Generative AI – the copyright issues

Introduction

Generative AIs such as ChatGPT, DALL-E, Midjourney, Bard and Tongyi Qianwen, are a type of artificial intelligence that can be used to create and generate new content including text, audio, images or even video. At a high level, once a generative AI algorithm has been trained on a dataset, it can produce new outputs based on the data on which it was trained, in response to a user’s prompt or input.

The viral use of new generative AI tools raises tricky copyright questions with untested answers. This article explores some of those issues and the current position.

Ownership of copyright in AI-generated outputs

In the UK, the ownership of copyright is governed by the Copyright Designs and Patents Act 1988 (the “CDPA 1988”). The general rule is that the author (i.e. the person who creates an original copyright work) is the first owner of any copyright. In many cases this is an easily identifiable person, for example the writer, the artist, the photographer, or the composer who created the work and its protectable elements. In other cases, this may be more complicated and while there are exceptions to the general rule (including works made in the course of employment), generally if a person creates a copyright work, they will be the first owner of any copyright subsisting in it.

In some situations, where a human author has used a computer or software to express their creativity and to produce the work, the author is still easily identified as the human. For example, the use of word processing software to write this article does not mean that the developer of the software would be considered a joint author. The use of a generative AI tool like ChatGPT can complicate matters. Authorship becomes more difficult when the software or AI is used as more than a tool to facilitate the creativity of human authors. Generative AI can create works with minimal input from its human users.

The UK is one of few countries that has a specific statutory provision in its copyright legislation that deals with the authorship of computer-generated works (although the term of that protection is 50 years from the end of the year in which the work was made, instead of 70 years from the end of the year in which the author died for other human generated literary or artistic works). These computer-generated works are defined as works that are generated by a computer in circumstances where there is no human author of the work. s.9(3) of the CDPA 1988 provides that the person who made the necessary arrangements for the creation of the computer-generated work to occur is taken to be the author.

The English Court has considered the scope of s.9(3) of the CDPA 1988 in the context of video games in Nova Productions Ltd v Mazooma Games Ltd [2006] EWHC 24, (and affirmed on appeal: Nova Productions Ltd v Mazooma Games Ltd [2007] EWCA Civ 219). In this case, the video game was based around a game of pool and the developer had created and designed the images of the table, balls and cue which were overlaid when the game was played. The player in this case moved the virtual cue to take shots on screen and to play the game.

The High Court held that the developer of the video game (Mr Jones) was the author of the composite images as artistic works under s.9(3) (at para 105):

“…he devised the appearance of the various elements of the game and the rules and logic by which each frame is generated and he wrote the relevant computer program. In these circumstances I am satisfied that Mr Jones is the person by whom the arrangements necessary for the creation of the works were undertaken and therefore is deemed to be the author by virtue of s.9(3).”

The Court also held that the player was not an author of the artistic works (para 106):

“The player is not, however, an author of any of the artistic works created in the successive frame images. His input is not artistic in nature and he has contributed no skill or labour of an artistic kind. Nor has he undertaken any of the arrangements necessary for the creation of the frame images. All he has done is to play the game.”

How these rules apply to works created with generative AI is untested in the UK, so there is some uncertainty. In the case of a work created by a generative AI, under English law, the person who made the necessary arrangements for the creation of the work to occur could be the developers who designed and built the AI (or more likely their employer). The user who inputted the relevant request may also contribute to the necessary arrangements depending on the level and creative nature of the user’s input. It remains to be seen how the English Courts will reconcile the contributions of the AI developer and the user.

Some developers of generative AIs are clarifying their position on copyright ownership in any inputs and outputs vis-à-vis their users in their terms of use. For example, OpenAI’s Terms of Use dated 14 March 2023 provide that, “As between the parties and to the extent permitted by applicable law, you own all Input. Subject to your compliance with these Terms, OpenAI hereby assigns to you all its right, title and interest in and to Output”. To the extent that OpenAI owns any copyright in ChatGPT’s output, this is purported to be transferred to the user.

Duplicate outputs

If the user is the owner of any copyright work produced by ChatGPT, this raises the issue of duplicative outputs. If two users input the same or similar requests into ChatGPT and the AI produces virtually identical outputs, who owns any copyright in the outputs then?

In order for copyright to subsist, the work must be original. Under English Law, this generally means that the work is the author’s own intellectual creation or a substantial amount of independent mental labour, skill, taste, or judgement has been exerted. Therefore, any simple output like the answer “blue” in response to a user asking, “What colour is the sky?” is unlikely to be afforded protection. The question of originality of computer-generated works has also not been tested in the English courts. In Laddie, Prescott, and Vitoria, “The Modern Law of Copyright”, one of the leading textbooks, the authors submit that originality should be judged objectively (i.e. it should not be relevant if the computer can exert in one millisecond what would be a sufficient level of skill and labour for a human authored work to be considered original). If this analysis is correct, then many of the outputs produced by generative AI will likely be considered original and afforded copyright protection under English law. As an example, the dataset for ChatGPT contains over 570GB of data and 175 billion parameters that are used in response to an input.

If copyright does subsist in two virtually identical outputs and the user owns the copyright in the output, both users could in principle own the copyright in their respective outputs. Both of these users would be able enforce their copyright. However, in order for there to be an infringement of a copyright work, it is necessary to show that the alleged infringing work derives from the work protected by copyright (to show there was some sort of “copying”). If the two users could show that the virtually identical works were generated independently there would be no infringement of the other’s work. Therefore, the copyright protection would not be enforceable against the other user (or any of their licensees or assignees), but could still be enforceable against third parties. This becomes particularly problematic if one of the users decides to publish and licence the work to the world for free. It is not clear how the courts will deal with this issue.

