The order that could not stand

The Supreme Court of Georgia decided Hannah Payne v. The State on May 5, 2026. It did not decide whether Payne should receive a new trial. It could not get that far because the legal record in front of it had been contaminated by AI-generated citation failures.

The court said the assistant district attorney assigned to the case, Deborah Leslie, filed a brief opposing Payne's motion for a new trial that contained nonexistent cases and cases that did not support the propositions asserted. The trial court then denied the motion in an order largely prepared by Leslie, and that order carried the same problem into the judge's ruling. The State's appellate briefing added more citations that failed when checked.

That sequence is worse than a bad brief. A party's AI-generated legal errors became part of a court order in a murder case. The Supreme Court vacated the order and remanded the case so the trial court could issue a new one without fake or misattributed citations. It also instructed that counsel for neither party may prepare the replacement order.

There is the graveyard entry in one sentence: an AI-assisted legal drafting workflow did not merely embarrass an attorney; it poisoned the document a judge signed and forced a redo in a high-stakes criminal appeal.

What the court found

The court's opinion lays out the timeline with the grim patience of judges who have been made to debug someone else's homework. After oral argument on March 18, 2026, the justices entered an order asking the State to explain how the September 12, 2025 trial-court order came to include nine citations that either did not exist or did not support the stated propositions.

On March 27, Leslie filed a supplemental brief and affidavit. She acknowledged using artificial intelligence software to draft the State's briefs opposing the new-trial motion and to draft the proposed order denying that motion. She also acknowledged that AI-generated citations were not independently verified before they appeared in the State's briefs or proposed order.

The damage count grew from there. In addition to the nine problem citations the Supreme Court had flagged, Leslie identified twelve more cases in trial-court briefing that were AI-generated, not independently verified, and did not support the propositions offered. She also withdrew reliance on nine authorities cited in the State's appellate brief because they were unsupported, nonexistent, or inaccurately quoted.

The court's language was controlled, but the point was blunt. Georgia has no blanket rule against responsible AI use by attorneys. Filing citations that do not exist, or citations that exist but do not support the claim, violates the court's rules and falls far below what Georgia lawyers owe the court.

Sanctions and a redo

The sanctions were case-specific but serious. The Supreme Court admonished Leslie and the Clayton County District Attorney's Office for failing to verify citations. It suspended Leslie's privilege to practice before the Supreme Court of Georgia for six months. Before reinstatement, she must complete twelve additional hours of continuing legal education beyond normal bar requirements, covering ethics, brief writing, and proper AI use in the legal system.

The court also vacated the order denying Payne's motion for a new trial. The trial court must prepare and issue a new order, and that order must not contain fictitious or misattributed case citations. The court strongly encouraged trial judges to review proposed orders carefully because AI may have been used to prepare them.

That last warning is doing real work. Proposed orders are common in litigation. A judge may ask a party to draft an order reflecting the court's ruling, especially in busy trial courts. If the party uses AI badly, and the court signs the draft without catching citation failures, the hallucination gains the authority of a judicial order. The mistake stops being "the lawyer's bad brief" and becomes part of the official case record.

This is why the Georgia case is more dangerous than the routine fake-citation sanction. The error crossed a boundary between advocacy and adjudication.

Why prosecutors make this worse

Most AI-citation sanctions involve civil lawyers trying to save time on motions. That is bad enough. A prosecutor's filing in a criminal case carries a different kind of weight. The State is not just another private party. It represents public power. When a prosecutor submits AI-fabricated authority, the harm is not limited to embarrassment or wasted legal fees; it can affect liberty, finality, and public confidence in the fairness of criminal proceedings.

The Supreme Court did not reverse Payne's conviction. It did not decide the merits of her new-trial request. The direct procedural harm was delay and a required new order. But delay in a criminal appeal is not trivial, and neither is forcing an appellate court to pause merits review because the record contains fake law.

Atlanta News First reported the practical outcome clearly: the case went back to a Clayton County judge not because the Supreme Court resolved the legal arguments, but because the order needed to be rewritten without bogus AI citations. That is the kind of procedural detour that courts are built to avoid.

The dissenting portion of the opinion objected to admonishing the elected District Attorney, noting that the office had apologized and reported internal discipline and new policies. The majority admonished the office rather than the DA personally. That distinction matters, but it does not save the workflow. The office's name was on the filings. The office's proposed order reached the trial court. The office bears institutional responsibility for making sure its attorneys verify the law before asking a judge to rely on it.

The failure mode

AI legal research failures are often described as hallucinations, which can make them sound like mysterious model behavior. The operational failure is simpler: nobody opened the cases and checked them.

A citation is not decorative trim. It is the load-bearing bolt in a legal argument. It tells the court where the law came from and lets the other side test the claim. If the citation does not exist, points to the wrong case, or quotes language the opinion never said, the argument cannot be trusted. That is not a formatting error. That is a candor problem.

The Georgia opinion also shows why "we have now implemented safeguards" is not a substitute for having had safeguards before filing. Leslie represented that she had added citation-verification steps. Fine. But the defective citations were already in the briefs and the trial-court order. A safety policy adopted after the appellate court discovers the issue is an incident report, not a control.

The practical rule for legal AI should be boring enough to print on office mugs: no case gets cited until a human has opened the opinion, verified the reporter citation, read the relevant passage, and confirmed the citation supports the proposition. If a lawyer cannot do that, the citation does not go in the document.

AI can help draft, organize, summarize, and brainstorm. It cannot be trusted to invent the legal authorities that decide a criminal appeal. Georgia's Supreme Court has now added its own reminder to the pile: if the brief contains fake law, the court may have to stop deciding the case and start deciding what to do about the lawyer.