Owning Your AI: The State of the Law
Since 2023, there has been a rapid and accelerating adoption of artificial intelligence (AI) systems by companies worldwide, with a particularly dramatic increase in the adoption of generative AI systems.1 As companies consider further investments in AI systems,2 key business stakeholders need to understand the opportunities and risks surrounding the ownership of AI systems. Of particular concern is ownership rights if a generative AI system acts as an inventor or author.
This chapter presents the primary options for protection and enforcement of a company’s ownership rights in an AI system, including the intellectual property in software, algorithms, configurations, models, training data, prompts, and output of such systems.
Patents
A patent offers the means to protect certain inventive concepts, and can be used as part of a company’s wider strategy to protect the value of its AI-related IP-assets. Companies seek patent protection for their inventions because patents offer a right to exclude others from practicing those inventions.
However, there are challenges to obtaining patent protection. Three stand out as significant hurdles with respect to AI-related inventions.
1) The invention must be disclosed to the public in detail
As an initial matter, companies seeking patent protection for their AI-related inventions must disclose their inventions to the public in exchange for the right to exclude others from practicing those inventions. This is the “patent bargain”, and is fundamentally incompatible with trade secret protection for those inventions. For this reason, given the rapid evolution of AI, many companies forgo patent protection and instead seek to protect their technology by keeping it secret.
2) The invention itself must fall within the permitted subject matter for a patent in the applicable jurisdiction
While patent laws differ by jurisdiction, a general principle is that the subject matter of an invention for which a patent is sought must be eligible for patent protection in the applicable jurisdiction.
In the United Kingdom and Europe, legislation excludes inventions which are implemented in computer programs as being eligible subject matter for patent protection.3 However, the development of case law in each jurisdiction has introduced criteria that provide applicants ways to avoid these statutory exclusions for AI-related inventions, or “computer-implemented inventions” (CIIs). In Europe, an applicant may avoid the exclusion of CIIs from patent eligibility by identifying the technical features of the claim that are new and that contribute to solving a technical problem (thereby satisfying the inventiveness test laid out in Comvik).4 In its recent statement on the patentability of inventions involving AI, the European Patent Office has affirmed that “patents may be granted when AI leaves the abstract realm of mathematical algorithms and computational models and is applied to solve a technical problem in a field of technology.”5
Similarly, in the United Kingdom, an applicant for patent protection may avoid the CII subject matter exclusions by demonstrating that the invention makes a technical contribution to the state of the art.6
In the United States, an application for patent protection may avoid the exclusion of CII subject matter as patent-eligible subject matter by satisfying the two-step Alice/Mayo test. First, the analytical framework of Alice/Mayo only applies when the invention is directed to patent-ineligible concepts such as an abstract idea or mathematical concept. If it is , then the patent application must identify additional elements that integrate a practical application in order for the subject matter to be patent eligible (for example, the improvement of a computer system or technical field).7 The US Patent and Trademark Office (USPTO) has issued guidance on the application of the Alice/Mayo test to determine the subject matter eligibility of CIIs, including worked hypothetical examples applying the subject matter eligibility test to claims relating to CIIs (for example, a method of using an artificial neural network (ANN) to detect malicious network packets).89 As reflected in US data from the USPTO (which classifies some inventions as AI-related), companies have been very successful in navigating these subject matter exclusions.10
Although there are similarities between the approaches adopted in each jurisdiction, demonstrating the “practical application” of a CII in the United States may not always satisfy the “technical contribution” tests in the United Kingdom and Europe (and vice versa). This area is continually evolving, and there have been recent changes in both the guidelines issued by the relevant patent offices and the case law in each jurisdiction that may affect whether patent protection of a CII may be available. By way of example, in November 2023 the High Court in England and Wales ruled that the subject matter of a patent application for an invention implementing an ANN11 did not fall within the statutory exclusion for computer programs, and that the test for technical effect was satisfied by the output of “a file that would not otherwise be selected… and when coupled with the purpose and method of selection it fulfils the requirement of technical effect in order to escape the exclusion.”12 However, this ruling was overturned by the Court of Appeal in July 2024,13 and it would appear that future patent applications for inventions involving ANNs will continue to be refused from being eligible for patent protection in the United Kingdom.14
3) A human inventor must be named on the patent application
At present, the laws of the United Kingdom, the United States and Europe do not permit an AI system to be listed as the “inventor” in a patent application.15
However, inventions created with AI may be patentable.16 As demonstrated by recent rulings made by the national courts and patent offices of the United Kingdom, the United States, Europe and Germany with respect to the patent applications made by Dr. Stephen Thaler for inventions made by his AI powered machine (which he calls DABUS), the clear guidance is that the decisions made in each of these jurisdictions are ““without prejudice” to the possibility of patents being granted to human inventors who use AI as a “sophisticated tool”… [at present, however] attribution of human inventorship remains the filing requirement across these major patent filing jurisdictions.”
