Courts Are Writing the AI Rules. They Won't Let AI in the Room.

On June 15, 2026, Chief Judge Dean Bucci stood before an assembled jury pool in the Paulding County Courthouse and instructed them to investigate nothing and discuss the case with no one. That instruction is among the oldest safeguards in the American trial. The case waiting on that pool was a murder prosecution, six and a half years old and finally reaching a jury.

The next day, Chief Deputy Clerk Lori Jennings informed the court that Juror #18 had violated it. Bucci called him back into the courtroom and questioned him. The State and the defense then moved jointly for a mistrial. Bucci granted it and sent the panel home.

The court's show cause order describes the conduct in neutral terms: performing research on the case and sharing it with other potential jurors. Court TV, which had a camera in Courtroom 1 under a Rule 22 order, reported that the research was a ChatGPT session.

No recording of the exchange exists. Bucci's Rule 22 order, entered six days earlier, forbade the cameras from capturing images of jurors or prospective jurors in any manner, and prohibited audio of their statements or conversations outright. The permitted exceptions ran to two: a foreperson announcing a verdict, and jury questions to the judge. The one proceeding that became national news was the precise thing the order sealed.

The case underneath

Lacy Jo Lee Boles stands charged with malice murder, felony murder and aggravated assault in the 2019 shooting death of her husband, Daniel Akers. She was indicted in December 2020.

The intervening years went to litigation. A pandemic delay came first, then a fight over access to the victim's cellphone. A hearsay ruling carved up a recovered audio recording. A firearm count was dismissed on the State's motion. A contested subpoena to the GBI medical examiner was resolved by in camera inspection weeks before trial.

Six years of that work, and jury selection was the last step. Trial is now specially set for November 9, 2026, with two weeks blocked out.

One procedural note for anyone reaching for the term. The panel had never been sworn, so jeopardy never attached. The word in the order is mistrial, and the consequence here is a calendar entry rather than a constitutional bar.

The recusal

On the day he granted the mistrial, Bucci did something else worth pausing on.

He recused himself from the contempt proceeding while remaining on the criminal case, and directed the clerk to randomly assign another superior court judge to hear it. Juror #18 was ordered to show cause before a different judge, in a different courtroom, on August 11.

The instruction that was broken was Bucci's own. The examination of the man accused of breaking it was Bucci's own. Having supplied both halves of the record, he would not be the one to weigh it.

A judge who has become part of the record cannot also weigh it. The bench enforces that on itself, and here it cost Bucci a proceeding he was entitled to keep. Ask a model the same question and it answers at once. Nothing about it can be recused, because there is no one in the proceeding to remove.

The referral worked the way referrals are supposed to. Judge T. David Lyles heard the matter. The juror appeared with counsel, admitted the conduct, and offered mitigation. On August 31, Lyles ordered him to pay $4,762.43 into the registry of the court, reimbursing the county for the out-of-pocket costs of the jurors who had been assembled and sent home. He has forty-five days. Nonpayment carries an order of incarceration.

Somebody in the clerk's office added up the mileage and the per diem of a discharged jury pool and arrived at a figure accurate to the penny. That is what the exercise of judgment looks like when it is working: a specific person, a specific sum, a deadline, and a consequence for missing it.

The rule already covers it

Jurors are instructed to decide the case on the evidence admitted at trial and the law the judge gives them. That instruction is old, and it is broad enough to reach generative AI without amendment. Federal model instructions prohibit using a phone, tablet or computer to research the case, look up a witness, or communicate about the proceeding. A juror who types the facts into a chatbot has done outside research. The rule reaches the conduct whether or not it names the tool.

The drafting still lags. The model instruction on electronic technology names its examples by brand, running through Twitter, Facebook, Instagram, LinkedIn, YouTube, WhatsApp and Snapchat. It names no AI product. That silence is where defense counsel will go on appeal, and it is where trial counsel should go first, by asking for an instruction that names the tools out loud.

Why the prohibition is substantive

A juror who Googles a term lands on a page that announces what it is. Wikipedia looks like Wikipedia. A message board looks like a message board. The juror keeps some sense that the source is unvetted.

