Copyright is the bundle of exclusive legal rights that the author of an original work of authorship holds over copying, distributing, publicly communicating and adapting that work for a limited time. In tech regulation, and specifically in the field of copyright and AI, it is the legal framework that decides whether AI models may be trained on protected works and whether AI-generated output can be protected or can infringe existing works.
Copyright arises automatically on creation, without registration, in every country that belongs to the Berne Convention. Its core concepts are old, but AI tests them at unusual scale: a single model may be trained on billions of works, and a single prompt can produce text, images or music in seconds. This description explains the general principles and is not legal advice; many AI questions are still being decided by courts and legislators.
What copyright protects and for how long
Copyright protects expression, not ideas. A novel’s wording, a photograph’s composition or software source code are protected, while facts, methods, plots in the abstract and artistic styles as such are not. The work must be original: in the EU, the author’s own intellectual creation reflecting free and creative choices; in the United States, independent creation with at least a modicum of creativity.
The following table lists the main exclusive rights and how they typically meet AI activities.
| Exclusive right | What it controls | Typical AI touchpoint |
|---|---|---|
| Reproduction | Making copies, including temporary or digital ones | Copying works into training datasets |
| Adaptation | Creating derivative works such as translations | Outputs closely based on a specific work |
| Distribution | Circulating copies to the public | Sharing datasets or model weights containing works |
| Communication to the public | Making works available online | Publishing infringing outputs or memorized content |
| Moral rights | Attribution and integrity of the work | Distortion or misattribution in generated content |
The Berne Convention sets a minimum term of life of the author plus 50 years, while the EU, the United States and many other countries grant life plus 70 years. After expiry, works enter the public domain and can be used freely, including for training.
Copyright and the training of AI models
Training a model usually involves copying works into a dataset, which touches the reproduction right. Whether this is lawful depends on exceptions and limitations that differ markedly between jurisdictions.
| Jurisdiction | Main legal basis | Practical consequence |
|---|---|---|
| European Union | Text and data mining exceptions of the DSM Copyright Directive | Commercial training allowed unless rights holders opt out in a machine-readable way |
| United States | Fair use, a flexible four-factor test | Decided case by case; early rulings differ |
| United Kingdom | Narrow TDM exception for non-commercial research | Commercial training generally needs a license; reform under debate |
| Japan | Broad exception for non-enjoyment uses | Training widely permitted, with limits for unreasonable prejudice |
In the EU, the DSM Copyright Directive of 2019 contains two text and data mining exceptions: one for research organizations and cultural heritage institutions, and one for anyone, which applies only if the rights holder has not reserved the use. The EU AI Act adds that providers of general-purpose AI models must adopt a copyright compliance policy respecting such reservations and publish a sufficiently detailed summary of training content.
In the United States, courts weigh the purpose and character of the use, the nature of the work, the amount used and the effect on the market for the original. Several district courts issued diverging fair use rulings in 2025, and appeals and settlements are ongoing, so the US position remains unsettled.
Copyright in AI-generated output
Two separate questions arise for AI output: whether it can be protected and whether it infringes someone else’s work. They are often confused, but they rest on different tests.
| Type of content | Typical protection status | Reason |
|---|---|---|
| Purely prompt-generated output | Generally not protected in the US and EU | No sufficient human creative control over expression |
| Human selection and arrangement | Arrangement may be protected | Creative choices in compiling outputs |
| Substantial human editing | Human-made parts protected | Original expression added by a person |
| UK computer-generated works | Protected for 50 years under a special rule | Statutory provision whose future is under review |
The human authorship requirement has been confirmed by the US Copyright Office and by US courts, and EU originality doctrine points the same way. For infringement, the usual test is copying plus substantial similarity of protected expression. A model that memorizes and reproduces training works can create liability for its provider or user, while output that only shares a style or genre typically does not.
Practical risk management for AI developers and users
Organizations reduce copyright risk through data governance and contractual controls rather than through a single technical fix.
- Document data sources and honor machine-readable opt-outs such as robots.txt rules and metadata.
- Prefer licensed, public domain or openly licensed data, checking Creative Commons conditions.
- Apply deduplication and memorization testing to limit verbatim regurgitation.
- Use output filters for known protected material such as lyrics and code.
- Record human creative contributions when protection of the result matters.
- Review indemnity clauses offered by AI vendors and their exclusions.
Common misconceptions
- Publicly available does not mean free to use. Online works remain fully protected unless an exception applies.
- A prompt is not authorship of the output. Simple instructions rarely give sufficient creative control.
- Fair use is not a global rule. It exists in US law; the EU relies on closed lists of exceptions.
- Style is not owned. Imitating a style is generally lawful, but copying specific expression is not.
The simple explanation
Copyright gives creators control over copies of their work. A writer, a photographer or a programmer decides who may copy, share or change what they made. This control lasts for many decades.
AI raises two new questions. First, may a company copy millions of works to teach a model? The answer depends on the country and on whether creators said no. Second, does an AI picture belong to anyone? Usually only the human-made parts are protected.
How it is used in practice
A European publisher plans to build an internal assistant that drafts book summaries. Its legal team checks the training plan against copyright rules. Licensed texts from its own catalog are used freely, while web-crawled material is filtered to exclude sites with machine-readable opt-outs, as the EU text and data mining exception requires.
Editors then rewrite and shape every draft summary before publication. Because the final wording reflects their own creative choices, the published summaries carry normal copyright protection. Memorization tests show the model does not reproduce long passages from source books. The publisher gains a productivity tool with documented compliance and protected final content.