Copyright
Where patent protection of a company’s AI-related assets is not available; for example, where a company is not an innovator in its AI foundation model or does not have any patentable inventions related to AI, a company may still wish to exclude others from developing similar AI-related assets. In such scenarios, the underlying elements of a company’s AI assets may be eligible for copyright protection as an original work, including, for example, the software code and output data of such AI systems. Other aspects of the system may also be subject to copyright protection such as the user interface. Unlike with patents, copyright subsists in the protectible elements of an AI asset automatically on an international basis by virtue of a variety of international treaties, without the requirement for application and examination at a territorial level. Such protection will enable the owner of the copyright to enforce its rights against third parties and to prevent copyright infringement.17 Given that the default owner of any copyright is the author of the original work in question (subject to work-for-hire exceptions), companies should ensure that the contractual terms governing the development of their AI-related assets (including agreements with any third-party contractors involved in the development of such assets) include appropriate terms vesting in the company ownership of any IP rights created by such contractors.18 Companies should also have appropriate policies with respect to the use of open source code in the development of their AI-related assets in order to protect the commercial value of a company’s proprietary source code; for example, by requiring that their AI software developers not use open source code that is subject to strong copyleft licenses.19
Companies will face similar challenges issues in seeking to obtain copyright protection of the output of a generative AI tool as they do in seeking patent protection for AI outputs. In both the United States and in Europe, the courts have held that an AI tool alone cannot be considered an “author” for the purposes of copyright.20 In its published guidance on copyright registration for works containing material generated by AI, the US Copyright Office has confirmed that a work containing output of a generative AI tool may also contain sufficient human authorship to support a copyright claim – for example, in the selection or arrangement by a human author of the output of a generative AI tool – however, “copyright will only protect the human-authored aspects of the work, which are ‘independent of’ and do ‘not affect’ the copyright status of the AI-generated material itself.”21 Following the decision of the Court of Justice of the European Union in Infopaq,22 where the court held that for a work to be original, and for copyright to subsist in such work, it must be “the author’s own intellectual creation,” to date, EU Member States have determined that human input is required for works to be copyrightable. In the United Kingdom, the courts have followed the Infopaq test with respect to the requirement for originality in copyrightable works. However, UK copyright legislation does explicitly consider the possibility of “computer-generated works,”23 and if copyright protection is available for such work, authorship of a computer-generated work is deemed to vest in the person “by whom the arrangements necessary for the creation of the work are undertaken.”24 In the case of a generative AI tool, this raises question as to whether authorship in the output of a generative AI tool vests in the user of the generative AI tool who provides the prompts, or the owner of the generative AI tool itself. To avoid ambiguity where works may be generated by computers, companies should seek to clarify where ownership lies in contractual terms; for example, in the terms and conditions of use of its generative AI tools.
The data used by a company to train its AI model may include copyrighted material of third parties. If a company does not have a license to use such copyright material, those third parties may have a claim that the company infringed those copyright rights and that the third parties have ownership rights in the resulting AI model. Although there are certain legislative exemptions that may allow a company to avoid liability for copyright infringement (for example, the “fair use” doctrine in the United States, and text and data mining exceptions in the United Kingdom and Europe), it may take “several years for the courts across various jurisdictions to resolve the arguments related to copyright infringement and the potential fair use of copyrighted material to train AI models”.25 In order to mitigate this risk, companies consider the provenance of the input data used for training their AI assets, and whether it is possible to use proprietary data or licensed data for such training. In addition to mitigating the risk of any third-party copyright infringement claims, companies that use proprietary data or exclusively licensed data for the training (or refinement) of their AI assets may also see benefits in the protection of AI assets and outputs. In particular, it may be more difficult for a third party to reverse-engineer or replicate its AI asset without access to the input data used by the company to train (or refine) its AI asset.
Trade secrets
A company may seek to maintain the value of its AI-related assets as trade secrets (for example where the company does not wish to disclose its invention to the public as part of any application for patent protection). Trade secrets are intellectual property rights on information that is secret, has commercial value because it is secret, and is subject to reasonable steps taken by the rights holder to keep it secret.26 Companies should ensure they have appropriate procedures in place to maintain the status of its AI-related assets as trade secrets, in particular its procedures to maintain the secrecy of such information and to demonstrate that it derives economic value due to its confidentiality (for example, by ensuring its employees and any third parties requiring access to its trade secrets execute appropriate confidentiality agreements containing obligations to maintain the secrecy of such information). AI components which may be well-suited for protection as trade secrets include, for example, the guardrails used by a company to moderate the behavior of, and maintain control over the output of, its AI asset.
Conclusion
In conclusion, companies will need to consider different strategies for the protection of its AI-related assets that are appropriate to the AI asset in question and its underlying components. By adopting a comprehensive AI and IP policy, companies will be best placed to prepare for the protection, and enforcement, of their AI assets, output and procedures, and any IP issues that are unique to the industry in which it operates.27
1 In the McKinsey Global Survey on Al conducted between February 22 and March 5, 2024, of 1,363 participant organizations, 72% reported the adoption of AI systems in one or more business functions (compared to 55% of participant organizations surveyed in 2023), with 65% reporting the adoption of generative AI systems (compared to 33% of participant organizations surveyed in in 2023). The state of AI in early 2024: Gen AI adoption spikes and starts to generate value, McKinsey & Company, 30 May 2024.
