IP and related contractual considerations for business users of generative AI
20/02/2024
In this article, IP & Technology partner Gina Lodge and Corporate partner Marc Nourse consider the IP and related contractual considerations for business users of generative AI, either using widely available open knowledge generative AI models such as ChatGPT, or other bespoke tech solutions which deploy generative AI.
What is generative AI?
In short, algorithms which generate the most probable answer in response to user input based on training data to which the algorithms have access. The output may be in the form of text, images, charts, code and other content.
Large language models such as ChatGPT are specific applications of generative AI that focus on natural language processing.
Generative AI can be used in a range of ways which may be helpful to businesses, such as performing data analysis, preparing summaries of complex information, and writing reports. Due to their computational power and ability to learn, these models can quickly perform tasks which would otherwise require significant human effort.
Who owns the IP in AI output?
Because generative AI can create original content, the question arises as to who owns the IP rights in that content. The user? The provider of the AI model? The AI itself?
This is a rapidly developing area and the answer is not yet completely clear. The AI itself has, at least for the time being, been ruled out but the law has not yet decided who should own the IP rights in the output as between the AI user and the AI provider. We explore this question in more detail below, together with related contractual considerations in relation to the ownership and use of AI outputs.
The two IP rights most likely to be relevant to generative AI content are patents and copyright.
Patents
A patent may be available where the output from the AI is an invention in its own right, for example a novel product or a novel process. (Note however that the starting principle is that a software product is excluded from patent protection. Specialist advice must be sought for an inventive software product developed using an AI tool.)
Under UK law, a patent can only be granted to the inventor(s) or to someone who is entitled to the invention by rule of law or prior agreement, such as an employment or other contract. In 2023, the UK Supreme Court confirmed that the inventor must be human. In other words, a patent will not be granted for an invention devised by AI on its own (as claimed by Stephen Thaler in respect of his DABUS AI system). A human inventor can use AI as a highly sophisticated tool for generating an invention, but there must be some human involvement. However, the question remains as to which human would be entitled to an AI-generated invention: the user of the AI who is responsible for the inputs to the AI, or the human creator of the AI model, or both as joint inventors. This question is highly likely to form the basis of patent entitlement disputes in the future.
Copyright
Where the output from the AI consists of an original expressive work (such as text, images, charts, code and other output), the relevant IP right is copyright.
Under UK law, for a work to be original it must be the author’s own intellectual creation, which means the author has made free and creative choices in developing the work. In order for AI-generated works to benefit from copyright protection, the work must express “original human creativity”. The more complex and/or iterative the human inputs to the AI are, the more likely the output is to overcome this hurdle. Once the threshold is met, the AI generated work is in principle capable of attracting copyright protection in the same way as a work created using any other tool, for example a photograph created using a camera.
But who owns the copyright in this scenario? The AI user or the AI provider? There is no answer yet as a matter of law. It may well be that the answer turns on the facts in each case: is it the AI user or the AI provider who is deemed to have exercised the necessary effort, skill or judgement in creating the output? Joint ownership is also a possibility.
If the threshold is not met in respect of any AI-generated work, it may be possible to claim copyright protection on an alternative basis, namely that it is a “computer-generated work”. (The UK is one of only a few countries which permits this.)
In this scenario, UK legislation provides that the owner of the computer-generated work is “the person by whom the arrangements necessary for the creation of the work are undertaken”. That person must have expended some skill and labour in making those arrangements but what might be needed to satisfy this requirement in an AI context is not yet known. Furthermore, it is not yet known who that person would be deemed to be. It must be a human, but which human (user or AI provider or both) remains the subject of debate.
Similar provisions exist in UK design law to permit ownership of computer-generated designs (a design is an IP right which protects the way something looks) by “the person by whom the arrangements necessary for the creation of the design are undertaken”. Similar considerations on ownership will therefore also apply to AI-generated designs.
A practical way to improve certainty
In neither of the foregoing cases is it clear that, as a matter of law and in the absence of a contract clearly specifying who owns the IP rights, the user of generative AI providing the input by which the invention, work or design is generated would automatically obtain the applicable IP rights. Nor is it clear that the provider of the generative AI would automatically acquire these rights. This could lead to a perverse outcome whereby the output is owned by no-one; but given that this option is likely to stifle commercial investment in, and use of, AI, this is the least likely outcome in practice.
Whilst uncertainty over IP ownership persists under IP law, one practical way to improve commercial certainty is to ensure that the contractual framework under which the AI is provided to the user clearly sets out who will be the owner of the AI output.
