The case for training AI’s on publicly available copyrighted material
Artificial Intelligence (AI) systems have taken the world by storm. The release of chat-gpt by OpenAI has catapulted a branch of computer science confined to conferences and journal articles into the mainstream. More recently a new branch of AI systems was introduced to the world, so called Generative AI. It allows users to generate images based on text-prompts and its overall prowess leads inevitably to various questions. What defines art? Can we categorize AI-generated art as true art? And how do we deal with the prior works the systems rely on? These AI models require extensive training on thousands, even millions, of images to ‘learn’ how to create their own art. But these images did not just pop into existence having been created through the labor of millions of artists around the world. Do developers need to seek consent from every owner of each image used for training? Furthermore, should these image owners be entitled to compensation? We believe they should not.
What is art?
A question with no definitive answer. The Merriam-Webster dictionary offers multiple definitions, being ‘skill acquired by experience, study or observation’, ‘a branch of learning’ and ‘the conscious use of skill and creative imagination especially in the production of aesthetic objects’. The creation of art is as old as time, with cave paintings dating back 45,500 years having been found in Indonesia. However throughout all these years it remained a strictly human affair. With technology advancing rapidly, it is now possible to teach generative AI to create new artworks by feeding the model hundreds of millions of images or paintings. The launch of Dall-E, a publicly available tool enabling anyone to create AI art with text prompts, has drawn increased attention to this topic and sparked a crucial discussion: do we consider AI-generated art as ‘art’?
We argue the answer to this question is ‘yes’. Art can be anything and the beauty of art lies precisely in this openness. The value of art is impossible to determine and lies in the eye of the beholder, no matter how it was created. Any work that resonates with a person, that makes you feel ‘something’ can be considered art. There should be no boundaries to what art can be. Furthermore, the way art is created has changed over the centuries as people found new techniques and materials. Why should it be any different with AI as a new technique for creating art?
A counterargument often heard is that AI cannot generate ‘new’ art, but only recombine and rearrange art from the training data. This is not true; already a specific method called Creative Adversarial Networks (CANs) can generate art that is also classified as ‘art’ by the model but that cannot be classified within an existing style, meaning that the art it outputs has to be novel. Another counterargument is that AI-generated art does not reach the quality of human-generated art, let alone surpass it. While this may have been true some years ago, technology has developed so quickly that studies now show that humans cannot distinguish human-generated art from AI-generated art when this new method of CANs is used. If this is the case, then why would AI-generated art be of lesser value than human-generated art?
Having established this, we shed light on the process of how art is created and we argue why the creation of art by humans is not all that different from the creation of art by AI. We as humans have always drawn inspiration from the things around us – nature, tales, and especially artworks from other artists. Whilst a painter needs to time to acquire their skills they do not pay the creators of the works they are studying. Artists have been using previous work as inspiration since time immemorial. As Pablo Picasso once put it “Good artists borrow, great artists steal”. Is training generative AI on millions of images of artworks not comparable to looking around, going to museums, and being inspired to create our own artworks? The extent to which we draw inspiration from different sources and artists is a ‘black box’: we do not know what happens between seeing art and creating new art, can rarely quantify to what extent a seen piece influenced another created. Generative AI have a very similar process. Images are used to learn and extract various features from, but in the end novel things can be created and exact influence of particular pieces is hard to pinpoint. If this is all so similar, then why should we hold machines to a different standard then we do humans?
“Good artists borrow, great artists steal”Pablo Picasso
Lets tackle this problem from another perspective and consider a world in which every copyrighted piece of art in a training set is due some compensation. A generative AI System like DALL-E 2 is trained on hundreds of millions of images. We argue that there is no level of compensation here that would lead to a desirable outcome. If you priced each access at say one dollar, the initial capital needed to train such a model would already price out small businesses or startups and not even give that much to each artist. If you contributed 100 artworks to the data set and got 100 dollars out of it, that would not offset any future economic worries of being outcompeted in the slightest. Moving the amount paid in both directions exacerbates both issues – either the payment becomes meaningless to the artists involved, or there is so much initial capital required you would seriously hinder innovation and only enable the biggest players in the economy to compete. Trying to pay people based on their contribution to a generated piece of art seems also unfeasible. These large models use billions of different parameters to eventually generate an output and are being called blackboxes for a reason. The try and pinpoint exactly what images contributed in what manner seems far fetched at best.
There is also the matter of innovation and the race for technological supremacy. AI will change the world in the coming years in ways we cannot even imagine right now. Whilst it is important to not forget about humans along the way, it is imperative to enable research and development as much as possible. And with nations such as Japan already confirming and codifying that training on images is no copyright infringement, getting left behind in the race is a real possibility.
Generative AI models have gone from a purely theoretical possibility to a concrete reality within the last few years. The cat is out of the box so to speak. And just because we don’t think training these models on publicly available material equates copyright infringement it does not mean that the artists perspective should be disregarded entirely. The impact these systems will have on artists all around the world is hard to predict. However allowing people to opt-out of their works being used to eventually replace them seems like the right thing to do, a practice open-ai has started doing already. Works that are out for the public eye to see can (and should!) be used in all sorts of manners, but if the artists explicitly does not want it to be used for some purposes, respecting their wishes is the least one can do. There is various lawsuits running right now that will determine how the courts are going to treat copyrighted training material, but regardless of how these turn out there is a conversation that will need to happen eventually. As generative AI models continue to improve they will take away an ever increasing amount of jobs from human artists. That will not stop people from expressing themselves, but if we value human artists we have to find some way to sustain them in the future. The great AI replacement will come. How we deal with the consequences is still up to us.