Artificial intelligence is quickly becoming a go‑to tool for aggregating and summarizing large volumes of data, formulating and testing arguments, and even sketching litigation strategies. But a recent ruling from the Southern District of New York serves as a stark warning: when clients turn to generative AI for legal strategy, they may be unknowingly turning privileged information over to a third party and then creating documents that may later be discoverable in litigation. In a closely watched bench decision, Judge Rakoff ruled that AI‑generated documents created by the target of a criminal investigation using Anthropic’s Claude were not privileged despite being generated with information learned from his attorneys to support his potential legal defense and then shared with his counsel. The decision highlights the unresolved and increasingly consequential intersection of AI, privilege, and discovery.

Facts

Bradley Heppner received a grand jury subpoena and hired attorneys at Quinn Emanuel to represent him. After learning he was a target of the investigation, but before he was arrested, he created 31 documents with Claude using information from his attorneys to outline a potential defense strategy. He was later arrested on charges of securities and wire fraud, and federal agents seized his electronic devices, which contained the 31 documents that had been provided to his attorneys. Mr. Heppner argued that the documents were created to prepare his potential defense strategy in anticipation of an indictment, but he conceded that he made the decision to prepare the reports on his own, i.e., not at the direction of counsel. He nevertheless claimed the documents were protected from disclosure by the attorney-client privilege and work product doctrine; the government moved to overrule the objections.

Decision

On February 10, Judge Rakoff heard arguments from counsel before ruling from the bench that the 31 documents at issue were not privileged or otherwise protected from disclosure. His decision appears to have been driven by two key factors.

First, Judge Rakoff noted that the 31 documents did not reflect the mental impressions, conclusions, and opinions of Mr. Heppner’s counsel, nor were they prepared at his counsel’s direction. Rather, Mr. Heppner made the decision to create them on his own volition.

Second, and more importantly, Mr. Heppner disclosed information to a third party (Claude) whose privacy policy explicitly warned users that data on the “prompts” entered, and “outputs” generated would be used to train the AI tool and subject to disclosure to regulatory authorities and third parties.

In sum, the “outputs” were not created by an attorney or otherwise reflective of an attorney’s mental impressions or legal theories, and any privilege that could have attached to the “prompts” was lost due to Mr. Heppner’s sharing of information with a third party.

Key Takeaways

As discovery disputes over AI-generated documents make their way through the trial and appellate courts throughout the country, we should expect to see divergent and even contradictory decisions on these nuanced privilege issues. However, the outcome of the dispute in Heppner provides two important takeaways for clients and attorneys as they face these issues.

1. AI Tools Should Be Treated as Third Parties: It is safe to assume that not all information shared with an AI tool (g., a privileged memorandum from outside counsel) or the resulting outputs will be privileged and that both the prompts and outputs could be subject to disclosure in litigation. While some AI companies offer enterprise-tier agreements that exclude your data from training and provide a level of confidentiality, the AI tool does not have a law license, and the attorney-client privilege may not extend to data voluntarily shared with the platform. If the outputs are created at the direction of counsel and the incoming/outgoing data is held confidential by the AI software, courts may find the data to be privileged and not subject to disclosure, but there are no bright-line rules or tests on this topic so far.
2. Source of Discovery: Companies should expect to receive preservation notices and document requests for their custodians’ data that were shared with and received from AI software. Sources of AI data will soon become the hot topic of discussion during early ESI and discovery planning conferences. This can create a logistical nightmare if employees are permitted unfettered access to AI tools on their company-furnished laptops and cell phones (g., Claude, ChatGPT, Gemini), as AI tools have different data retention policies and settings. We expect to see motions for sanctions and spoliation arguments surface as litigants discover, for example, that the prompts used to generate default notices to a series of subcontractors were destroyed after the general contractor was on notice of its duty to preserve ESI. Training employees on permitted uses of AI and their obligations to retain data when doing so is critical.

It is important that attorneys and their clients are mindful of the potential ramifications when sensitive (and perhaps once privileged) data is shared with third-party AI software. To conclude, it is important to consult with your attorneys if and when questions arise surrounding obligations to preserve data or the use of AI to summarize documents to assist with potential or ongoing litigation.

Please contact Christopher Olsen, Freddy Muñoz, or Gary Stein with any questions.