different is the ultimate commercial purpose really?

And this is where we need to separate training from output. 

If Suno generates a song that reproduces protectable elements of an existing song so closely that it constitutes copyright infringement, that is principally an output infringement question: has the generated work copied protected expression from the original? That is different from asking whether Suno was entitled to copy the original work during training under the doctrine of fair use. The two questions are connected, but they are not the same.

What is the Trump administration saying and how does it affect the music industry?

The Trump administration said, “This Administration will never let our Nation be at a disadvantage relative to our foreign adversaries based on a plainly incorrect understanding of copyright law.” Relating to the ongoing case between New York Times v OpenAI. A case where The NYT, in a copyright infringement claim against OpenAI who trained its models on a large quantity of NYT newspaper articles without prior consent. The NYT is citing, unauthorized training, direct competition and loss of revenue as consumers can stay on OpenAI search engine rather than looking at the articles on NYT website or other forms of media provided by the NYT.

The administration is siding with OpenAI stating; “Rules of law that make it significantly more difficult to develop a robust AI industry in the United States, threaten national security and give a competitive advantage to foreign adversaries who are not so encumbered,” - the Justice Department 

How do political statements like this affect the music industry?

Although every fair use case is determined on its individual facts, favourable rulings in these cases could have significant long-term consequences for the music industry — particularly when the US government itself enters the debate.

A DOJ filing does not change copyright law. Nor does it determine how a court must rule. But government intervention matters. It can influence the wider policy debate, signal how an administration believes existing copyright law should apply to AI, and ultimately contribute to the direction of future legislation and regulation.For the music industry, the stakes are enormous.

If US courts increasingly accept that copyrighted works can be copied at scale without permission for the purpose of training commercial generative AI systems, the question quickly becomes: what happens to the emerging licensing market?

Why would an AI company negotiate a licence for training data if the law ultimately tells it that no licence is required?

And there is a wider cultural issue here too.

If the political and commercial narrative becomes that copying copyrighted works without permission is acceptable provided the ultimate technological purpose is sufficiently transformative, we risk normalising an idea the music industry has spent decades trying to move away from: that access to music somehow removes the obligation to obtain permission from the people who created and own it.

We have seen versions of that mentality before.

Napster. LimeWire. The Pirate Bay.

The technology is completely different, and the legal questions are not the same. But the underlying tension should feel familiar to the music industry: technological innovation moves faster than licensing, enormous businesses and consumer behaviours develop around access to copyrighted works, and rights holders are left trying to establish where permission and payment should sit after the technology has already achieved scale. The danger is therefore not simply one court deciding that one particular training activity constitutes fair use.

 It is the precedent, legally, commercially and culturally, that may follow.

If permission becomes the exception rather than the starting point for building commercial AI systems from copyrighted creative works, the consequences could extend far beyond today's AI litigation. They could fundamentally reshape the value of music rights in the next generation of technology.