Original article:
Summary: visual artists accuse Stability AI, Midjourney and DevianArt of misusing their copyrighted works in connection with the companies’ generative artificial intelligence systems.
Allegations: Stability (AI system) used billions of images “scraped” from the internet, including theirs, without permission to teach Stable Diffusion to create its own images.
Outcome: judge dismissed all allegations against Midjourney and DevianArt. Two out of three plaintiffs had their claims dismissed, but the third was allowed to pursue claim of alleged use of her work to train AI systems.
Key ratio: Judge Orrick said allegations are unlikely to survive if they’re based on the AI systems’ output without showing that the images were substantially similar to the artists’ work.
My commentary:
Orrick J is, in essence, confirming the “artist as rearranger” view of creativity. To situate the principles of this case, I will use a common model of creativity. The training of visual artists, much like that of musicians, is comprised of some version of the following stages: emulation, variation, innovation.
The first stage involves learning to copy exemplary masterworks and to master the techniques that went into their production. Output at this point is largely replicatory. At the second stage, the artist innovates by changing some essential constitutive elements of the work. Not all elements, but just enough to keep the work grounded in an artistic tradition, while being modified through the artist’s own experience and skill. The third stage – one that very few artists ever reach – is one of innovation and transformation. Here, the artist has so thoroughly mastered their craft, that their output completely redefines the genre.
Some examples: Johann Sebastian Bach revolutionized every musical genre he composed in. He was masterful at all three stages. On the other hand, Wolfgang Amadeus Mozart made brilliant, genius, contributions to several musical genres, but he contributed very little to their development. Essentially, opera was the same before and after Mozart. The church cantata as a genre, on the other hand, was radically different after Bach wrote hundreds of them.
Judge Orrick seems to indicate that the output of Stability AI falls within stage three of the model described above (innovation). Yes, the output is – by definition – comprised of constitutive elements of other artists. However, the end result is substantially different from those originating works.
What would Orrick J say about stage one (emulation) and stage two output (variation)? Pure emulation, it seems, could most likely be seen as copyright infringement, due to the lack of original contribution from the artist and a purely mechanistic reproduction of other works. Stage two output will be the most problematic. How much of the final product is identifiable with another product? To what extent? Has the artist exercised enough skill and judgement? How much skill and judgement is sufficient to push the product closer towards stage three output on the spectrum.
Hopefully future case law will delineate this for us.
Or I could ask ChatGPT for an answer…
Communications Law