The UK government proposed a broad copyright exception in June 2022 that would have allowed text and data mining for any purpose, including commercial artificial intelligence. The plan was intended to reduce friction for researchers and technology companies that train systems on large collections of text, images and other protected works.
Its significance reached far beyond a technical change to copyright law. Training modern machine-learning models requires access to enormous datasets, but many of the works inside those datasets belong to writers, artists, publishers and other rightsholders. The proposal placed Britain at the centre of a debate that would become increasingly important as generative AI expanded.
Text and Data Mining Was Already Permitted
UK law already contained an exception for computational analysis of copyrighted material. It was limited to non-commercial research and required lawful access to the works. A university researcher could therefore analyse a licensed collection under certain conditions, while a company developing a commercial model faced a more complicated licensing question.
The proposed change would remove the non-commercial restriction. Organisations with lawful access could copy works for automated analysis without obtaining an additional licence specifically for mining. Rightsholders would not be able to contract out of the exception, although security measures could still protect systems and databases.
Supporters argued that the existing distinction no longer matched how research reaches the market. University projects often involve commercial partners, and successful experiments may become products. Determining where non-commercial work ended could consume time and legal budgets before technical development had even begun.
Government Wanted a Competitive AI Environment
The policy formed part of a wider attempt to make the UK an attractive place to build AI. Easier access to data could lower entry barriers, speed experiments and give smaller firms a better chance of competing with companies that already controlled vast proprietary datasets.
The government also considered how patent law should treat inventions involving AI. It maintained that an inventor must be a human, while acknowledging that software could assist the inventive process. For computer-generated works, the UK retained a distinctive copyright protection lasting 50 years.
Together, these decisions showed a preference for adapting existing intellectual-property rules rather than inventing a separate legal system for AI. The approach promised continuity, but it also exposed unresolved questions about who should benefit when machines learn from human creativity.
Creators Faced a Loss of Control
For publishers and creative professionals, a mandatory commercial mining exception could eliminate an emerging licensing market. A company might subscribe to a database or obtain access to a website, then use the material to train a valuable model without negotiating compensation for that additional use.
Lawful access was therefore not the same as informed permission. Buying a book allows someone to read it, but it does not intuitively imply permission to reproduce it at scale for product development. Rightsholders worried that a broad exception would transfer value from those producing source material to companies building automated systems.
The problem was especially acute for independent creators, who lacked the technical tools to discover whether their work appeared in a training corpus. Enforcement would be difficult even where access itself was unlawful.
Europe Took a More Limited Route
The European Union’s copyright framework provided mandatory mining rights for scientific research organisations and cultural heritage institutions. It also created a broader exception that rightsholders could reserve against, giving commercial users more access while preserving an opt-out mechanism.
Britain’s proposed model was more permissive because commercial mining could not simply be blocked by contract. That difference could make the UK more attractive to developers, but it risked creating tension with European partners and international content businesses.
Rules also needed to work across borders. A model can be trained in one jurisdiction and offered globally, making national exceptions only one part of a much larger compliance picture.
Access Alone Would Not Guarantee Better AI
More data can improve model coverage, but quantity does not ensure quality. Datasets may reproduce discrimination, contain personal information, omit important communities or include material with unreliable labels. A copyright exception would not remove privacy, data-protection or safety obligations.
Developers still needed documentation, filtering, testing and clear governance. They also needed sustainable relationships with the people and institutions producing high-quality content. An ecosystem that weakens the incentive to create could ultimately reduce the material available for future systems.
The 2022 proposal captured a genuine policy challenge: AI research benefits from the ability to analyse information at scale, while creative economies depend on control and compensation. Durable rules would have to support both. Treating copyright as nothing more than an obstacle to innovation would miss the economic and cultural value that training data represents.
