Large language models learn by processing enormous amounts of text. A lot of that text was written by people who never agreed to be part of a training set, and some of it came from sources that were not lawfully obtained. That tension is now playing out in courtrooms, negotiations and legislatures. Whatever happens in any single case, the underlying questions will shape how AI is built for years. Let’s untangle them.
The core question
Does using copyrighted works to train an AI model count as infringement, or is it a permitted use? Copyright law gives creators exclusive rights to copy and distribute their work, with exceptions. In the United States, the main exception relevant here is fair use, a flexible doctrine that courts weigh using several factors: the purpose of the use, the nature of the work, how much was used and the effect on the market for the original.
AI developers often argue that training is transformative: the model doesn’t store books as books, it learns statistical patterns, and the output is something new. Creators respond that their work has real market value, that models can compete with them and that they should have a say, and a share of the profits.
Two different issues often get mixed up
- How the data was acquired. Was it purchased, licensed, scraped from the open web, or taken from pirated collections? Even people who accept that training can be fair use may object if the source material was obtained illegally.
- What the model does with it. Does it memorize and reproduce passages? Can it generate content that substitutes for the original?
Courts can treat these separately. A judge might see training as acceptable while still finding fault with how a dataset was assembled. This split helps explain why legal outcomes seem mixed and why settlements can happen even when the broader legal question remains open.
Why settlements matter
A settlement is an agreement to resolve a dispute without a full ruling. It does not establish binding precedent on the underlying question. But it can still move the industry in practical ways:
- It signals that data sourcing is a financial risk companies need to manage.
- It encourages licensing deals as a cheaper, safer path.
- It sets informal benchmarks for what creators might receive.
- It pushes developers to document and clean their training data.
For creators, the practical impact shows up in claims processes, payments and new contract terms.
What creators can do
- Register your works where registration is available, since it strengthens your legal position.
- Keep records of publication dates and ownership.
- Read platform terms. Sites where you post may grant themselves broad licenses to your content.
- Use available opt-out tools, such as technical flags that ask crawlers not to collect your pages, understanding that compliance is voluntary.
- Join collective bodies like author associations or collecting societies that negotiate for members.
- Watch for class action notices that might apply to you.
What AI developers and businesses can do
- Know your data. Keep documentation on where training sets came from.
- Prefer licensed or clearly permitted sources.
- Read vendor terms. If you use a third-party model, check whether the provider offers indemnity for copyright claims.
- Filter outputs. Reduce the chance of generating near-copies of protected works.
- Plan for variation by country. Rules in the EU, UK, Japan and elsewhere differ from those in the US, including provisions on text and data mining.
The licensing path
Many observers expect much of this to be resolved through licensing. Publishers, news organizations, stock image libraries and music companies have been striking deals with AI developers. This creates a market for high-quality data and compensates rights holders, but it raises its own concerns: licenses may favor large publishers, leaving independent writers and artists out. It also risks entrenching big AI companies that can afford licenses while smaller competitors cannot.
Open questions
- Should creators be able to opt in rather than opt out?
- How should compensation be divided among millions of contributors?
- What counts as a substitute for a work, and who decides?
- How do we treat models trained abroad but used locally?
Where this leaves us
No single case will settle everything. What is emerging is a more careful industry where data provenance matters, licensing is normal and creators have more leverage than they did a few years ago. That’s not a complete victory for either side. It is a sign that the rules of the road are being written, and that anyone who publishes or builds with AI should pay attention to them. If you create, know your rights. If you build, know your sources.
Featured image: Allegory of war and Law – Prunksaal – Austrian National Library.jpg by Daniel Gran, CC BY 2.5, via Wikimedia Commons.
