Copyright and AI training data: fair use, data mining and authorship
AI models are trained on copyrighted material. In the US, fair use may apply but is not guaranteed. The EU has a harmonized copyright framework, while the UK is consulting on changes. Authorship of AI-generated works raises questions under copyright law.
Fair use and AI training
In the United States, the fair use doctrine permits the unlicensed use of copyrighted works in certain circumstances. Section 107 of the Copyright Act specifies four factors that courts consider: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use on the potential market for the work, as described in the U.S. Copyright Office's Fair Use Index. The Index is a searchable database of court opinions that helps make the principles and application of fair use more accessible, but it does not include all opinions and is not a substitute for legal advice.
The paper on foundation models and fair use explains that in the US and other jurisdictions, the fair use doctrine can allow the use of copyrighted material to build foundation models, but this protection is not absolute. If a model produces output closely resembling copyrighted data, especially in ways that affect the market for that data, fair use may no longer apply, according to the paper on Foundation Models and Fair Use.
Data mining and exceptions
The European Union has harmonized its copyright law through a framework of 13 directives and 2 regulations, which aim to reduce national discrepancies and guarantee protection that fosters creativity and investment, as stated on the European Commission's copyright legislation page.
In the United Kingdom, the government held a consultation on copyright and artificial intelligence. The consultation sought views on how to ensure the legal framework supports both the creative industries and the AI sector, addressing transparency, control over the use of works, and access to high-quality material, as described on the UK Government's consultation page.
Authorship of AI-generated works
The U.S. Copyright Office launched an AI initiative and published a notice of inquiry that received over 10,000 comments. It is issuing a report in several parts; Part 2 addresses the copyrightability of outputs created using generative AI, as noted on the U.S. Copyright Office's AI page.
The World Intellectual Property Organization (WIPO) observes that AI intersects with intellectual property in many ways, including raising questions about authorship, ownership, and remuneration for training models, according to WIPO's page on AI and intellectual property.
Transparency and documentation
Large webtext corpora used for training AI models are often scraped from the internet with minimal documentation. A case study on the Colossal Clean Crawled Corpus (C4) found that the data contains text from unexpected sources like patents and military websites, and that the filtering process disproportionately removes text about minority individuals, as documented in the arXiv paper on documenting large webtext corpora.
The paper recommends that creators of web-scale datasets document where the data came from and how filtering was applied, to improve transparency and understanding of the training data.
| Factor | What courts examine |
|---|---|
| Purpose and character of the use | Whether the use is commercial or nonprofit educational, and whether it is transformative |
| Nature of the copyrighted work | Whether the work is creative or factual, and whether it is published or unpublished |
| Amount and substantiality | The quantity and quality of the portion used in relation to the whole work |
| Effect on the potential market | Whether the use harms the existing or future market for the original work |
Frequently asked questions
What are the four factors of fair use?
The four factors are the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use on the potential market for the work.
Is fair use guaranteed for AI training?
No, fair use is not guaranteed. If a model produces output that is similar to copyrighted material and affects its market, fair use may not apply.
How does the UK consultation address AI training?
The UK consultation seeks views on how to ensure the legal framework supports both creative industries and AI developers, covering transparency, control, and access to material.
Why is documentation important for training data?
Documentation helps understand the composition and filtering of webscale datasets, which can contain unexpected content and have unintended filtering effects.