The Artificial Intelligence (“AI”) revolution is upon us. With the introduction of tools such as Dall-E, Midjourney, GPT 4 and Microsoft Copilot into mainstream consciousness, the use of AI has rapidly evolved. Initially promised as a solution to more repetitive and mundane tasks, the scope and reach of AI has resulted in unexpected implications for many industries, including art.
A number of these Generative AI models are able to, within seconds, create original images informed by succinct, text-based descriptions (“prompts”) or reference images supplied by the user. The possibilities for the future of the creative process are endless, with the average layperson, one who has little to no background in art, now being capable of creating professional-standard works through nothing more than a sentence.
How is this possible?
Generative AI art is made possible through machine learning models. Although the images generated by AI may be perceived as original creations, these models are not sentient and as such creativity is not an inherent characteristic. Rather, these models are trained (programmed) on large data sets of existing artworks scraped from across the internet. These models use this data to generate mathematical representations of visual patterns in different objects, themes and artistic styles. This process enables the Generative AI models to make increasingly accurate predictions in response to text prompts input by users when generating images. GPT 4, for example, is trained to match certain words with certain prompts, intending to provide answers which are the most probable correct response to the input.
The implications of this process are especially relevant for the commercialisation of artistic works, commercialisation which is made possible through the exclusive rights derived from the copyright in one’s work – which an artist should be able to assign or license in order to generate income. In the context of Generative AI art and the data collection process that enables it, questions around human artists’ consent, credit and compensation are salient to the use of machine learning models and have already found resonance in other jurisdictions such as the US and the UK. Read about some of those ongoing legal battles here and here.
A brief introduction to Copyright Law in South Africa
In South Africa, there is no registration system for the subsistence of copyright, with copyright automatically coming into existence once the statutory requirements for them are met. These requirements are found in the Copyright Act 98 of 1978 (“The Act”).
In terms of s3 of the Act, the default position is that copyright is conferred upon the creator, or “author”, of the work. Deciding who to credit as the author in the context of Generative AI art has already proven to be a legal conundrum. Legal questions have arisen regarding who the copyright derived from Generative AI art vests in: the user who inputs the request for generation of the art, the AI itself, or the company that controls and owns the AI. Whilst these questions are valid, they are beyond the scope of this article. This article focusses on individual artists’ whose works are currently being used in the process of creating Generative AI art. It will first consider the possibilities for the compensation of individual artists whose work is used by Generative AI in creating works and secondly how human artists could use Generative AI and avoid potential liability for copyright infringement.
AI and potential licencing agreements
Of concern is the manner in which these machine-learning models are being trained. As mentioned previously, the platforms scrape information from available resources online in order to train the AI, including existing copyrighted works in the data sets. Because of the vast amount of data being scraped, this is done, more often than not, without the consent of the affected artists.
In an ideal world, the companies that create and control AI models should have a duty imposed by law to examine the exact information and parameters of what the AI model is being fed, ensuring that they obtain the consent of the artists whose work make up these datasets. S22(1) of the Act allows for the licencing of one’s copyrighted work. Hence, artists should have the option to opt-in to their work being included in the dataset being used to train AI programmes.
This sort of licensing arrangement is not a foreign concept and has already been successfully applied in other creative industries. Platforms such as Tracklib allow musicians to upload their works through licencing agreements and be fairly compensated for their songs being sampled by other artists. In turn, subscribers to the platform are able to incorporate elements of existing songs into their own without fear of being held liable for copyright infringement further down the line. It is likely that a similar arrangement is eventually deployed for the datasets of AI models.
Navigating the murky waters of copyright infringement
In the absence of licensing arrangements, users of Generative AI must consider potential copyright infringement. According to s23(1) read with s1(2A) of the Act, copyright infringement may occur where one, without the permission of the copyright owner, performs an act that said owner has an exclusive right to do in relation to the whole or a substantial part of the owner’s copyrighted work.
For artistic works, these exclusive rights are outlined in s7 of the Act and include acts of reproduction. In determining whether there has been reproduction of a copyrighted work, the court in Laubscher v Vos & others 3 JOC (W) stated the following:
“The question whether there has been a reproduction is a question of fact which must be taken in two stages, one objective and the other subjective. In order to constitute reproduction within the meaning of the Act, there must be (a) a sufficient degree of objective similarity between the original work and the alleged infringement; and (b) some causal connection between the plaintiff’s and the defendant’s work”.
In summary, the infringing work must not only be derived from the original copyrighted work but must also be objectively similar to it. Therefore, because Generative AI users may not know the extent to which a certain copyrighted work informs the eventual output generated by AI, it is advisable that these users engage with this generated artwork cautiously.
In light of the above, a suggested approach for artists is to interact with the art after it has been generated by the AI in applying their own mind and skill. Used in this manner, the AI may be viewed as an intuitive collaborator that human artists can bounce ideas off of in the early stages of the creative process, after which each individual artists own, unique worldview may inform the final product thereby making the ‘objective similarity’ as referred to above less likely.
Conclusion
Whether solutions to the problems outlined in this article are capable of being found in the existing legal framework or through the potential introduction of new laws (such as the Copyright Amendment Bill) remains to be seen and is best left for the courts and regulators to decide.
It is certain however that AI is here to stay, and therefore, whilst uncertainty around the future of the industry persists, it is imperative that artists become more proactive in protecting and monitoring their intellectual property moving forward.
Contact us today if you have any queries in connection with the above
*Image credit: Westworld Opening Titles, HBO