AI & IP consultation - focus on copyright for computer-generated works
In view of the increasing use of artificial intelligence (AI) in creative fields, the UKIPO considered that a consultation focussing on potential changes to copyright, as well as patent, law was needed. In this article we explore the issues considered by the consultation and explain the reasoning behind D Young & Co’s input to the consultation, with a particular focus on copyright of computer-generated works.
In an earlier newsletter we explored the issues surrounding patents in more detail: https://dycip.com/ai-ip-patents.
Section 9(3) of the Copyright, Designs and Patents Act 1988 (CDPA), accords authorship of a copyright work to the person “by whom the arrangements necessary for the creation of the work are undertaken”.
As AIs become increasingly autonomous and complex, the current consultation seeks to establish whether this definition is fit for purpose, or whether changes are required to account for the increasing creativity of AIs.
As part of the consultation the UKIPO has identified three potential ways forward for protection of computer-generated works. These are as follows:
|Computer generated works|
|Option 0||Make no legal action.|
|Option 1||Remove protection for computer-generated works.|
|Option 2||Replace the current protection with a new right of reduced scope/duration.|
Making no legal change is the preferred option
Contrary to the situation for determining inventorship of patents, in our opinion the UKIPO’s current system for determining authorship of copyright works made using an AI is fit for purpose, such that no legal change is required. In our response to the consultation we therefore advocated for option 0.
The definition provided by Section 9(3) of the CDPA appears to offer a reasonable approach and treats the AI as a tool like any other computer based art tool whose programming contributes to the overall aesthetic of a copyright work. An AI may be a more sophisticated (or perhaps just more opaque) tool, but it is a tool nonetheless. Therefore, even once AIs advance to become more autonomous, there would still be “arrangements necessary for the creation of the work”, as recited in Section 9(3). This definition would therefore appear to render the current provisions fit for purpose now, and in the future.
Removing protection for computer-generated works should be discouraged
In our response to the consultation we argued that removing protection for computer-generated works would seem to be a disincentive to the production of such works, and also seemingly an unjustifiable punishment inflicted because of an apparent intractability with the law rather than because of any lack of artistic merit in works developed using an AI.
Removing, or limiting, protection for computer-generated works would also lead to a significant problem in defining which works should have their protection altered. Here it will be understood that many works generated using a computer do not use an AI, and it is unclear whether the present consultation intends all computer-generated works to be considered together, irrespective of whether such works require the use of an AI or not.
It is also unclear at what point a work should become a “computer generated work”, regardless of whether AI is used. For example, it will be appreciated that many artworks, and indeed photographs, are routinely generated using computational methods. On the one hand, computers are utilised to determine how digital brushstrokes blend within an image, or how a model is airbrushed or warped by algorithms, and it is likely that these could be considered to contribute to the work being considered “computer-generated”. However, it is also commonplace to use an in-camera computer for dynamic range computation, aperture selection, exposure compensation, colour saturation, focus selection, and many other elements of composition – including in some cases when to take the photo itself – that would previously have been used in support of according copyright to a photographer. It is significantly less clear which of these should be considered to contribute to a work being considered “computer-generated”, even if the selections are made by something claimed to be an AI.
It does not seem justified to remove protection for a certain category of artistic works merely because new developments in technology make it more difficult to determine the appropriate level of protection.
Replacing the current protection with a new right of reduced scope/protection should also be discouraged.
The final option proposed in the UKIPO consultation is to replace existing copyright protection with a new right for computer-generated works of reduced scope/duration. Whilst we consider that this may represent an acceptable middle ground if it is deemed that amendments to the current legislation are required, the introduction of a new right may lead to additional issues making it an unattractive proposal.
One may assume that an AI has no inherent artistic capability, and/or that it is not entitled to moral rights or rights that reflect a personal investment in the artistic properties of the work, as is the aim of copyright protection. As such, an AI could be treated as a design engine, whose output is commercialised by another party, and so the rights afforded to its output may be more suitably protected with a right which is similar to a registered design.
If we were to equate rights provided for computer-generated works with design right, a duration of about 25 years, similar to design right, would seem appropriate. Even if the commercial exploitation is to sell it as an artwork, it will be appreciated that an AI can produce an endless stream of such works without tiring (or without new inspiration) by randomly exploring the training space its internal weights represent, and a reduced duration of right for such works would seem appropriate.
