Protecting AI with IP: Comparing approaches taken in the US and UK

The swift evolution in artificial intelligence (“AI”) has prompted discussion as to how IP law can protect AI assets and AI output. The issue has attracted much debate and attention from both stakeholders and policymakers, as exemplified by the protracted back and forth between the two houses of Parliament regarding the Data (Use and Access) Act (the “DUA Act”) in relation to the training of AI tools on copyright works. (Read our Legal Update on the DUA Act, copyright and generative AI.)

This article discusses the current position in the United Kingdom and United States on copyright and AI, as well as the different ways in which IP law can protect AI assets, especially through trade secrets and patents. Companies are increasingly using trade secrets to protect AI assets, as this is an effective way to maintain the secrecy of information related to AI and to avoid certain drawbacks related to patenting AI technology.

1. The text and data mining (“TDM”) Exemption and Copyright

In the United Kingdom

Rights holders are becoming increasingly concerned that AI models are being trained with input that includes works protected by copyright. For example, an AI model that creates images could be trained on images associated with the descriptive words input into the AI model prompt. In the United Kingdom, the Copyright, Designs and Patents Act 1988 allows for text and data mining of copyrighted works for non-commercial purposes on the condition that the user has lawful access to the works, for example, through a subscription or licence. “Data mining” refers to automated techniques that analyse large volumes of information and data mining can involve the extraction of information from copyrighted works.

In many cases, however, rights holders believe that AI models are being trained on copyrighted works without the copyright owner’s permission. In its consultation from December 2024, the UK government signalled that in its view, changes need to be made to copyright law in the context of AI. The government indicated that it may opt to amend the existing TDM exception to allow the use of TDM for commercial purposes, but only in cases where the rights holder has not opted out of their consent from being used in that way. Regardless of the outcome of the consultation, the government is proposing that changes be made through legislation rather than a code of practice, which will bring welcome clarity and more certainty to this area.

The English courts are considering the issue of AI and copyright, where a company brought claims alleging copyright, trademark and database rights infringement against a UK-based company. The High Court is yet to give its judgment. One of the original three claims was that there was a primary copyright infringement by the AI company copying its works during the training and development stage of their AI tool. This part of the claim, as well as its additional claim of primary copyright infringement, has been withdrawn. Therefore, it is unlikely that the judgment will deal with these issues and it remains to be seen how primary copyright infringements will be dealt with by English law. The High Court’s ruling on the secondary copyright infringement, however, will determine whether other rightsholders can successfully claim copyright against AI developers in the United Kingdom. Primary copyright infringement involves the unauthorised copying or communication to the public, while secondary copyright infringement concerns secondary acts such as selling, distributing, importing and marketing.

The claimant’s decision to drop its two main claims, which are both primary copyright infringement claims, highlights the difficulties that claimants face in successfully claiming copyright infringement in AI assets. The claimant dropped its primary infringement claims because the evidence supporting its claims involved acts of infringement outside of the United Kingdom’s jurisdiction. UK copyright law does not have extra-territorial reach, and therefore some infringing acts must take place in the United Kingdom for UK copyright law to bite. This presents a challenge for rights holders looking to enforce their rights under UK copyright law because AI is frequently trained outside of the United Kingdom. Nevertheless, the ruling will provide some clarity for both copyright rights holders and AI developers going forward.

In the United States

US law takes a different approach to dealing with the training of AI models on copyright works. The equivalent to the TDM regime in the US is the fair-use defence. US courts decide on the facts of each case whether the training of the AI tool on copyright works qualifies as “fair use.” The courts have been more active in the United States than the United Kingdom in considering this issue. Three US district courts have issued opinions analysing the fair-use defence in the context of training an AI model. The first decision issued analysed a non-generative AI model finding the defence inapplicable. In limiting the analysis to non-generative AI, the court appeared swayed by the commercial nature of the use and the lack of transformation of the output from the AI model. Two other cases analysing generative-AI models held that the fair-use defence applied to the training of these AI models.

Differing from the earlier case, these cases found the AI-generated output transformative compared to the copyrighted work. The analyses differed on key aspects of the fair-use defence. The same appellate court was set to review both of these decisions. It would have been interesting to see the appellate judges’ views on the district court analysis. However, that will not occur with at least one case, because the parties informed the district court that the parties have settled that case and district court has approved the settlement. For now, it remains to be seen whether other litigants find a business resolution to avoid the possibility that an appellate decision is issued that is problematic to either content owners or model developers.

In the United States, the court has ruled on a whole host of other issues related to copyright and AI. For example, in another case, the court held that a work created solely by an AI programme does not qualify for copyright protection, although the court acknowledged that copyrightability may be possible for aspects of a work that includes human contribution.

Congress has also considered copyright issues. Bills directed at copyright and AI transparency are currently being heard by Congress, such as The AI Accountability and Personal Data Protection Act¹ and the TRAIN Act.² The AI Accountability and Personal Data Protection Act prohibits companies from using personal data or copyrighted works to train AI models without permission and would create a federal cause-of-action to address misuse of personal data or copyrighted materials. The TRAIN Act would establish an administrative subpoena process, enabling copyright owners to request district court clerks to issue a subpoena to AI developers for disclosure of copies or records to identify with certainty the copyrighted works they believe to have been used in AI training. 

Both the United States and United Kingdom are grappling with how best to balance the interests of AI innovation with the rights of copyright holders. Developers and deployers of AI should closely monitor new developments in the copyright space, as we expect many developments to come on this issue.