2 While total global private investment in AI in 2023 (USD 95.99 billion) decreased slightly in comparison to 2022, total global private investment in generative AI increased nearly eight-fold in 2023 (USD 25.23 billion). Nestor Maslej, Loredana Fattorini, Raymond Perrault, Vanessa Parli, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, and Jack Clark, The AI Index 2024 Annual Report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April 2024.
3 Section 1(2), UK Patent Act (1977), and Article 52(2), European Patent Convention.
5 Artificial intelligence and patentability: statement from the EPO, European Patent Office. Accessed August 9, 2024.
6 Aerotel v Telco Holdings Ltd [2006] EWCA Civ 1371 (27 October 2006)
7 Alice Corp. v. CLS Bank Int’l, 573 U.S. 208 (2014). USPTO guidance confirms that if the patent application demonstrates that the additional elements integrate the judicial exception into a practical application, the claim will not be directed to the judicial exception and the claim will be subject matter eligible. If not, the claim will be found to be directed to a judicial exception, and the claim will need to show that the additional elements amount to significantly more than the judicial exception itself to be subject matter eligible for patent protection.
8 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence, Federal Register (17 July 2024)
9 July 2024 Subject Matter Eligibility Examples, United States Patent and Trademark Office. Accessed August 9, 2024.
10 Giczy, A.V., Pairolero, N.A. & Toole, A.A. Identifying artificial intelligence (AI) invention: a novel AI patent dataset. J Technol Transf (2022).
11 Namely, the patent application concerned an invention which used an artificial neural network to distinguish the semantic similarity of media files from their physical properties to make recommendations to the end user.
12 Emotional Perception AI v Comptroller-General of Patents [2023] EWHC 2948 (Ch), paragraphs 76-78.
13 Emotional Perception v Comptroller 18 July CA-2024-000036.
14 In response to this ruling, the UK Intellectual Property Office has re-issued its guidance for the examination of patent applications involving ANNs. This guidance confirms that patent examiners “should treat ANN-implemented inventions like any other computer implemented invention for the purposes of section 1(2) [of the UK Patent Act 1977”]: Examining patent applications involving artificial neural networks, UK Intellectual Property Office, 25 July 2024.
15 Richard M. Assmus, Brian W. Nolan, Dr. Ulrich Worm, Benjamin Beck, Mark A. Prinsley, Oliver Yaros, Reece Randall, Ondrej Hajda, Ellen Hepworth, Alasdair Maher, Can AI Be An Inventor? The US, UK, EPO And German Approach, 9 January 2024.
16 AI-assisted inventions may be patentable. Brian W. Nolan,USPTO: AI Use in Invention Process Does Not Foreclose Patentability, February 15, 2024.
17 In the United States, registration is a prerequisite to an enforcement action.
18 In the United Kingdom, there is a (rebuttable) presumption that intellectual property created during the course of employment by an employee is owned by the employer. In the United States, such works are owned by the employer automatically.
19 The use of open source code which is subject to strong copyleft licenses (for example, GNU GPL v3.0) in the development of a company’s proprietary code may require a company to make such proprietary code freely available under the same licensing terms as the open source code itself.
20 Thaler v. Perlmutter, No. 22-CV-01564-BAH (D.D.C. Aug. 18, 2023), and No. 10 C 13/2023-1 of 11 October 2023.
21 Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, Federal Register, 12 March 2023.
22 Infopaq International A/S v Danske Dagblades Forening (Case C-5/08).
23 s178, Copyright, Designs and Patents Act 1988.
24 s9(3), Copyright, Designs and Patents Act 1988.
25 Corporate Governance Spotlight: Considerations For Protecting Valuable AI-Related Assets, 6 February 2024. By way of example, Getty Images has filed and is pursuing parallel proceedings against Stability AI in the United Kingdom and the United States for breach of copyright (namely, in Stability AI’s usage of over 12 million copyrighted images from Getty’s website for the purpose of training its Stable Diffusion AI image-generation system).
26 The legal definition of “trade secrets” is not uniform across jurisdictions. In Europe and the United Kingdom, the definition of “trade secrets” is derived from the definition of “undisclosed information” in international law under Article 39 of the Agreement on Trade-Related Aspects Of Intellectual Property Rights. This is contained in the EU Trade Secrets Directive (2016/943) in Europe and the implementing Trade Secrets (Enforcement etc) Regulations 2018 in the United Kingdom respectively, which define “trade secrets” as information which: (i) is secret in the sense that it is not, as a body or in the precise configuration and assembly of its components, generally known among or readily accessible to persons within the circles that normally deal with the kind of information in question; (ii) it has commercial value because it is secret; and (iii) it has been subject to reasonable steps under the circumstances, by the person lawfully in control of the information, to keep it secret. In the United States, the federal statutory definition of trade secrets is contained in the Uniform Trade Secrets Act (UTSA), which defines “trade secrets” as information that derives independent economic value because it is not generally known or readily ascertainable, and it is the subject of efforts to maintain secrecy (UTSA § 1.4).
27Corporate Governance Spotlight: Considerations For Protecting Valuable AI-Related Assets, 6 February 2024.
