A court AI policy should not answer every use with one word: yes or no. It should route different work into different lanes.
One lane covers ordinary, non-adjudicative court operations. The other covers work connected to a pending case and the substantial rights of the people in it. The distinction is not a technical nicety. It tells a court when ordinary guardrails may be enough and when the work needs tighter controls, independent judgment, and a clear record of responsibility.
Iowa's Supreme Court adopted Chapter 27 of its court rules on July 21, 2026, with an effective date of January 1, 2027. The rules define adjudicative duties as judicial duties in a pending legal proceeding that affect substantial rights. They include presiding over hearings and trials, deciding motions, cases, and appeals, and issuing rulings, orders, decisions, and opinions.
For that adjudicative lane, Iowa does not permit a judicial officer to delegate or defer independent decision-making responsibility to AI output. It requires independent verification of factual and legal sources in AI-prepared work product. And it requires an approved tool meeting confidentiality, security, and reliability requirements before judicial case information is used with AI for adjudicative work.
California's Rule 10.430 makes a related distinction from the operations side. A California court that allows generative AI must have a policy covering court staff for any purpose and judicial officers for tasks outside their adjudicative role. The policy must prohibit confidential, personally identifying, and other nonpublic information from being entered into a public generative-AI system. It must also require reasonable steps to verify AI material and correct erroneous or hallucinated material that is used.
Neither source is a Tennessee court rule or a Tennessee legal-ethics opinion. But together they offer a usable policy design: do not treat every AI task as though it carries the same risk.
Four routing questions before AI touches court work
- Is the task tied to a pending proceeding? If it affects a party's substantial rights, it belongs in the heightened lane. The question is not whether the output looks routine. The question is what legal work it helps perform.
- Will the task use case information or other nonpublic data? A tool that may be acceptable for public information cannot be assumed safe for case-specific material. Data boundaries should be decided before information is entered, not after it leaves the court's control.
- Who keeps the independent judgment? The person with adjudicative responsibility must remain able to evaluate the work without deferring to the tool. A clear answer from an AI system is not a substitute for a judicial decision.
- What is the verification and correction path? A policy should say who checks factual and legal sources, who corrects an error, and how the court responds when a tool produces material that should not be used.
This two-lane approach does not prevent innovation. It makes routine uses easier to manage because the policy gives them a clear home. It also keeps case-connected work from drifting into the same low-control workflow simply because the software looks familiar.
A good policy need not predict every future tool. It needs to make one present judgment clearly: ordinary operations and adjudication are not the same kind of work, and they should not carry the same AI permissions.
Question for local discussion
If a proposed AI use moved from public-facing or administrative work into a pending case, would your policy tell the court when the rules must change?