HELD: LLM prompts are not attorney work product, particularly when the output of the LLM in response to certain prompts is a key issue in the case. Here authors claimed OpenAI illegally reproduced the copyrighted work of authors. In discovery, OpenAI requested production of the prompts utilized by plaintiffs to generate the LLM output that they proffered to the court as evidence. No doubt this discovery request was driven by a learning in the The New York Times case against OpenAI, where it appears the NYT may have fed lengthy segments of articles into ChatGPT and then asked ChatGPT to finish the article. Plaintiff's counsel refused to produce the ChatGPT prompts used to obtain its evidence, claiming that such prompts are attorney work product because they would reveal the attorney's thinking and strategy. The court (correctly) didn't buy it because: a. the LLM output was introduced into evidence, and thus all the accompanying queries need to be made available in order to evaluate whether such evidence is a likely output of the model under a typical use case. [This argument reminded me of the SNL courtroom skit, which I will link in comments below, where the defense attorney for Jurassic Park shows a 5-second clip of the plaintiffs having fun at the park and then wants to stop the video before it shows the dinosaurs attacking.] b. "[O]pinion work product consists of the attorney's interpretation of legal theories and the application of facts to those theories, rather than the bare facts or legal theories alone. The account settings and negative test result sought here are more in the nature of bare facts." c. The suggestion that OpenAI "can simply interrogate ChatGPT themselves" is laughable when OpenAI doesn't know the prompts used by the plaintiffs to achieve their presented output. I think we will find that the plaintiff authors (i) engaged in elaborate, uncommon prompt engineering to achieve their results that were put into evidence, (ii) violated OpenAI's Terms of Service in using ChatGPT to achieve their results, and (iii) attempted many other prompts that failed to produce sufficient copyright infringement.
Elizabeth Rader
5mo
Fed. R. Evid. 106 comes to mind, though the issue here is discovery not admissibility.
Neil Peretz Very insightful. Thank you for sharing
Scaling the Law
5moThe great SNL courtoom skit I referenced above: https://youtu.be/6F8mJZkP-Hg?feature=shared