Copyright and Artificial Intelligence

AI and Copyright Consultation: Ministerial Foreword

Sets out the government's aim to balance growth in creative industries and AI sectors through a copyright framework that lets right holders reserve rights and receive payment, while enabling AI developers to access high-quality training data. The framework should address current legal uncertainty over whether using copyrighted works to train AI models requires a license, and proposes increased transparency from AI developers alongside a copyright exception for text and data mining. Notes that delivering this approach will require practical and technical solutions developed through collaboration between AI companies and creative industries to create systems enabling creators to exercise their intellectual property rights.

A.1 Executive Summary

The government backs growth in both creative industries (worth £124.8bn GVA annually) and AI sectors but identifies current problems: creative organisations cannot control use of their works in AI training or secure remuneration, UK legal ambiguity discourages major AI developers from training models domestically leading them to train elsewhere where copyright rules are clearer or more permissive, and this disadvantages UK SMEs unable to train abroad and leaves UK rights holders unprotected. The government proposes balancing measures including requiring AI developers to disclose training datasets, introducing a text and data mining exception to copyright that permits right holders to prevent their content's use in AI training, and improving transparency. Eight areas for stakeholder consultation are identified: transparency standards for AI training use, technical tools for right holders to reserve rights, licensing including collective and individual creator licensing, labelling of AI outputs, ownership of computer-generated works, digital replicas and deepfakes, emerging issues including inference-time data use and synthetic training data, and next steps for legislation and implementation.

A.2 How to engage with this consultation

Invites stakeholders to submit views on the proposals through designated channels, runs for ten weeks from 17 December 2024 to 25 February 2025, and welcomes both quantitative data and qualitative evidence.

A.3 Data protection and confidentiality

States that consultation responses and respondent names or organisations may be published publicly and made available under freedom of information laws, though respondents may request information be treated as confidential.

B.1 Background

Explains that the UK leads in both AI development and creative industries, notes copyright laws' difficulty in controlling AI training uses and enabling creator remuneration, and emphasises the need for solutions respecting creators' rights while allowing AI developers access to high-quality training data.

B.2 AI training and copyright

Copyright protects creative works and related rights, granting owners the right to authorise or prohibit copying, distribution, and public communication; AI models are trained on large volumes of data from copyright works through automated data mining, which requires permission from copyright owners unless an exception applies, often without express licensing or creator compensation. AI developers and copyright holders dispute whether existing UK copyright exceptions permit such training, with ongoing litigation, and the government intends to clarify copyright rules through potential legislation to enable creators to control and receive remuneration for works used in AI training while ensuring AI developer access to quality content, supported by transparency requirements about model inputs and outputs. Generative AI outputs that include substantial parts of copyright works without a licence may infringe copyright, though co-creations with human input credit the human as author, and computer-generated works without human authorship may be protected with their creator being whoever arranged their creation.

B.3 Our objectives

The government aims to balance innovation in UK AI development with safeguards for creators and the creative sector. Policy proposals will be assessed against three objectives: ensuring rights holders can license and profit from AI model use of their works, enabling developers to obtain and utilise substantial amounts of copyrighted material for model training in a lawful manner, and establishing a transparent copyright system that clarifies which works are used to train models and what they produce. Implementation will respect international copyright treaties and maintain a competitive global position.

B.4 Policy options

The government considers four policy approaches: maintaining status quo despite legal uncertainty, mandating licenses for all AI training on copyrighted works to protect right holders but risking UK competitiveness, adopting a broad mining exception allowing commercial mining with few restrictions, and implementing a data mining exception allowing training on lawfully accessible material unless right holders expressly reserve rights through mandatory transparency measures and effective rights-reservation technologies. The preferred option is a data mining exception with rights reservation supported by developer transparency requirements, designed to provide legal clarity, enable licensing opportunities, maintain creative industry viability, balance AI innovation with creator control, and ensure right holders receive payment where appropriate, though the government acknowledges uncertainty about practical implementation and seeks feedback on this option's suitability and preferred specific implementation details.

C.1 Exception with rights reservation

A data mining exception applies to any commercial or non-commercial use where users have lawful access to works including those on the internet and accessed through subscriptions, but right holders can prevent use by reserving rights through an agreed mechanism requiring developers to obtain a licence, supported by transparency measures on training sources. Rights must be reserved using standardised machine-readable formats that are accessible and effective, and where any right holder applies such a reservation, the exception does not apply and licensing is required. A reservation on one copy of a work may extend to other copies, requiring developers to avoid all expressions of that work.

C.2 Technical standards

Various technical approaches for reserving rights exist, including robots.txt standards, metadata-based approaches, and direct notification registries, but they lack standardisation, are not universally adopted by developers, and are difficult for right holders to use at scale, particularly for individual work-level control; the government believes further standardisation is needed to enable content publishers to easily reserve rights while allowing developers to respect these decisions, and regulation may be required to support adoption and compliance with standardised protocols and metadata. The government seeks views on whether greater standardisation of rights reservation protocols is needed, how compliance can be encouraged, and what role government should play in ensuring standardisation.

