This issue raises developments concerning the intersection of evidentiary privileges and use of AI tools; public record requests about AI use; and AI hallucinations in legal research.

Issue 1: Opposite Court Rulings on AI‑Generated Materials and Privilege

It was only a matter of time before a court ruled on whether prompting an AI tool would waive privilege. But no one expected two federal courts to issue rulings on the same day. That happened in February, and the courts reached opposite conclusions. Below is a brief description of the cases, takeaways, and excellent articles on the topic.

United States v. Heppner, No. 25 Cr. 503 (JSR), 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026)1 In November 2025, the defendant, Bradley Heppner, was arrested and charged with securities and wire fraud. Federal agents seized devices from Mr. Heppner’s home containing AI-generated documents. After Mr. Heppner received a grand jury subpoena and spoke with his legal counsel, he used the free consumer version of Anthropic’s Claude, prompting the model to outline his potential arguments and strategy, which he later shared with his attorney. On February 10, Judge Jed Rakoff of the Southern District of New York issued a decision from the bench, finding that documents the defendant created using Claude and sent to his attorney were not protected by the attorney‑client privilege or the work‑product doctrine. Judge Rakoff concluded that the AI‑generated documents were not privileged communications for three reasons. First, the court noted that Claude is not an attorney, so no duty of confidentiality was owed to Mr. Heppner. Second the AI‑generated documents were not confidential because Anthropic’s privacy policy states that data entered by consumers is used for training its AI models unless users opt out2 and is subject to disclosure in response to valid third party process or in the event of “claims, disputes, or litigation.”3 And finally, because Mr. Heppner used Claude independently and not at the behest of counsel, the attorney-client privilege did not extend to a third party—in this instance Claude—under the Kovel doctrine. The doctrine protects communications with third-party experts when they assist counsel in providing legal advice.4

Warner v. Gilbarco, Inc. et al, No. 2:2024cv12333 – Document 94 (E.D. Mich. 2026)5 Pro se plaintiff Sohyon Warner filed a civil employment discrimination case using ChatGPT to help prepare and draft filings. When defendants discovered Warner’s use of the tool, they moved to compel production of all AI logs and responses. The court found that the AI logs and generated materials reflected the plaintiff’s mental impressions in preparation for anticipated litigation and were therefore protected by the work-product doctrine. The court also concluded that ChatGPT is a tool rather than a person, and therefore its use did not constitute a waiver of that protection under Sixth Circuit jurisprudence, which requires disclosure that directly or indirectly exposes the information to an opposing party.

More to come on this issue. I will note that some commentators, including Georgetown Law School Professor Jonah Perlin, have raised questions about the Heppner court’s decision to apply Anthropic’s terms of service—such as provisions on training data and on disclosure of user data to third parties, including “governmental regulatory authorities”—to confidentiality. They argue that this approach could implicate other digital tools and third‑party software, including virtual assistants, search engines, and cloud‑based services.

Here are a few quick takeaways:

  • Both courts agreed that AI tools are not attorneys;
  • Both cases involved consumer‑grade AI tools intended for public use, where user inputs and outputs may be used for model training and disclosed pursuant to valid legal process. It remains unresolved how courts will evaluate confidentiality when counsel directs a client to use an enterprise‑grade AI system6  with robust data‑protection safeguards;
  • As always, review the terms of service of generative AI tools to evaluate their policies on the use and disclosure of user data; and
  • Set clear policies with client agencies on using generative AI tools (consumer grade or enterprise) in anticipation of litigation

Further Reading

The following articles provide excellent breakdowns of the issue and cases, including ways in which the judges may have erred in their reasoning. My thanks to Professor Rebecca Fordon of Ohio State University Moritz College of Law for bringing these articles to my attention:

Issue 2: Public Records Requests Target Government AI Use

In 2025, journalist Nate Sanford wondered whether, and in what ways, local government officials in cities across Washington state were using generative AI. He decided to file Freedom of Information Act (FOIA) requests to find out. To his surprise, not only did several cities comply with his request, but the responses showed that government use was more widespread than he had anticipated. The disclosures have led to further reporting, including how AI was used by an employee to award a vendor a city contract. This is an opportunity to update administrative processes around public records requests.

Here are some quick takeaways:

  • AI-generated content, whether through public or enterprise models, may be a public record in your jurisdiction;
  • State agencies, including AGOs, must prepare for AI-related FOIA requests by establishing clear policies regarding disclosure;
  • Records officers will need training on AI-specific requests; and
  • Use-policies must clearly state what is acceptable use, and that AI user logs and AI-generated content may be subject to public disclosure.

Articles

Issue 3: Hallucinations in Legal Research

Every week brings new instances of legal filings or court decisions containing fake cases generated by AI. As of publication, attorney Damien Charlotin’s AI Hallucinations Case Database has tracked over 1,000 cases globally since 2023 in which “a court or tribunal has explicitly found (or implied) that a party relied on hallucinated content or material generated by generative AI.” Professor Rebecca Fordon of the Ohio State University Moritz College of Law decided to dig into the current research on hallucinations in legal tasks to discern emerging patterns. Her findings thus far (which are being published as part of a three-part series) are important food for thought for attorney general offices.

Here are six takeaways she recently published on LinkedIn:

  1. Models and data access matter;
  2. AI agrees with you even when you’re wrong (sycophancy);
  3. Performance drops sharply outside major jurisdictions;
  4. Knowledge cutoffs create invisible blind spots7;
  5. Complex tasks fail at much higher rates than simple ones; and
  6. AI sounds equally confident whether it’s right or wrong.

Article

Rebecca Fordon, Part 1: What the Science Says About Hallucinations in Legal Research, AI Law Librarians Blog (Feb. 19, 2026): https://www.ailawlibrarians.com/2026/02/19/what-the-science-says-about-hallucinations-in-legal-research/

Endnotes

  1. The written opinion that followed Judge Rakoff’s ruling from the bench is available at courtlistener.com. [↩]
  2. Anthropic Updates to Consumer Terms and Privacy Policy, (updated August 28, 2025) (last visited March 11, 2026).  [↩]
  3. Anthropic’s Privacy Policy (Effective, January 10, 2026) (last visited March 11, 2026). [↩]
  4. United States v. Kovel, 296 F.2d 918 (2d Cir. 1961). [↩]
  5. The decision is available at law.justia.com. [↩]
  6. Enterprise‑grade AI systems are designed for organizational use and incorporate enhanced security, privacy, and reliability controls. Unlike many consumer generative AI tools, which may use user data for model training by default unless individuals opt out, enterprise models typically do not use customer data for training unless explicitly authorized by the organization. For additional context, see IBM’s “What is Enterprise AI?” [↩]
  7. Fordon notes that “AI tools trained on historical data will apply outdated law unless they actively search for current information,” thus creating knowledge cutoffs. Consequently, users should “ask vendors whether their tool uses web search or real-time database access” to stay abreast of legal developments. In instances where this does occur, users should still exercise caution and prompt tools to verify the accuracy of their responses. [↩]

Author

Related Posts