2. Trade Secrets

In the United Kingdom

Trade secrets protect information that is “secret.” They are a useful tool to protect AI technologies. Trade secrets in the United Kingdom are governed by the Trade Secrets (Enforcement, etc.) Regulations 2018, which reflects the EU Trade Secrets Directive. The law requires the individual or entity controlling the information to take reasonable steps to maintain its secrecy in the relevant circumstances. The definition of a “trade secret” under UK law is broad and can encompass a wide range of information, including source code, training data, model weights, and proprietary algorithms. Trade secret protection is not time-limited and can, in theory, last indefinitely, provided the information remains secret and continues to have commercial value.

In order to protect trade secrets, it is important for parties to use express confidentiality provisions in contracts and non-disclosure agreements (NDAs). Contractual agreements that make clear that the parties agree to the confidentiality and value of algorithms, machine learning and AI data will minimise risk of the information leaking. It is also important for organisations to implement robust internal secrecy measures. Without such secrecy procedures in place, there is a risk that other individuals could gain access to the data.

A key benefit of protecting AI in this way is that trade secrets do not require making the information publicly available, as is the case with a patent registration. Therefore, trade secrets are an effective way to safeguard sensitive AI technology. Moreover, trade secrets are useful in a fast-moving industry, such as the AI industry, when inventions protected by patents may become obsolete as the patented AI becomes outdated. Trade secrets can easily protect know-how, which is difficult to protect with a patent registration, and contracts can quickly be updated to include new technologies. This is significant as know-how is important in the design, data collection, training, algorithms, output of AI systems and source code of AI.

However, organisations seeking to rely on trade secret protection in their AI assets need to be alert to the challenges posed by enforcing those trade secret rights. For instance, businesses need to meet the requirement that reasonable steps have been taken to protect the trade secret. This is an area of uncertainty, as at the time of writing, there is no judicial guidance in the United Kingdom on this requirement. As well as ensuring that NDAs and the appropriate terms in employment contracts are in place, it is important that organisations enforce and police these terms, for example, by giving employee training. Moreover, for a successful trade secret claim, trade secret holders need to prove that the information is not fully or in part in the public domain and that the trade secret holds commercial value because it is a secret. Organisations should consider these requirements before deciding to protect their AI assets as trade secrets.

In the United States

Similarly in the United States, trade secrets offer an easier path to obtaining protection than patents. When information has economic value due to not being generally well known or ascertainable, trade secret protection is available. This is because statutes on both a federal and state level give AI developers and deployers the ability to maintain trade secret protection. Trade secret laws require owners of the information to take reasonable steps to keep the information secret.

One difference in the United States between trade secrets—as compared to copyright or patent protection—is that the ownership of a trade secret is not derived from the way in which the information was created, but rather upon the lawful ownership of the information. Therefore, the use of trade secrets circumvents questions surrounding their ownership, which is a concern that may arise if patent or copyright protection is sought.

Trade secrets accompanied by well-drafted contracts are a powerful tool for protecting AI technologies in both the United Kingdom and United States. There are clear benefits to protecting AI assets by trade secrets, particularly assets for which the ability to obtain other IP protection is in doubt. Entities will do well to keep these trade secrets and contractual provisions in mind when developing their IP strategies for AI assets and advancements created by AI.

3. Patents

In the United Kingdom

Another way to protect AI assets is through patent law. However, patenting AI technology is a lengthy procedure and there is some legal uncertainty in this area. In a previous Supreme Court case, the UK Supreme Court confirmed that an AI system cannot be named as an inventor under UK patent law, as the law presently requires a human inventor. The court did not consider the wider question of whether AI-generated outputs themselves are patentable, leaving this area of law unclear. Companies looking to patent AI technology may struggle to meet the requirement that the inventor must be a person, which presents an issue for those looking to patent technology that is generated by AI. Following the case, there is debate surrounding whether the definition of an “inventor” in UK legislation should be updated to cover a computer-generated invention, as AI-generated output is often generated by a machine, rather than a person. The purpose of patenting technology is to promote innovation; therefore, the government may legislate in this area if the ruling starts to hinder the advancement of AI in the United Kingdom.

In the United States

As in the United Kingdom, US courts have ruled that AI cannot be listed as an inventor. The invention may be patented when there has been significant human input in the creation or use of the invention. In this respect, trade secrets offer a more effective way to protect AI technology that has been generated by an AI inventor.

In the United States, an advantage of protecting AI with patents is that independent development is not a defence to infringement, unlike trade secrets and copyright. Therefore, patent protection remains a tool that companies may consider to protect AI technology in which it can be shown that there was significant contribution to the creation or use of the tool by a human inventor.

Overall, patenting AI technology remains a legal tool to protect AI. However, the current framework in the United Kingdom and United States requires a human element in the inventorship, and those developing AI technologies should be mindful of this when seeking patent protection.

Conclusion

AI tools are advancing rapidly and businesses will need to consider how to best protect the IP that is contained within their AI tools. Content owners will continue to rely on the protections provided by copyright, particularly if they can demonstrate that infringing acts have been committed in the United Kingdom. Trade secrets may be an effective mechanism to protect AI assets and maintain the secrecy of the AI assets in both the United Kingdom and the United States, provided that organisations negotiate confidentiality provisions within their contracts and ensure that internal measures are taken to keep the information secret. Patent protection, on the other hand, requires a human inventor in both the United Kingdom and United States, so patent protection is a less effective tool to protect technology that is generated by an AI inventor. However, this “inventor requirement” may be amended and overridden by legislation. Given the pace of legal change, AI developers and deployers should keep up with the legal developments in both the United Kingdom and United States when considering how to safeguard their IP assets.