Photo: "Closed Book With Clasps On The Lap Of Robert Ingersoll Aitken's 'Past,' On The North Side Of The National Archives (Washington, DC)" by takomabibelot, CC0 via Openverse

Confidential, Except It Wasn't: A Federal Court Just Ruled Typing Into a Public AI Tool Can Waive Attorney-Client Privilege

A federal court has just ruled on a question every company's AI policy needs to answer: type something privileged into a public AI tool, and the privilege may be gone.

That's the holding in *United States v. Heppner*, according to Katarina (Kate) Polozie, a partner at Woods Oviatt Gilman, who said the court found a user waived attorney-client privilege by entering privileged information into a public generative AI platform — because sharing it with the platform counted as disclosure to a third party. "The provider's privacy policy (which allowed it to collect inputs, train on them, and disclose them) meant the user had no reasonable expectation of confidentiality," Polozie said. She was careful to note the limits of that logic: "Of course, while every instance is fact specific, the same logic threatens any confidential, trade secret or privileged material an employee pastes into a public tool." That's not a hypothetical about legal memos alone — Polozie said feeding customer lists, proprietary code, pricing, or strategy into a public tool can undermine trade-secret protection outright, since that protection depends on a business taking reasonable steps to keep the information secret. Once it's in the tool, she said, "there is often no practical way to claw it back."

The ruling lands on top of a broader governance gap: nearly all companies are investing in AI, but only 1% consider themselves mature in how they've implemented and governed it, according to a 2025 McKinsey report. Read together, those two facts suggest the exposure Heppner creates isn't a narrow edge case for a small number of unusually careless employees — it's a mismatch between how fast AI tools are being adopted and how few organizations have the policies in place to know whether a given prompt is safe to send. Troy Lieberman, counsel at Nixon Peabody, framed the underlying problem as one of pace: "Perhaps the biggest challenge for companies adopting AI is implementing guardrails at the same pace businesses are scaling their AI tools." He said organizations need clear policies on which AI platforms employees may use, what information can be entered into them, and when human review is required — without that, he said, "employees may inadvertently expose confidential information, rely on inaccurate AI-generated content or use AI in ways that create legal or regulatory issues," including unknowingly entering trade secrets.

Governance doesn't stop being the company's problem once the AI tool is running, either. Lieberman said using AI generally does not shift legal responsibility to the technology provider — if an AI tool improperly screens out a job applicant, denies a loan for discriminatory reasons, or mishandles personal information, blaming the technology isn't a legal defense. He singled out employment decisions, lending, and patient care as areas drawing particular regulatory attention over fairness, bias, and transparency. Tim Plunkett, senior counsel at Harris Beach Murtha and a member of the firm's Artificial Intelligence Industry Team, put the underlying tension bluntly: "AI can do so much so fast, but it can also do a lot of really bad things really fast, too. It's a force multiplier for your workforce." His point wasn't about the tools themselves — "no matter how fast the market's making tools, you've got to have really good people to use them," he said, pointing to employee training and AI literacy as the actual bottleneck.

None of the attorneys framed this as a problem solved by one policy document. Plunkett recommended organizations first inventory the AI tools already in use before writing governance rules around them, and revisit those policies as the technology and regulations change. Polozie recommended due diligence before adoption — "read the terms and do your due diligence," including negotiating contractual protections and confirming vendors maintain proper confidentiality, security, and data-retention practices. Lieberman similarly described governance as an ongoing process, not "a one-time project." Plunkett's closing framing is the one worth sitting with: "The companies that I think that'll succeed with AI won't be the ones that adopt it the fastest. It'll be the ones that govern it the best."