Ever since the emergence of generative AI, a major concern for all participants has been the extent to which copyright works can and should be used in training AI models.
The application of UK copyright law for this purpose is disputed, leading to inevitable high-profile tension between, on one hand, rights holders keen to control and be paid for use of their work and, on the other, developers who argue that this legal uncertainty is undermining investment in, and the development of, AI in the UK.
Whilst cases are making their way through the courts in the UK and further afield (such as in Germany[1] and the US[2]) on this issue, there have been frequent calls for specific legislation (including by the UK government itself, which has publicly stated that the status quo cannot continue).
As a result, the UK government has launched a consultation[3] open until 25 February 2025 inviting interested parties to submit feedback on potential changes to UK copyright legislation in light of AI. The options set out in the consultation, and on which feedback is sought, range from doing nothing through to the introduction of broad data mining rights which would allow use of copyright works for AI training (including for commercial use), without rights holders’ permission and subject to few or no restrictions.
The Options
The consultation invites feedback on four potential options under consideration:
- Do nothing and leave UK copyright and other related laws as they are – essentially this would defer the matter to the courts to resolve on a piecemeal and ad-hoc basis. Whilst feedback on this option is invited the consultation makes clear this is not an option looked upon favourably by the government, given that it would prolong the current legal uncertainty.
- The opt-in model requiring licensing in all cases – this would strengthen copyright protection for rights holders by providing that AI models could only be trained on copyright works in the UK if an express licence to do so has been granted. This option is likely to be popular with rightsholders, but at odds with the government’s desire to turbocharge the AI economy in the UK.
- A broad data mining exception – this would follow a similar approach already seen in Singapore[4] (and to some extent in the US under its “fair use” standards) and allow data mining on copyrighted works in the UK, including for AI training, without the rights holder’s permission. Under this approach, copyrighted material could be used for commercial purposes, and would be subject to few, if any, restrictions. Needless to say, this option is likely to be very popular with AI developers but is the least favoured by rights holders.
- Allow data mining but copyright holders to reserve their rights along with increased transparency measures – this is the middle ground between options 2 and 3, and would allow AI developers to train AI models using material to which they have lawful access, but only to the extent that right holders have not expressly reserved their rights. Any such use would also be subject to robust transparency measures requiring developers to be transparent about what material their AI models have been trained on. For rights holders, this means an “opt-out” as opposed to “opt-in” model and pro-active monitoring to identify unauthorised use.
The Unanswered Questions
Option 4 broadly follows the approach which has already been seen in the EU under the not uncontroversial text and data mining exception in the EU Directive on the Digital Single Market,[5] which has been further enhanced by the EU AI Act[6] which declared these text and data mining exceptions to be applicable to general-purpose AI models. The government’s view expressed in the consultation is that option 4 is the option which would balance the rights of all participants, although the EU approach was rejected by the previous government as being a threat to rightsholders interests.
However, at this stage it does not represent a “silver bullet” as many issues would still need to be resolved, including those set out below:
- It is unclear how a “rights reserved” model would work in practice and how exactly copyright owners would be able to reserve their rights. The EU equivalent provision requires opt-outs to be machine readable, but query how this works once multiple copies are available. There is also the question of what “machine-readable” means in the context of machines designed to read anything (including handwriting).
- How would such a model apply to works that are already publicly available and how does it address works which have been previously used to train current AI models? It is uncommon for legislation to have retroactive effect. This would then leave it open to debate what will happen with works that have been mined prior to the effective date of the legislation, and would leave early entrants into the AI market with a huge advantage (or disadvantage) depending on what shape any future legislation and court cases take.
- Does this apply to works in non-digital formats? The EU legislation on data mining specifically refers to “automated analytical technique aimed at analysing text and data in digital form.” But how does this apply to books which are scanned in (a process which Google went through many years ago)?
- What happens to AI models if all or a significant volume of rights holders opt-out? An AI opt-out could soon become ubiquitous at which point developers could find themselves wading through a significant volume of claims making the UK an unpopular location for AI development.
- How will rights holders know that their material has been used? The consultation states that robust measures will be put in place to ensure that developers are transparent about the works their models are trained on, but what will be the penalties for failing to be transparent and will there be robust enforcement against non-compliance?
- To what extent can and should any new legislation have extraterritorial application? With many major AI players headquartered outside of the UK, any new legislation which is limited to those based in the UK may have limited impact and an increased legislative burden in the UK could make it a less attractive location for AI businesses.
Ultimately the outcome may be collective licensing deals between rights holders and AI developers as has already happened for a number of news outlets and websites. However, that will be reliant on collective will and action by rights holders and a willingness to embrace AI, which so far has not been forthcoming.
[3] Copyright and Artificial Intelligence – GOV.UK
[5] EU Directive 2019/790.
[6] EU Regulation 24/1689.
Sumaiyah Razzaq contributed to this article