In its Terms of Use published on 14 March 2023, OpenAI has sought to address this issue by including a clause that states: “Other users may also ask similar questions and receive the same response. Responses that are requested by and generated for other users are not considered your Content.” This suggests that OpenAI recognises the challenges posed by duplicate outputs and competing claims by multiple users. This raises other difficult questions. For example, users have no way of knowing if another user has received the same response leading to uncertainty as to ownership of these works. In practice, users will interact with generative AIs to refine and iterate the initial output to create modified versions of the output – this will complicate matters further.

Aside from the copyright ownership issues, AI models use the input and data provided by their users and the outputs to train further and to improve the capability and performance of the models. OpenAI allows its users to opt-out their content from being used to improve model performance, but this may not be available for all generative AIs.

Copyright ownership in other jurisdictions

Most jurisdictions do not have an equivalent of s.9(3) of the CDPA (Hong Kong, India, Ireland and New Zealand are the notable exceptions). Therefore, ownership of copyright in works created by or with the assistance of AI are subject to the usual rules of authorship and ownership.

In the US, there is a registration process for US copyright works created by US citizens. In February 2023, the US Copyright Office revoked copyright protection for images in a graphic novel generated by the AI “Midjourney” as it determined that the images “are not the product of human authorship” as the specific output from Midjourney could not be predicted by its users (distinguishing it from other software used by artists). The US Copyright Office held that the human-authored text combined with the images in the graphic novel constituted a copyright work, but the individual AI-generated images themselves could not be protected by copyright. Following this decision, the US Copyright Office published guidance confirming that in order for copyright to subsist in a work in the US, it must be created by a human author. Where AI technology determines the creative and expressive elements of its output, the US Copyright Office will consider that generated material is not authored by a human and it must be disclaimed in a registration application.

Regarding the position in Europe, the EU Commission published a news article on its “IP Helpdesk”. This article advocates the view that for a work created by generative AI to be protected by copyright, it must reflect a human author’s choices and creativity or the author’s own intellectual creation in accordance with CJEU case-law on subsistence of copyright. This will likely depend on the original request posed to the AI and any amendments or changes suggested by the user to the AI’s output.

This approach may result in AI-generated works being protected in those countries that have a provision similar to s.9(3) of the CDPA and not protected in countries that do not have such a provision.

Generative AI and third-party datasets

Generative AI is trained on pre-existing datasets. In many cases, these datasets will contain works protected by third-party copyright. This issue garnered some attention when Getty Images sued Stability AI for copyright infringement in the UK (and later in Delaware). Getty Images alleged that Stability AI copied millions of images owned or represented by Getty Images, in order to train its Stable Diffusion model.

However, one issue for rights holders is proving that their works are being used by the AI and proving that their work was used to produce the particular output (i.e. the proof of copying issue mentioned above). Getty Images has been able to point to reproductions of modified versions of its watermark on the output generated by Stable Diffusion. However, such evidence may not always be available and copyright holders may not know if their works have been used to train generative AI models. In addition to the risk of copyright infringement, AI developers may also infringe database right in the UK and Europe and breach any applicable licence terms.

The UK Government has recently scrapped its plans to introduce a new copyright and database right exception which would have allowed text and data mining for any purpose in the UK. At present, the text and data mining exception in the UK is limited to non-commercial research only (although this may change, for which see below).

If the output from generative AI is identical or includes a substantial part of a third-party pre-existing work then, depending on its use, the user may also be liable for copyright infringement. The ChatGPT Terms of Use are clear that any such liability will fall on the user rather than OpenAI.

AI and patents

While patents are governed by a distinct regime, similar thorny issues have arisen. On 2 March 2023, the Supreme Court heard submissions in Thaler v Comptroller General, a third appeal of the decision of the UK Patent Office. The Court has been asked to consider if an AI can be named as the inventor on a patent, or if there must be a human inventor (the English Court of Appeal unanimously agreed that UK patent legislation requires patent inventors to be humans). The Supreme Court is also considering if the creator or user of the AI would be entitled to the grant of any patent created by the AI. The Supreme Court decision is expected in the coming months.

Conclusion

In 2021, the UK IPO consulted on the effect of AI on copyright (and patents). The outcome of this consultation was published in June 2022. Ultimately, no changes to the law on computer-generated works were proposed because there was “no evidence at present that protection for [computer-generated works] is harmful, and the use of AI is still in its early stages”.

However, since the publication of the consultation outcome this technology has developed quickly (and going forward will continue to move faster and faster). In March 2023, a policy paper entitled “Pro-innovation Regulation of Technologies Review: Digital Technologies” was published. The author, Sir Patrick Vallance, recommends a “a clear policy position on the relationship between intellectual property law and generative AI to provide confidence to innovators and investors.” The paper notes the uncertainty in the relationship between intellectual property law and generative AI. It highlights the risk that AI companies may infringe copyright or database rights when using publicly available works as inputs to train generative AI models.

In the Government’s response, it has stated that the UK IPO will produce “a code of practice by the summer which will provide guidance to support AI firms to access copyrighted work as an input to their models, whilst ensuring there are protections (e.g. labelling) on generated output to support right holders of copyrighted work.” It also states that the UK IPO will: provide guidance to AI firms on enforcement, coordinate intelligence on systematic copyright infringement and encourage the development of AI tools to assist with enforcement.

In the UK (and around the world) there remains uncertainty on how best to address works created by AI. This technology is only going to become more ubiquitous and the challenges described above show that the position is evolving and further clarification is needed. In the UK, both copyright holders and AI firms will have to wait for the UK IPO’s new code of practice and guidance.