What happened
KOLD reports that the Arizona Supreme Court rejected a petition filed on Jan. 12, 2026, that sought to create Rule 135 and impose a moratorium on generative AI for core judicial work through 2029. Instead, the court will continue a closely supervised testing phase.
KOLD reports that the Arizona Supreme Court rejected a petition seeking to ban judges from using artificial intelligence in their core work. The petition, filed Jan. 12, 2026, proposed “Rule 135,” which would have imposed a strict moratorium on generative AI use for core judicial work through 2029. The court instead chose to continue what KOLD describes as a highly regulated testing phase. The source does not include the court’s written order, the full text of the proposed rule or a detailed explanation of the justices’ reasoning, so those elements are not independently confirmed here. The reported choice leaves supervised testing as the operative approach described by KOLD, while the proposed moratorium remains rejected.
According to KOLD, the court’s current guidelines allow judges to use approved AI tools for limited purposes, including reducing case backlogs, improving administrative efficiency and expanding public access to justice. The report describes the policy as a test-and-monitor approach rather than unrestricted adoption. KOLD also reports a firm boundary: judges may not use AI to write actual legal rulings or make final decisions in cases. The source does not specify which systems qualify as approved tools, what data protections apply, or what documentation judges must create when they use AI. Those omissions leave the practical limits of approved use unresolved.
KOLD interviewed Jaime Ibrahim, deputy director for Southern Arizona Legal Aid, and Keith Swisher, a University of Arizona law professor, about the decision. Ibrahim characterized AI as another tool in a legal environment that has long used case-law search and law-clerk assistance, while Swisher said completely blocking access could leave courts unable to respond to emerging challenges. Swisher specifically pointed to identifying deepfake evidence and navigating cases in which self-represented people rely on AI-generated legal advice. These are expert perspectives reported by KOLD, not findings by the court or independently verified conclusions. Their comments help explain why the debate includes both possible benefits and possible risks.
Why it matters
The decision establishes a state-level approach to courtroom AI that permits limited operational use while reserving legal reasoning, rulings and final decisions for judges. It also places practical importance on oversight, approved tools and future evaluation.
The decision matters because it separates administrative assistance from the exercise of judicial authority. Under the framework described by KOLD, AI may help with certain court operations, but responsibility for rulings and final case outcomes remains with human judges. That distinction is practically important: an efficiency tool can influence what information is surfaced or how quickly work is processed even when it cannot formally decide a case. The source does not establish how much AI is currently being used or whether the policy has already changed processing times. The reported framework therefore makes human responsibility the stated dividing line while leaving actual effects on court work unmeasured.
KOLD’s reporting also highlights a growing governance problem for courts: they may need AI-related capabilities to detect synthetic or manipulated evidence and to understand filings prepared with AI assistance, while the same systems can introduce factual errors or unsupported claims. Swisher told KOLD that AI can exaggerate points, miss material details or make mistakes, and that trained human review remains necessary. The report provides no performance testing, error-rate data or examples from Arizona courts showing how often these problems occur. The need for review is thus clear in the reported discussion, but the scale of the risk is not established by the available account.
The policy could become a reference point for other courts weighing bans against controlled experimentation. A total prohibition would provide a clear rule but could limit access to tools that court administrators or judges consider useful. A supervised-testing model may allow learning, but its credibility depends on transparent safeguards, consistent review and meaningful consequences when systems fail. KOLD reports the policy and the experts’ views; it does not independently demonstrate that Arizona’s approach improves justice, reduces backlogs or protects litigants. Its significance will consequently depend on how the experiment is evaluated and what information becomes available about its operation.
What to watch next
The Arizona Steering Committee on Artificial Intelligence and the Courts is scheduled to submit its next major status report on March 1, 2027, according to KOLD. Key unknowns include which tools are approved, how usage is audited, and whether the policy changes after testing.
The next formal checkpoint is the Arizona Steering Committee on Artificial Intelligence and the Courts’ major status report, which KOLD says is due to the Supreme Court on March 1, 2027. That report could clarify what the testing phase has shown, whether approved tools have been expanded or restricted, and whether the court will retain its current limits. The source does not say whether interim reports, public hearings or other disclosures will occur before that date. The March 1, 2027, report is therefore the clearest scheduled opportunity identified in the coverage for assessing the policy’s direction.
The most important operational questions remain unanswered in KOLD’s report. It is not clear which AI products judges may use, whether those systems can process confidential case information, how outputs are preserved for review, or who is responsible for checking citations and factual assertions. It is also unknown whether the court has established testing standards, incident reporting requirements or public metrics for backlog reduction and access to justice. Until those points are explained, the phrase supervised testing describes an intended control structure more clearly than it describes its day-to-day operation.
Future scrutiny should focus on the boundary between assistance and influence. Even if AI cannot draft a final ruling or make a formal decision, its summaries, searches or suggested language could shape what a judge notices and how a case is understood. KOLD reports that human oversight is required, but does not describe an enforcement mechanism or independent audit. Those details will determine whether the court’s experiment is a controlled evaluation or simply a permission structure with limited visibility. They will also show how the stated limits operate when judges use approved tools in actual judicial work.