However, there are many works which an AI would not produce itself if the user had not provided the necessary prompt; for example the snail harps and avocado chairs generated by GPT-3: https://dycip.com/avocado-armchair. There is therefore a question of whether the user is a co-author of the work, or whether they simply provided the system with a design brief. For human artists, the latter may be true, but for an AI where the input directly influences the AI’s search within its state-space, this is less clear-cut, particularly since the phrasing of the same basic requests to GPT-3 would result in different works being produced. As such the user’s input may again be equated with the “arrangements necessary for the creation of the work” under Section 9(3) of the CDPA.
We note that, as acknowledged in the consultation itself, typically the dissemination of a computer generated work by the owner of the computer will itself attract additional copyright protection that greatly exceeds the existing 50 year term available to the person running the computer. As a result, any changes to the duration accorded directly to that person are likely to be ineffectual unless they also percolate downstream to other uses of the work (for example, dissemination by the owner of the computer).
If the duration of protection afforded to a computer generated work was reduced without alteration of the downstream rights, an unnecessarily complex scenario is likely to develop where a piece of computer aided art comprises within itself a mix of rights of different durations (for example where there are computer-generated special effects in a movie). This could lead to significant enforcement issues.
Hence this option appears to suggest either a relatively ineffectual change to the duration of the right, to reflect the diminished artistic stature of the source, or the creation of complex networks of rights within a work, which may have different durations. This would appear to be the case unless corresponding changes to reduce the duration of protection were allowed to pass through other phases of the creative and commercial processes of making the art accessible. This would appear to represent a more fundamental, and far reaching, change to copyright protection.
Further, as discussed above, removing computer-generated works from copyright protection and setting up a new right would require an accurate definition of a “computer-generated work”. Since not all computer-generated works involve the use of an AI, careful consideration would also need to be given to whether the new right would apply to all computer-generated works or only those made using an AI, as well as how an AI should be defined. In a field which is developing as quickly as AI, providing these definitions represents an ongoing and substantial challenge.
Finally, it will also be appreciated that such a new right may not transfer smoothly to other Berne convention countries, making the dissemination of such works even more complex.
Copyright in text and data mining (TDM)
As well as the copyright in computer-generated works discussed above, the current UKIPO consultation also considers issues of copyright in text and data mining. In our response to the consultation we advocated that the UKIPO should extend the existing TDM exception to cover commercial research and databases.
The present framework of TDM allows non-commercial research. However, the findings of such research are often subsequently used for, or inform, commercial purposes.
It will be understood that many “big data” AIs ingest huge quantities of data, and it will be frequently impractical to determine the copyright status of every training item, or to subsequently separate an AI from the source material or training set. For example, where a commercial AI is the result of any non-commercial research phase of developing such an AI, or its training set, it is not then possible to separate the AI from the source material in a sense that safely makes the exploitation of the AI a subsequent and separate commercial act.
In order to allow the commercialisation of AIs that are trained on datasets which include copyright works, we therefore consider that the current TDM exception should be extended to cover commercial research and databases. Notably, however, this should not remove from the persons responsible for the AI any liability for copyright infringement by the results of the AI; an AI should not be used to “wash” the copyright from an earlier work it has been trained on.
It is apparent that AIs will make a comprehensive contribution to computer-derived works over the coming years. There is therefore a need to ensure that such works are appropriately protected. We believe that in most cases where a person makes the arrangements necessary for the creation of the work, the current law is fit for purpose. However, if the production of artistic works by an AI is considered akin to using the AI as a design engine, it could be appropriate for the duration of protection to be similar to design right, that is, 25 years.
If changes in the scope of protection afforded to such works are going to be made, it will be important to distinguish between works derived using an AI, and computer-derived works which do not use an AI. This may also be affected by the definition of an AI, which is likely to change over the coming years as AIs develop and become more autonomous, and as AIs are increasingly used in both the creative and functional aspects of a computer-generated work.
With respect to TDM, copyright provisions with respect to AI will also need to be considered as the use of AI expands. Specifically, it will need to be appreciated that an AI is inextricably linked to its training set, and that removing a training set for commercial exploitation of an AI, which was permitted during non-commercial development, may be particularly difficult.