The latest briefing in Silverman v. OpenAI reads like that old REM song, “The End of the World as We Know It.” OpenAI has responded to the Plaintiffs’ claims that OpenAI’s popular platform ChatGPT has infringed their copyright with disaster-laden references to Michael Jordan and “the future of artificial intelligence.”
As we’ve previously written, actor/comedian Sarah Silverman and author Paul Tremblay are among a number of plaintiffs who have bought lawsuits alleging that use of their creative content to train large language model (“LLM”) platforms like ChatGPT violates the U.S. Copyright Act.
OpenAI moved to dismiss the majority of claims, arguing, among other things, that ChatGPT’s output cannot be infringing, because it is not “substantially similar” to Plaintiffs’ works. Plaintiffs hit back, noting that OpenAI hasn’t moved to dismiss the “core claim” in the lawsuits—direct infringement by OpenAI—and arguing that “substantial similarity” is irrelevant because OpenAI copied their works wholesale, without permission, in order to make ChatGPT work. Plaintiffs further argued that the LLMs incorporate their works, and thus are infringing “derivative works.” Because ChatGPT’s answers in response to user prompts come from those LLMs, Plaintiffs contended that ChatGPT’s output must necessarily be infringing too.
OpenAI has responded to Plaintiffs’ arguments with a doomsday scenario, arguing that “the future of artificial intelligence may turn” on how the court ultimately decides this case. In its reply brief, OpenAI contended that, under Plaintiffs’ theory, every single answer given by ChatGPT to a human user is infringing. For example, OpenAI argued, if a user asks ChatGPT who the greatest basketball player of all time is and ChatGPT responds “Michael Jordan,” the response would constitute infringement even if His Airness was never referenced in plaintiff Sarah Silverman’s book. OpenAI urged the court to reject this approach, and to focus on the fact that the plaintiffs cannot point to any ChatGPT output in response to individual user prompts that is substantially similar to Plaintiffs’ works. Indeed, OpenAI argued, Plaintiffs’ theory flies in the face of well-established copyright law, and the manner in which courts evaluate copyright claims.
OpenAI used another illustrative example, this time of a film student who records a copy of the film Citizen Kane to study its cinematography. The student watches the film, takes notes, and creates their own film as a tribute to Orson Welles. The fact that the student “directly copied” the original broadcast has absolutely no bearing on the question of whether the student’s film is ultimately “substantially similar” to Citizen Kane, OpenAI posited.
According to OpenAI, Plaintiffs’ own pleadings doom their claims, both as to copyright infringement and as to violation of Section 1202(b) of the Digital Millennium Copyright Act, which prohibits the removal of copyright information to conceal infringement. As OpenAI pointed out, Plaintiffs themselves admit that ChatGPT outputs are not verbatim copies of the original books used for training—and often do not even slightly resemble the originals. Thus, they cannot be infringing. Further, OpenAI contended that Plaintiffs “ignore that the ChatGPT outputs they attached to their Complaints include the very [copyright information] they allege OpenAI ‘removed,’” a fatal blow to the Section 1202(b) claim.
Ultimately, it seems that OpenAI is taking the position that Plaintiffs are trying to shoehorn their claims into a legal framework that does not support them. In OpenAI’s view, adoption of their position would both run contrary to longstanding legal precedent and also ring a death knell for LLMs. These platforms are not trying to supplant the original authors in the marketplace, OpenAI argues; they are simply using material available for public consumption in order to “understand human knowledge.”
It will be interesting to see how the court views these issues in the motion hearing noticed for December 7, 2023.