Are major changes to copyright law coming to boost UK development of AI models?

On 17 December 2024, the UK government published a consultation paper on changes and clarifications to the law, aimed at making the UK more attractive to the tech companies developing AI models, but  balancing this with supporting the creative industries upon whose data many AI systems are trained.

One of the key proposals is an "opt-out" text and data mining (TDM) copyright exception, allowing use of content for training AI systems where IP holders have not reserved their rights. Another is the removal of copyright protection for computer-generated works. As well as these eye-catching proposals, several other related changes are floated, such as obligations to label so-called deepfakes (or "digital replicas") and other AI-generated or AI-manipulated content. What are the implications?

Opt-out TDM exception

The government's rationale for changes to copyright law is that AI developers feel the UK legal position is unclear, that they do not know how to legally access UK data and so are shunning the UK in favour of training their models in countries whose rules are "clearer or more permissive".

The government considers that the current copyright position is unsatisfactory, leaving developers and rightsholders uncertain as to the extent of their legal positions.

The main way the government wants to address this is by bringing in a new TDM exception to copyright law, which would allow the use of content for training AI systems for any purpose, unless the owner has expressly opted-out of allowing that use.

This new exception would apply only where the user already had lawful access to the content such as via a paid for subscription service, or it had been made freely available online. It would not, for example, allow unauthorised access to paywall-protected content.

The opt-out rights reservation would need to be notified to potential users in a machine-readable format. The government acknowledges the shortcomings of current technical ways of notifying, notably a lack of mechanisms with the granularity to offer differential consents for different situations and purposes, and for different works on a given website, as well as the need for standardisation.

The consultation references similar or overlapping provisions under EU copyright law (in particular Article 4 of the Digital Single Market Copyright Directive) and indicates consciousness of the benefits of alignment with relevant laws in the EU and other jurisdictions.

Disclosure of data sources

An important quid pro quo of the new exception is the introduction of transparency measures for AI developers. These would likely take the form of obligations on them (and possibly others involved) to disclose details of:

• the copyright content and datasets used in training the AI systems;
• web crawlers used to obtain the training data;
• measures taken to ensure compliance with opt-out expressions;

together with supporting measures such as record-keeping requirements.

The consultation acknowledges that compliance with transparency requirements could cause practical difficulties given the huge amounts of data involved, as well as issues with commercial confidentiality and trade secrets, noting that the intention is for measures to be 'proportionate and justified'. Again, the consultation refers to relevant disclosure obligations in copyright and AI laws in other countries and the desire for international engagement and possible interoperability.

The expectation is that the combination of the above measures will together facilitate the negotiation of licences for data access, with collective licensing schemes and data aggregation services playing a major role alongside private contractual licensing. The government seeks views on how licensing may work in practice, and whether the anything is needed from the government in order to facilitate the development of a robust licensing market.

Training AI models abroad

The consultation raises the question of how the UK can "encourage" developers whose AI models are trained in other jurisdictions to nonetheless comply with UK law on model training. This reflects a key issue in the high-profile Getty Images v Stability AI UK litigation, namely whether the distribution and use of AI models trained in more permissive jurisdictions (which would constitute copyright infringement if the models had been trained in the UK) amounts to infringement under UK copyright law.

An extension of the existing law to achieve this (whether via amendment to copyright law or as part of separate AI-specific legislation) would be quite a radical move. Teasingly, the consultation does not include any suggestions on how this might be achieved, merely seeking views.

Copyright for computer-generated works

Existing UK copyright law (unlike US and the EU copyright law) expressly allows copyright protection for works which have been generated solely by computers (that is, where the computer program generates the work with a sufficient degree of autonomy, rather than just being used as a tool to assist the human who is the real creator of the work). In those cases, the copyright is owned by the human who made the arrangements for the creation of the work (the consultation gives the example of a person who inputs a prompt to a general purpose AI system which generates output in response).

The consultation questions the need for this separate protection, noting that its absence does not appear to have inhibited AI innovation in other countries that do not have this protection. The government's preference is to remove the protection unless sufficient evidence of its benefits is received. If it is decided to retain this type of copyright protection, the law would be clarified (in particular, to define what degree of originality was required, or to abolish the requirement for originality).

Other existing exceptions

The consultation asks for views on whether some existing copyright exceptions need amending in the light of developments in AI capability, in particular:

• The temporary copies exception, which allows the making of temporary copies of a copyright work where required for technological purposes, such as internet browser caching or for display on a device screen, but it has been argued may also permit the making of copies for the purposes of AI model training.
• The non-commercial research exception, which permits copying in the course of text and data analysis as part of "research for a non-commercial purpose", but is narrower than the equivalent EU exception (which permits commercial research, and covers database rights as well as copyright).

It is sensible to clarify the scope of these exceptions to ensure they are interpreted in alignment with the government's overall objectives of permitting use of data for AI training (subject to rights reservation). Many AI companies would also welcome the broadening of the non-commercial research exception to cover commercial research.

As well as the above more striking proposals, the consultation addresses and seeks input on other topical AI-related copyright issues, including:

• Copyright infringement by generative AI outputs – the government feels that this is dealt with adequately by copyright law as it stands but wants views on any barriers to enforcement, especially in light of its TDM proposal.
• Labelling AI-generated outputs – the government is broadly supportive of introducing laws to make it mandatory to label certain types of AI-generated or AI-enhanced content.
• Digital replicas/deepfakes – the consultation acknowledges that many artists feel that existing intellectual property law does not allow them to effectively prevent the unauthorised use of realistic digital replicas of their images or voices. The consultation stops short of explicitly consulting on the creation of new type of right, but does "welcome" views on whether the existing law is adequate.
• Infringement during inference/synthetic data – the consultation asks for views on any other areas where copyright law may need to be developed or clarified in order to be applicable to developments in AI capabilities. One example is use of third parties' copyright works during "inference" and other phases of AI systems' operations which lie outside of the pure training and output aspects dealt with elsewhere in the consultation. Another example cited is the increased use of synthetic data (that is, artificial data created in order to train AI models).

Osborne Clarke comment

This is an important and unexpectedly wide-ranging proposal and consultation paper. It will be of particular interest to:

• the creative industries;
• publishers and other content owners (particularly owners of content made available online);
• developers of AI systems which utilise third party content for model training; and
• deployers/users of generative AI systems, especially those who deploy AI-generated content externally.

They will wish to consider their organisation's copyrights risks in relation to AI, whether from the point of view of protecting their own content, avoiding infringement of third party content, or (as is increasingly the case) both.

The government has been unusually forceful in expressly stating that it is minded to abolish protection for computer-generated works unless it receives compelling evidence that the current provisions are positive, for example in terms of contributing to the use or development of AI systems. This would bring the UK in line with other jurisdictions such as the USA and many in Europe. 

As expected, the consultation aims to strike a balance between rightsholders and AI developers, but industry reactions are so far mixed on whether it achieves this. Previous governments have unsuccessfully sought to achieve a balanced approach: for example, following the previous government's proposals on some of the same issues in 2022/23, the two sides spent over a year failing to reach agreement on a way forward for licensing of content for AI training. While it remains to be seen to what extent this government's proposals will survive contact with what will be very strongly expressed views on both sides, it is likely that at least the main proposals will bear fruit.

The new proposals are at an initial stage, and are relatively light on detail, with the government apparently not yet having a settled preferred approach to many of the possible changes. Themes of standardisation, international alignment and proportionality echo throughout the proposals, and it will be interesting to see to what extent these are emphasised and developed by those responding to the consultation.

The consultation is open until 25 February 2025, with the government keen to receive input from interested parties.