Prompts, privilege & discovery
Tremblay v. OpenAI, Inc.
Attorney-crafted prompts and negative tests were opinion work product; only the interactions underlying examples used in the complaint had to be produced.
Attorney prompts protected01
Case background
Material facts
OpenAI sought authors’ pre-suit testing prompts, outputs, settings, and unsuccessful results after the complaint disclosed selected examples of allegedly infringing output.
02
The decision
The court’s ruling
The district judge set aside a broader magistrate order. Attorney-crafted prompts and failed tests reflected counsel’s thought process, and disclosure of successful examples did not waive protection for unrelated testing. Prompts, outputs, and settings for examples actually used in the complaint remained producible.
Why it matters
What the decision means
Tremblay v. OpenAI draws the foundational distinction. Counsel’s exploratory testing qualified as opinion work product, but the inputs and settings behind examples used in a pleading remained discoverable.
Limit of the ruling. This is a trial-level, fact-specific ruling. The action was later transferred into the Southern District of New York MDL.
Primary reading