Artists Up in Arms Over New AI Model That Can Generate Similar Works

Is Stable Diffusion Violating Copyright Law?

A new controversy is brewing in the world of art, as traditional artists are concerned about a new AI model called Stable Diffusion. This model, developed by the startup Stability AI, is trained on an internet-scale dataset of images, and is able to generate artworks in the styles of specific artists. Some of these artists are legitimately worried about their livelihood, as the AI model is capable of creating works that are very similar to their own.

Indeed, it has been clear for a while that many of our jobs will be replaced by intelligent machines in the future. We’ve seen it happen in other industries, such as manufacturing and agriculture. But it is surprising to see that even the creative arts are not immune to this trend.

Some outraged artists are now claiming that stable diffusion is violating copyright by generating works that are too similar to their own. However, it is important to understand that copyright only protects an artist from someone else reproducing one of their actual artworks. It does not protect them from someone else coming up with their own work in the same style. So, legally, it is okay for the AI model to examine someone else’s work, absorb the concepts contained in it, and then apply those same concepts, techniques, colors, and styles to similar subject matter to come up with original works.

In other words, as an artist, you should know that copyright only protects you from someone else reproducing one of your actual artworks - not from someone else coming up with their own work in the same style.

Whether the result is a copy depends on whether what is taken from the original work is a ‘substantial part’ or not. Could Stable Diffusion copy a substantial part of a specific artwork?

While in theory “overfitting” is possible, where AI models memorize part of their training set, common sense - backed by real world experience with this tool, agree that stable diffusion, and other contemporary AI models such as DALL-E and Midjourny, are not capable of substantially memorizing works of art. In fact, the size of the stable diffusion model is only a couple of gigabytes, while it is trained on hundreds of terabytes of source images. In the words of Stability AI CEO and founder Emad Mostaque, this is the collective imagery of humanity compressed into files a few gigabytes big. This immensely lossy compression from the source image to the model, which reduces the volume of data by at least four orders of magnitude, implies that the probability for stable diffusion faithfully memorizing ‘substantial’ parts of specific works of art is extremely slim.

From the perspective of artists, this is a worrying development. Their livelihoods depend on their ability to create original works, and if AI is able to generate similar works, it could devalue their work. If an artist comes up with a new and unique style, it is only a matter of time before an AI model can learn to generate works in that style just as well, if not better.

Artists are important to society, as they help us to see the world in new and different ways. They help us to understand our own emotions, and to empathize with others. If we lose the economic incentive to create original art, it will be a great loss for humanity.

From this perspective, it’s possible that we may need to make some changes to copyright law in the future to protect artists from this new threat. What those changes might be, however, is still up for debate.

Similarity is not enough.

The degree of similarity between two works must be “striking or substantial.” There is no exact formula for what is considered substantial, but merely emulating the style of an artist does not meet the threshold.

It is not impossible for an AI to create a copyright violation, and much larger models in the future may pose real risks. However, given the very small model size used by Stable Diffusion compared to the massive number of images it is trained on, it is extremely unlikely that the model can reproduce a prior artwork in a way that constitutes a substantial copy. It’s akin to asking someone to recreate a painting they saw years ago from memory—our memories are considerably larger and more detailed than Stable Diffusion, but are still usually too vague to result in a direct copyright violation. In my own attempts to reproduce very famous paintings with Stable Diffusion, the results were always substantially different from the originals.

Of course, this discussion is highly simplified. There can be broader protections for specific things, such as fictional characters, which Stable Diffusion could potentially generate. For specific situations, it’s always best to consult a lawyer.