Dear Doc,
Many questions about artificial intelligence (AI) remain up in the air as judges try to keep up with technology that is developing faster than the speed of light. It seems to me that many of the issues might be solved if we just kept the AI companies from putting copies of other peoples’ materials in the AI models that they build. Can’t we just do that?
Signed,
Every Author, Photographer, Artist, Musician, etc.
Dear Every:
That’s a great thought, and it’s central to a question recently addressed by the High Court of Justice of England and Wales in the case of Getty Images v. Stabiility-AI. In a 201 page opinion, Mrs. Justice Joanna Smith, DBE, addressed a complicated case in which Getty Images, which owns a database containing over 477 million photographs and other images, charged Stability-AI with copyright infringement. Stability-AI markets a visual artificial intelligence program called “Stable Diffusion” that used Getty’s images to “train” Stable Diffusion to reproduce nearly identical images from the Getty database when prompted to do so.
As Justice Smith explained,
Stable Diffusion is a type of generative AI model known as a diffusion model, or more specifically a latent diffusion model. Broadly, it transforms an input (a user command or “prompt” in the form of written text or a “seed” image) into a desired output in the form of a synthesized image by modeling a probability distribution based on its training data and then sampling from that distribution. The development of a stochastic model of this type typically involves designing and building the architecture for the model which is then trained by repeated exposure to massive quantities of data, in this case in the form of human generated digital images contained in datasets created by crawling and scraping images and associated descriptive captions from the Internet.
The model parameters (the “model weights” or “biases”) define the network connections in the model and are learnable parameters controlling the functionality of the network. Before training begins, the network’s weights are initialized with random values. As the network is exposed to the training data, the weights are iteratively updated to better satisfy an optimization criterion specified by engineers. Once the model is trained, running the network, referred to as inference, is (in simple terms) an input-output system in which the user specifies inputs, the trained network performs computations on those inputs and then generates the desired output.”
(Opinion, p.1)
The central question that the case addressed was whether the trained model contained copies of the images used in training, and was thus an infringement of copyright.
The Court explained at great length about the processes used in training. This included that compression of digital images, the introduction of random noise into each compressed image, and the care taken not to “over expose” the model or to permit simple memorization by the model. There were other issues brought to the Court, including a charge that use of Stable Diffusion was capable of producing images that contained trademarks owned by Getty (which were used as “digital watermarks” on the images that were part of the training data. The Court agreed that this was possible, but held that there was no way to know how many images had been generated that contained the trademarks.
In the end, the Court concluded, “[A]n AI model such as Stable Diffusion which does not store or reproduce any Copyright Works (and has never done so) [and thus,] is not an “infringing copy”.
The Doc has no doubt that this is not the last we have heard of this issue, as AI models continue to devour all types of information. The Doc is amazed by what today’s AI systems can do that is undeniably useful, but he still believes that people should be taught research and critical thinking skills that do not depend on AI before they use these new digital assistants. Stay tuned here for more updates. (This was not written using AI in any manner, though the Doc did recently give a long lecture about how AI was changing the field of law, for which he used several AI tools to create his Keynote slide deck!)
Have a question about intellectual property law? Give the attorneys at LW&H a call or an email. They keep up to date on this stuff (and have their own Ai models to prove it!)
Until next month,
The “Doc”
Using Stable Diffusion, prompted with “Generate a portrait of an attorney who is known as “Doctor Copyright” in the style of Rembrandt”
— Lawrence A. Husick, Esq.