Open AI’s terms of use for ChatGPT is referred to in this article for ease of reference as these terms are publicly available and ChatGPT is widely used. ChatGPT’s non-commercial terms of use (for EEA, Switzerland and the UK) can be found here and business terms here. Users should refer to terms of use for their own specific situation.
While business terms are relevant for businesses using ChatGPT Enterprise, the non-commercial terms may be relevant where staff have signed up an individual account “off the side of the desk”.
Under Open AI’s terms of use, for both non-commercial and business users, the user is the owner of the output, which Open AI assigns to the user. Importantly, this is only between Open AI and the user, and Open AI excludes and disclaims responsibility for any third party services or output included in ChatGPT.
Third party rights and training data
Once ownership of AI-generated output has been established, there arises a second question as to whether that output can be freely used without infringing the IP rights, in particular copyright and database right, of third parties. This is because the AI will have been trained on data inputs, which may include copyright works or database rights owned by third parties.
Copyright owners have already raised objections against AI providers on this basis. For example, Getty Images is currently involved in UK litigation against Stability AI, owner of the Stable Diffusion AI system, for using copyright protected images in training the Stable Diffusion model.
It follows that an AI user who generates an image (or other work) using AI trained on infringing data might itself be liable for infringement.
The UK government initially confirmed its intention to legislate to permit text and data mining for any purpose, including for the purposes of training commercial generative AI (adding to the current exception for non-commercial mining), but has since reversed its position.
Attempts to produce a voluntary code of practice aimed at making licences for data mining for AI purposes more available have also stalled. An update from the government on where it might seek to go next is expected in the coming months.
In the meantime, one practical way for an AI user to deal with this issue would be to obtain contractual protection for third party IP infringement from the AI provider. This is not uncommonly provided in negotiated contracts with vendors of conventional tech solutions, although it remains to be seen how market practice develops with tech solutions with gen AI features.
As noted above, ChatGPT’s terms of access expressly disclaims any responsibility for third party services or output. To minimise risk, this may practically limit the extent to which its services can be used only to internal usage, as external usage increases the likelihood of an infringement claim, which the user is unable to pass back to the AI provider.
In terms of having lawful access to underlying data, users may also note that ChatGPT’s terms of use provide that non-commercial users represent and warrant that users have all rights, licences and permissions to provide all inputs, and provide that all input and, importantly, outputs will be used to improve services (i.e. to train models, with a right to opt out).
Model training is useful for large language models to improve accuracy of responses to inputs. However, individual users with non-commercial accounts concerned about IP rights in their content (both inputs and output) may consider opting out. In contrast, ChatGPT’s business terms provides that no input and output for enterprise users will be used for its development or improvement.
Responsibility and accuracy
A third question arises as who takes responsibility for the accuracy of AI outputs. Machine learning is probabilistic in nature, and large language models are prone to “hallucination” which describes the phenomenon where the model perceives patterns or objects that are non-existent or imperceptible to humans, creating the possibility of nonsensical or inaccurate outputs.
This is particularly the case for open knowledge AI providers that access vast amounts of disparate and widely available data, although AI providers providing a tech solution for bespoke use cases may aim to technically reduce or eliminate this effect by accessing only discrete, internal datasets and restricting access to unrelated data.
Nonetheless, because AI is a developing field, and machine learning is probabilistic in nature, users should be aware that AI providers are likely to disclaim accuracy of outputs in their contracts. For example, ChatGPT’s non-commercial terms of use disclaim the accuracy of outputs, require users to perform their own evaluation, and prohibit its use to make decisions which could have a legal or material impact on a person; and its business terms provide that users are solely responsible for the use of outputs and the evaluation of the outputs for accuracy and appropriateness.
Final comments
As powerful and convenient as generative artificial intelligence promises to be, businesses must also be informed and intelligent in its use. In this article we have discussed in general terms some areas where contractual terms can and should address issues such as IP ownership, rights of use, third party rights, and responsibility for accuracy. We’d be delighted to discuss any of these points in your particular situation.
This article is not legal advice, which it may be sensible to obtain before you take any decisions or actions in the areas covered. Please do contact me if you would like an initial discussion of your situation.
Related articles:
Should I use generative AI in my business? by Temple Bright Employment partner Rosie Evans, discussing the pros and cons and practical considerations for businesses using generative AI, including the development of organisational policies for its use by staff.
A short introduction to intellectual property: what is it, and why does it matter to my business? by Temple Bright IP partner Gina Lodge (co-author of this article), considering the wider IP issues affecting businesses.
