30 November 2023

Does Artificial Intelligence infringe copyright?

On 30th October 2023, Judge William H. Orrick of the Californian District, with his Order, ruled on Stability AI’s motion to dismiss the appeal presented by Sarah Andersen and other authors, who accused the famous company, to which belongs also Midjourney, for copyright infringement for copying thousands of works they created.

The appeal reported the complaints of the authors on the violation of many rules and the request to sentence Stability AI, Midjourney and others, for direct infringement, indirect infringement – acted through the intervention of a third party – for removing the so-called “CMI” – copyright management information – which consists in the indication of the name of the author and of the information on the work, as well as for unfair competition.

The judge granted all motions, nevertheless allowing the authors to be able to amend their claims, but dismissed the most important one: copyright infringement.

This decision has been considered by many as a victory of the authors against artificial intelligence producers, but it is not quite so.

It is necessary to point out, first of all, that this measure was issued in a preliminary phase of the proceeding. At this stage the Judge is expected to decide on the basis of the acts and must consider as true the statements of the plaintiff, with respect to which s/he must only evaluate if enough indications of credibility have been provided.

Therefore, not only the judgement is not final, but there is still a lack of a proper investigation, that is the stage of gathering the evidences necessary to prove the merits of the parties.

The consequence of Judge Orrik’s Order is not a punishment of Stability AI, but only the affirmation that the process will happen and that in the next months they will enter in the thick of the discussion.

That being said, what is really important to highlight in this decision is that it was issued on the basis of the plaintiffs’ statement that seems to be technically incorrect and, not coincidentally, they contradict in the context of their legal acts.

The premise is that the system of artificial intelligence, during the training phase would create “compressed copies” of the copied works on the basis of which, thus, the user by asking specific questions to the system could generate other works (the outputs of AI) that, moreover, would come out from those compromised copies.

Stability AI highlighted how all this is untrue. In its motion it is said:

“Stable Diffusion had been trained on billions of images available on the Internet. To be clear, to train a model does not mean to copy or memorize images for a later distribution. As a matter of fact, Stable Diffusion does not “memorize” any image. The training, rater, implies the development and refinement of thousands of parameters which define collectively – in the learned way – how things appear. Lines, colors, nuances and other attributes associated to countless subjects and concepts. The aim of this is not to enable models to reproduce copies of the training images. If someone wants to copy images from the Internet integrally, there are many easier methods to do it, including copy-paste of an image and maybe the use of tools such as Photoshop to modify them. Moreover, Stable Diffusion is not a “collage tool”. Stable Diffusion, however, allows users to create completely new and unique images by using simple words suggestions. As recognized in the same opposing appeal “none of the Stable Diffusion output images provided in response to a particular text suggestion is likely to closely correspond to a specific image in the training data”. Therefore, if this court will have to examine the merits of any claim of copyright infringement complained by the plaintiffs involving Stable Diffusion, Stability AI is confident that there will be one only possible conclusion: Stability AI enables creation; it is not a copyright infringer.”

This statement was not considered and explored in this phase of the proceeding for procedural grounds, but it will be during the debate.

It will be very interesting to find out more about the functioning of these systems that, with the exception of their developers, no one knows in depth.

Categories

• Digital & IP
• Eventi
• Non categorizzato
• Turini News