Tennessee Edition
Welcome to Issue 008 of The Bench & Bar AI Brief, Tennessee Edition.
Planned publication date: July 14, 2026.
The legal profession has spent much of the past year warning that AI can invent cases. That warning remains true, but this week's verified court record shows a broader problem. AI can also expose confidential material, defeat a privilege claim, and create a false sense that legal judgment has been delegated along with the drafting.
The clearest statement comes from the Eleventh Circuit's July 10 published opinion in Akerlund v. Atlas Air. The court said an opening brief relied on eight nonexistent cases. When the problem was identified, counsel sought to withdraw eight different authorities. The court said those authorities did not match the original eight and were also hallucinated.
The court's lesson was not limited to citation checking. It said completely outsourcing legal work to artificial intelligence is not competent and described that conduct as a dramatic violation of the client's interests. It also stated the operational rule plainly: when lawyers use AI, they still must verify that the software got it right and keep what it got wrong out of court filings. The opinion says the matter will be referred through the Chief Judge to the court's Committee on Lawyer Qualifications and Conduct. That is an announced referral, not a completed disciplinary result.
A separate federal decision shows why tool use must begin before the prompt is typed. In United States v. Heppner, the Southern District of New York held that a defendant's exchanges with a publicly available AI platform were not protected from government inspection by attorney-client privilege or the work-product doctrine in the facts before the court. Counsel had not directed the defendant to run the searches, and later sharing the material with counsel did not convert previously nonprivileged communications into privileged ones.
That decision should not be stretched into a claim that every AI-assisted document is unprotected. Its value is narrower and more useful. Lawyers must decide who is directing the work, what information may be entered, what relationship the tool has to counsel, and whether the workflow satisfies ordinary privilege and work-product rules. Labels such as "research," "draft," or "for my lawyer" do not create protection by themselves.
Two discovery orders add the confidentiality layer. A Northern District of West Virginia standing order for criminal cases warns that data-retentive or consumer-tier AI tools can threaten the integrity and security of discovery. It requires defense counsel to give notice to and obtain written government consent before entering sensitive material into an AI tool. The request must identify the tool and address whether it retains, trains on, or exposes the material.
A case-specific stipulated protective order from the District of Nevada draws a related line in civil discovery. It permits AI assistance for tasks such as drafting, summarizing, organizing, and document review, but it conditions use of protected material. A closed system may be used only when the user has a good-faith basis to believe the system will not use the material to train publicly accessible models and will restrict access consistently with the order. Open-system use may not disclose protected material or violate restrictions involving sealed documents and personal identifiers.
None of these decisions or orders is Tennessee law. The sources come from the Eleventh Circuit, the Southern District of New York, the Northern District of West Virginia, and a District of Nevada case. But the duties they identify are familiar everywhere: competence, confidentiality, privilege analysis, candor, and accountable judgment.
For Tennessee courts, lawyers, agencies, and vendors, the safest workflow begins with the duty, not the product.