U.S. District Court Excludes Expert Testimony After AI Agent Prompts Withheld in Flexport v. Freightmate AI
Bot Mutiny |
A federal judge has barred expert testimony regarding AI-driven data extraction after the defense failed to disclose the specific prompts generated by a 2026 AI agent to conduct the experiment.
Judge Rita F. Lin of the U.S. District Court for the Northern District of California issued an order on September 18, 2026, granting a motion to exclude specific expert testimony in the trade secret dispute between Flexport, Inc. and Freightmate AI, Inc. The decision, documented in a pretrial order first reported by CourtListener, centers on the failure of a defense expert to disclose the underlying prompts used by an AI agent to conduct a technical experiment. The ruling addresses an experiment conducted by a defense expert named Armstrong. He used GPT-4o to ingest shipping documents, claiming the model could extract data with up to 85% accuracy using industry-standard metrics. To run this experiment, Armstrong used a 2026 AI agent, Claude Sonnet 4.5, to generate the prompts and the 33-field schema provided to GPT-4o. ## The Black Box Problem Flexport challenged the reliability of the experiment because Armstrong didn't disclose the prompts the AI agent sent to GPT-4o. The court agreed. Without these prompts, it's impossible to know what the agent actually told the model to do. Judge Lin noted that a new version of the experiment might result in the AI agent sending different prompts. If the agent provided detailed instructions that made the extraction easier, the experiment's results would be skewed. The judge likened it to an expert telling a human assistant to perform a task without knowing what steps the assistant actually took. Because the defendants didn't provide the "verbatim extraction prompts" generated by the AI agent, the court ruled they failed to meet the burden of showing the testimony was reliable. The motion to exclude Armstrong's testimony regarding the experiment was granted. ## Trade Secrets and Token Usage The order also addressed motions regarding another expert, Zeidman. The court denied Flexport's request to exclude Zeidman's opinions on whether Flexport's information was secret or "readily ascertainable."
Zeidman is also permitted to testify about the lack of literal copying in Freightmate's source code. The court found this relevant to whether the defendants relied on Flexport's trade secrets. Additionally, the court upheld Zeidman's analysis of OpenAI tokens consumed by Freightmate, which relied on testimony from the startup's CTO. ## Corporate Size and Animosity Flexport successfully blocked Freightmate from using a "David and Goliath" narrative at trial. Judge Lin ruled that references to Flexport being a large company bullying a startup are not relevant and are unfairly prejudicial. However, the court won't exclude all mentions of company size. Freightmate's resources and stage of development may still be relevant to liability and damages. The judge also ruled that evidence of animosity from Flexport higher-ups toward Freightmate could be used to show witness bias, though evidence of Flexport's "motive to sue" remains excluded. The trial involves claims of misappropriated trade secrets, including Flexport's platform source code. While Flexport voluntarily withdrew its copyright claim, the court ruled that evidence overlapping with the remaining trade secret claims is still admissible. The case is Flexport, Inc. v. Freightmate AI, Inc., Case No. 25-cv-02500-RFL, in the Northern District of California.