C.3 Contracts and licensing

Material licensed through contracts with creators or intermediaries including collective management organisations represents an important AI training source, but creators and performers report insufficient control and broad contractual terms, with questions about whether blanket licensing by collective organisations meets creator needs and whether these organisations can effectively reserve members' rights. The government will support fair licensing between right holders and AI developers, explore text and data mining exceptions with rights reservation and transparency measures, and work with the Department for Education on ensuring high-quality AI tools for teaching while protecting and managing pupils' intellectual property rights without excessive burden on the education sector.

C.4 Transparency

Right holders struggle to determine whether their works train AI models because developers often do not disclose or limit information about training sources; the government believes increased transparency from AI developers is crucial for copyright compliance and enforcement and may regulate to require disclosure of specific works, datasets, web crawler details, and records, though practical challenges exist for small businesses and some information may need protection. California's Assembly Bill 2013 requires producers of generative AI systems to publish information about datasets used for developing, testing, and validating models, including a summary of those datasets. The consultation invites views on whether developers should disclose training material sources, the appropriate level of detail needed, handling of web crawlers, proportionate implementation, and costs to developers.

C.5 Wider clarification of copyright law

The government will clarify copyright law for models trained outside the UK jurisdiction and will engage internationally with the EU, US, G7 and G20 to align approaches and establish level playing field between providers training models in and outside the UK. It will also consider whether the temporary copies exception requires clarification regarding its application to generative AI model training.

C.6 Encouraging research and innovation

Proposes that regulatory burdens on AI developers should be proportionate to the issues addressed, and considers whether copyright rules should differentiate based on AI model purpose or firm size. Examines whether the existing non-commercial research data mining exception remains appropriate, given differences with the EU's approach regarding eligible organisations and coverage.

D.1 Computer-generated works: protection for the outputs of generative AI

AI outputs may receive copyright protection as authorial works if original and reflecting human creative choices, as entrepreneurial works like films or sound recordings regardless of originality and owned by the person arranging production, or as computer-generated works lasting 50 years with no human author and authored by the person making necessary arrangements or, for general-purpose AI, the person providing the user prompt. The government proposes assessing three options against objectives of legal clarity, appropriate incentive structures, and balance between human creativity and technological progress: maintaining current provisions; clarifying the originality requirement by either removing it entirely or redefining it such as protecting works that would be original if made by humans; or removing protection for wholly AI-created works while allowing continued protection for AI-assisted works showing human creativity and for AI-generated sound recordings, films, and broadcasts. The government currently lacks justification for the current provision as drafted, notes that most countries including the US and EU do not provide such specific protection, has found no evidence that the UK protection increases AI outputs or investment compared to jurisdictions without it, and seeks consultation feedback and current evidence on whether protection is being used, its practical effects, and its economic impact on stakeholders.

D.3 Computer-generated works interaction with designs

UK design law, both registered and unregistered, includes provisions for computer-generated designs where the author is identified as the person responsible for making arrangements for design creation, similar to copyright law. No significant concerns about these provisions were raised in the previous 2021 consultation, and no changes are proposed here, though the framework will be reviewed separately in the future.

D.4 Infringement and liability relating to AI-generated content

AI-generated content infringes copyright if it contains a major portion of a copyrighted work without permission. While AI models learn from many protected works, their typical outputs do not replicate original creative expression. However, evidence shows that generative AI occasionally does produce existing copyrighted works or material parts thereof, which constitutes infringement.

D.5 AI output labelling

Industry labelling initiatives that disclose whether outputs are AI-generated inform consumers and support source attribution, benefiting copyright owners and service providers. The government believes labelling should be encouraged and may be necessary to ensure consistency, considering approaches in other jurisdictions like the EU's requirement for machine-readable AI output detection. Technical challenges include defining what degree of AI use requires labelling, ensuring labels resist manipulation, and scalability of content provenance systems, with the government considering support for research and development on labelling tools.

D.6 Digital replicas and other issues

Digital replicas are realistic images, videos and audio made by technology to copy someone's appearance or voice, often created without consent. Existing UK legal protections for personality include intellectual property rights, passing off law, performers' rights in Chapter 2 of the CDPA 1988, and data protection law, though some creative industry respondents propose new personality rights to give individuals greater control over use of their likeness or voice by AI. The government recognises concerns about individual control over creation of content depicting their personality or likeness but does not propose new personality rights in this consultation, welcomes views on whether current legal frameworks adequately protect such control, and notes that the Information Commissioner's Office is reviewing how data protection law under the UK GDPR and Data Protection Act 2018 applies to generative AI.

D.7 Other emerging issues

The government commits to monitor new developments as AI technology advances and will consider further action to clarify legal requirements for AI developers and protect copyright holders' rights. The section seeks public input on whether copyright law's application to emerging AI practices such as inference, retrieval-augmented generation, and training on synthetic data requires clarification or regulatory response.