AI and copyright
In short
AI and copyright is the legal question of how copyright applies when protected works are copied to develop an AI system, or when a person uses AI to make something new. The answer depends on the country, the work and the facts of the use.
- Also called
- Copyright and AI, Generative AI copyright
- Topic
- Safety, ethics and risk
- Level
- Beginner
The US Copyright Office says several stages of building a generative AI system use copyrighted works in ways that implicate the owner's exclusive rights. Developers make copies by downloading works, moving them, converting formats and creating filtered subsets. Training can add further copies. An output that replicates or closely resembles a protected work is, the Office says, likely to infringe. Uses during development and uses during deployment need separate consideration.1
Part 3, released on 9 May 2025, is still the pre-publication text. The Office's AI page says a final version will be published later, without any substantive change expected in the analysis or conclusions.4
In the United States, fair use is the defence the Office analyses. It says various training uses are likely to be transformative, but whether they are fair depends on the works, the source, the purpose and the controls on outputs, because those facts can affect the market. Commercial use of vast amounts of copyrighted works to produce expressive content that competes in existing markets, especially where the access was illegal, goes beyond established fair-use boundaries.1 The Office's Fair Use Index calls fair use a case-by-case doctrine under section 107, not a fixed amount of copying.2
Fair use is a US doctrine. The Office says the EU, Japan and Singapore have text-and-data-mining exceptions that may bear on AI training. In the EU, the broader exception depends on lawful access and on respecting a copyright owner's opt-out. The Office says how that opt-out applies to generative AI is still disputed.1
An output is not automatically unprotected, and it is not automatically protected. The Office says a generative output can be copyrighted only where a human author has determined sufficient expressive elements. That can include a human work still perceptible in the output, or a human's creative arrangement or modification of it. Merely providing a prompt does not qualify. Using AI as an aid, or placing AI-generated material inside a larger human work, does not by itself bar protection.3
Example
Someone who only types a prompt has not, on the Office's account, determined the expressive elements copyright requires. A creative rearrangement of the output, or their own work still visible inside it, can be protected.3
Why it matters
The Office treats training copies and the copyright in one output as different questions. One decides whether a developer needed permission. The other decides whether a human contribution in the result can be protected.13
Questions people ask
Is training an AI model on copyrighted works illegal?
There is no single worldwide answer. The US Copyright Office says many training uses are likely to be transformative, but fair use still depends on the works, the source, the purpose, output controls and the market. It says some commercial uses that compete with existing works, especially after illegal access, go beyond established fair use.1
Can I copyright an AI-generated image?
In the US, the Copyright Office says an output is protected only where a human determined sufficient expressive elements. A creative arrangement or modification can qualify, as can a human work that remains perceptible in the output. A prompt, on its own, does not.3
Does public online content have no copyright?
No. The Copyright Office says 'publicly available' is not the same as authorised. It may only mean the material is on the internet. In the US, fair use can allow some unlicensed uses, but the Fair Use Index says each case turns on its facts, not on a fixed amount of copying.12