Chat GPT Is Eating the World

Here’s this week’s status report of all 24 copyright lawsuits against AI companies.

(1) The biggest development since last week’s report was Judge Orrick’s tentative rulings on the defendants’ motion to dismiss the first amended complaint in Sarah Andersen v. Stability AI.  

Judge Orrick indicated that he’s “inclined to DENY all motions to dismiss the direct and induced infringement claims under the Copyright Act. Beyond the Training Images theory (that suffices for direct infringement as to Stability, Runway, and Midjourney), plaintiffs have plausibly alleged facts to suggest compress copies, or effective compressed copies albeit stored as mathematical information, of their works are contained in the versions of Stable Diffusion identified. At this juncture, plaintiffs should be allowed to proceed with discovery. The facts regarding how the diffusion models operate, or are operated by the defendants, should be tested at summary judgment against various direct and induced infringement theories and precedent under the Copyright Act.”

At the hearing this week, a key point of contention between the two sides was whether the plaintiffs must allege examples of specific outputs that are allegedly substantially similar to the plaintiffs’ works to support their allegation that the Stable Diffusion models actually contain “compressed copies” of the plaintiffs’ works. Or can the plaintiffs rely instead on other, more general allegations, such as alleged admissions by Stability CEO Emad Mostaque regarding compression in its models and a research paper discussing compression by Yadong Yu et al, as alleged in the first amended complaint.

Even if Judge Orrick allows this theory to proceed, the same question would arise at the summary judgment stage, albeit there with what evidence must be presented to raise a genuine issue of material fact that so-called “compressed copies” of the plaintiffs’ works actually exist in the models.

(2) Judge Chhabria was assigned to preside over Dubus v. Nvidia Corp.

(3) There’s a major discovery dispute in Authors Guild v. OpenAI. The plaintiffs want to compel OpenAI to produce its interrogatory responses submitted to the Federal Trade Commission’s investigation of OpenAI, as well as document related to the identities of 2 former OpenAI employees who supposedly were responsible for creating the “books1” and “books2” datasets that were used to train OpenAI’s model, but were later destroyed. OpenAI opposes the motion to compel. Business Insider reported this controversy earlier this week. In a research article published in 2020, OpenAI researchers revealed that books1 and books2 datasets composed 16% of the weight in training GPT-3 and consisted of 57 billion tokens.

The new developments from last week’s report are listed with an asterisk(*).