Issues and Concerns in the AI Era

Data, Copyright, and Ownership

---

Training data and copyright concerns

Large language models are trained on vast amounts of text scraped from the internet and other sources: books and academic articles, news and blog posts, websites and forums, software repositories, and transcriptions of audio and video.

Some of this material is in the public domain or shared under open licenses that permit reuse. However, much of it is copyrighted.

Studies have shown that LLMs can sometimes reproduce copyrighted passages nearly verbatim — for example, generating content closely resembling popular books or articles when prompted in specific ways. This raises fundamental questions: Is training on such data without permission or compensation legal? Who, if anyone, has rights over what the model produces? Should creators whose work trains these systems receive compensation?

---

Legal disputes and lawsuits

Because these questions remain largely unresolved, numerous lawsuits have emerged testing the boundaries of copyright law in the AI context:

The Authors Guild class action lawsuit against OpenAI, joined by prominent authors, claims that training on copyrighted books without permission constitutes copyright infringement and that the resulting models can reproduce substantial portions of those works.

The GitHub Copilot lawsuit against Microsoft, GitHub, and OpenAI alleges violations of software licenses when training on open-source code, particularly concerning attribution requirements and copyleft provisions in licenses like GPL.

Similar cases have been filed by visual artists against image generation systems, by news organizations against companies training on their articles, and by other content creators across various media.

These disputes involve different claims and uses of copyrighted material. Their outcomes must be read with their jurisdiction, date, and particular facts in mind. A settlement resolves claims between the parties; it does not by itself establish a general rule about whether AI training is lawful.

---

Fair use and global perspectives

In the U.S., the doctrine of fair use allows limited use of copyrighted material without permission for purposes such as criticism, commentary, news reporting, teaching, scholarship, and research. Courts evaluate fair use claims based on four factors:

1. The purpose and character of the use (particularly whether it’s transformative)
2. The nature of the copyrighted work
3. The amount and substantiality of the portion used
4. The effect on the potential market for the original work

AI companies, including OpenAI, Anthropic, Google, and others, generally argue that training on publicly available copyrighted text qualifies as transformative fair use. Their reasoning: models learn statistical patterns from training data rather than storing and retrieving works, they create entirely new outputs rather than reproducing originals (in typical use), and the purpose — enabling AI systems to understand and generate language — is fundamentally different from the purpose of the original works.

However, this position is contested by many creators and legal scholars, and fair use determinations are highly fact-specific and unpredictable. Courts might consider whether models can reproduce substantial portions of training data, whether AI-generated content substitutes for human-created work in ways that harm creators’ markets, and whether the enormous commercial value derived from these systems should require licensing or compensation.

The four-factor framework above describes U.S. law and should not be assumed to apply worldwide. Different jurisdictions have different frameworks:

• The EU Copyright Directive, Articles 3 and 4, contains two different text-and-data-mining exceptions. Article 3 covers scientific research by qualifying research organizations and cultural heritage institutions with lawful access. Article 4 is broader, but applies only where rights holders have not appropriately reserved the relevant rights. That opt-out condition does not apply to Article 3. Other conditions also apply, and national implementation matters.

• The UK’s Copyright, Designs and Patents Act, section 29A, permits copies for computational analysis for non-commercial research where the person has lawful access, subject to its conditions, including sufficient acknowledgement unless impractical. This is a separate provision, not the EU Article 4 opt-out framework.

• Japan’s copyright law includes relatively broad exceptions for machine learning that are often cited as more permissive than U.S. fair use.

• Where no applicable exception or license covers a use, permission from the rights holder may be required. The answer depends on the jurisdiction and the particular use.

The lack of global consensus means companies operating internationally must navigate a complex patchwork of potentially conflicting legal frameworks, creating uncertainty for AI development worldwide.

---

Who owns AI outputs?

Beyond training data, another critical question is the ownership of model outputs — the text, code, images, or other content that AI systems generate.

Three questions need to be kept separate: whether an output qualifies for copyright, how a provider’s contract allocates any rights, and whether using the output affects someone else’s rights.

In its January 2025 report, the U.S. Copyright Office explained that copyright protects sufficient human-authored expression, including qualifying creative selection, arrangement, or modification of generated material. Purely machine-generated expression without human authorship does not qualify under that U.S. framework; merely providing prompts is not sufficient. Using AI as an aid does not automatically disqualify a person’s own creative work. Other jurisdictions may take different approaches.

A provider’s contract can allocate whatever rights the provider has. For example, OpenAI’s terms assign its rights in outputs, if any, to users to the extent permitted by law. Such an assignment cannot create copyright where the legal requirements are not met.

An output may also contain human-authored or third-party material. A lack of copyright in its generated elements does not make the entire output free of all rights or restrictions. Users still need to consider how they use it.

If an LLM generates text substantially similar to copyrighted work, using that output commercially could still constitute infringement regardless of who “owns” it.