Artificial General Insanity: Concluding thoughts in 2025

As the year comes to a close, of course we have to talk about AI (don’t hate me) Personally, I take the view that the technology jumped ahead without the necessary safeguards. Because the organisations involved were “too big to stop” and “pursuing the greatest frontier of human existence”, or whatever other grandiose terminologies have been used to describe the competition in the AI arms’ race, we now have ourselves in a situation where creators, artists and writers have essentially been left holding the short end of the stick as the “weaker” parties in the scheme of things. Why do I say this? Companies like OpenAI might say that they obtained access to the books, art, or other original works by legal means, e.g. by purchase of an e-book or scanning a physical copy of a book, or purchasing a copy of an artwork. But such use is not as innocuous as it sounds. Why? Because the purchase of the book does confer on a reader any copyright at all. You are not paying to get a license of an author’s copyright. Instead, you are paying to obtain a copy of the book (which is the author/ publisher’s right to copy and sell). Many GenAI companies have however circumvented this entirely, and claim that we have entered “a new area of boundless creativity” (whatever that means), where anyone can be an artist, a musician, a writer. The questions we are left grappling with are of course multifaceted: from an ethical viewpoint, is there an issue with the large companies building models on the backs of thousand or millions of artists and creators for little to no remuneration? From a legal viewpoint, does such use infringe existing copyright laws, or are changes required to properly find a balance in the mix? One school of thought is that generative models are not infringing because they usually provide substantially dissimilar outputs and combine features from multiple sources instead of copying chunks from a single source. Whether this is indeed true is open to debate however, and could depend on the nature of a prompt provided to a model and the availability of alternative sources in response to the prompt. It may also depend on the guardrails (if any) implemented in a model specifically on the question of copyright infringement. For example, a substantial similarity comparison may be conducted on generated output against source material to ensure substantial/sufficient dissimilarity. On the other hand, as text and data mining used for generative models usually involves the making of copies or extractions of copyrighted materials as inputs, this may itself also be regarded as infringing conduct if one takes a strict interpretation of copyright law that does not only focus on the outputs, but the creation of copies in any stage of the content creation process for the model. Regardless, given the speed, volume and scale of these models, this thrusts robustness of existing copyright laws into question all over the world, with litigation abounding in many jurisdictions as a result. Several countries such as Japan and Singapore have implemented rules pertaining specifically to the treatment text and data mining (“TDM”) under copyright law. The U.S and the EU also possess TDM frameworks for this purpose. Each jurisdiction in doing so is trying to find the right balance between technological advancement on the one hand, and the legitimate protection of copyright holders on the other. Copying and the training process A relevant consideration proposed by Dr. WooJung Jon in his article, Comparative Analysis of Text and Data-Mining Exception Clauses in South Korea, Singapore and Japan (2025) 37 SAcLJ 832 touches on the actual technical aspects of how models produce images. He highlights that rather than storing entire images these models learn statistical patterns – such as edges, colours, and shapes – and encode these features into high-dimensional weight matrices (this generally applies to diffusion models). As for text-to-image models, the model identifies associations between textual descriptors and underlying visual representations and encodes statistical patterns and features, rather than the exact pixel arrangement of the original training image. In other words, the model identifies patterns from the training data, which are then encoded and used to generate new content. When a prompt is entered, the model identifies associated patterns and produces the output based on the encoded patterns, not based on the actual raw images in the training set. Dr WooJung therefore points out that even where a substantially similar image is produced, this is not a result of copying the original image, but a product of the encoded statistical patterns. An analogy here might be if two people went to the Grand Canyon and ended up drawing substantially similar images without having referred to one another’s work. In this case, no act of copying has taken place and thus no infringement can be found regardless of how similar the images are. On the other hand, it may be argued that this does not bring models outside the purview of copyright simply because a similar image was produced by reference to encoded patterns rather than the original image. While the source of the output is not the “original image”, it may be argued that such output is nonetheless derivative of or a result of the use of the original image. An example that springs to mind is the recent experiment where a starling had recreated a digital image from sound hours after hearing it : https://www.yahoo.com/news/articles/scientists-fed-birdsound-made-202800622.html. While the reproduced image would be a result of the sound code repeated by the starling, to state that the final reproduced image is not properly regarded as a “copy” of the original might not also be entirely accurate. Of course, this is not a perfect analogy, but the larger point here is that these models may arguably be making use of technicalities to circumvent existing copyright restrictions, or may otherwise reveal lacunas in existing copyright laws, which may not be designed to deal with the technical aspects of models like diffusion or text-to-image models. The CDA exception in Singapore

Another point that arises therefore are the exceptions that appear to cover TDM and other machine learning uses of works. Section 244 of the Copyright Act allows for the use or making of copies of works for the purposes of “computational data analysis”, which is non-exhaustively defined in Section 243 as including “using a computer program to identify, extract and analyse information or data from the work”. CDA also includes using a work “as an example” or as “information or data” to improve functioning of a computer program in relation to that type of information and data”. The illustration in Section 243 refers to the example of training a program to recognize images. While the broadness of the CDA exception lends to the view that use of works or recording for the purposes of training models is a permitted use, the question arises whether this would still be the case where a substantially similar output is generated from the model (since this goes beyond the bounds of data analysis, and instead goes towards the use of such data for output generation).

On one view, as indicated by Dr. WooJung, because of the technical aspects of these models which look at statistical patterns for output generation as opposed to the original images, this last step in the content creation process may be regarded as non-infringing, while the model training portion of the process is protected by the CDA exception. The question of course is: is this the desired or desirable balance to be struck in copyright law? And should different types of content be regarded differently? Many are of course in favour of untrammeled expansion of such exceptions as “drivers of innovation and creativity”. But this begs the question, what incentive is there for artists to create new art when it will just be fed into the machine? Will this spell the end of new content altogether? Or will all new content come from generative models, giving new meaning to someone’s work being described as “derivative”? The longer we ruminate over this, the further away we are from being able to change the course of progress. As it is, we have already seen volumes and swathes of data being swallowed up by the “giants of AI”, with little to no transparency on the sources of data and how these were obtained. This is not a situation where we can turn back the clock and undo the use of works after the fact. New content is being generated every single day based on the works of human creators. Don’t get me wrong, I’m not saying AI is bad or should be abolished; but until a framework that properly balances the interests of creators is established, charging ahead with a “move fast and break things” mentality that characterises the tech ecosystem might not be the best answer, and could cause irreversible damage to an already struggling group of creators who pursue their craft with passion and originality. Sorry for the doom and gloom – here’s to a better 2026! Stuart Peter, Human

*This entry was written at the end of 2025 as a reflection on the state of the art at the time and the legal challenges and issues that had arised over the course of the year.