A juror who asks a chatbot for the standard of care receives something else entirely. The answer arrives structured, confident, and dressed in citations. It does not know which state's law governs. It does not know whether the case is civil or criminal. It may cite authority that was invented three seconds earlier. None of that is visible to a layperson reading it at the kitchen table.

The result is a juror who believes he has been educated when he has been given unauthorized instructions on the law from a source no judge approved and no lawyer can cross-examine. The model heard no witnesses. It observed no demeanor. It received the juror's own compressed retelling of the testimony rather than the testimony itself.

Courts have declared mistrials over far less. A single juror looking up one definition has been enough.

The expert's own file

The apparatus around an expert witness is the most elaborate verification machinery the law has built. An expert is sourced, disclosed, deposed and cross-examined, and as Josh Hyatt of Expert Institute put it in an interview on this site, the prior transcript sinks more experts than the thin CV does, because every answer he has ever given sits in the record waiting to be used against him. The whole design assumes that what reaches a jury from the witness stand has survived an adversary.

Four cases from the last two years show that machinery meeting generative AI at the one point where nobody was looking.

In Kohls v. Ellison, Minnesota defended its deepfake statute with a declaration from a Stanford professor who studies misinformation. The declaration cited two academic articles that do not exist and misattributed a third, the product of drafting with a large language model. Judge Laura Provinzino excluded the declaration in its entirety and refused to accept an amended version. Rule 11, she wrote, imposes a “personal, nondelegable responsibility” to “validate the truth and legal reasonableness of the papers filed.” The citations to fabricated sources, she added, shatter the expert’s credibility with the court.

In Concord Music Group v. Anthropic, a defense expert’s declaration cited an article carrying a fabricated title and authors who had never worked together. Counsel acknowledged that Claude had produced the citation while formatting references, and that a human had afterward located the real underlying article. Magistrate Judge Susan van Keulen struck the paragraph. “That is a plain and simple AI hallucination.”

In Matter of Weber, an expert used Microsoft Copilot to cross-check a damages calculation and could recall neither his prompts nor how the tool reached its outputs. Surrogate Jonathan Schopf rejected the calculation on grounds that had nothing to do with AI, then addressed the tool separately. Counsel carries “an affirmative duty to disclose the use of artificial intelligence,” he held, and such evidence “should properly be subject to a Frye hearing prior to its admission.” The court had run the figure through Copilot itself and received three different answers.

The fourth is the one where somebody went looking.

The January 2020 explosion at Watson Grinding and Manufacturing in Houston killed three people and damaged roughly two hundred homes. In the litigation that followed, 3M retained Josh Autenrieth of KnightHawk Engineering, an engineer who had never before served as an expert witness, to opine on its standard of care. Plaintiffs' counsel Will Moye noticed a document in the production that he recognized as chatbot output, and demanded the prompts behind it. He received hundreds of pages of ChatGPT logs. One prompt asked the model to show how 3M was zero percent at fault.

Put to him at trial that eighty-five to ninety percent of the thirty-page report had come from the model, Autenrieth declined to give a number and answered, “If you say so.” Asked why an impartial witness would instruct a chatbot to demonstrate zero fault, he said the prompts “were biased because I had opinions,” and that his opinions were “already set.”

One detail carries the section. ChatGPT’s draft asserted that 3M bore zero percent of the responsibility, and Autenrieth cut that sentence. Striking a line like that is what an expert is supposed to do with a tool’s output, until you reach his reason for striking it.

He had asked the model to argue against his own report, and it came back with two objections. A flat claim of zero percent would be easy for opposing counsel to attack. Apportioning fault between defendants is the jury’s job rather than the engineer’s, so a percentage had no place in an engineering opinion at all.

Both objections were sound. Neither was his. The one act of professional judgment visible anywhere in that report was performed by the chatbot, on request.

3M moved to disqualify its own expert. On August 10, 2026 the jury returned $61,578,273.19 and put thirty percent of the responsibility on 3M.

There is a further wrinkle in who it worked on. The check Hyatt describes runs on history, and Autenrieth had none. No prior testimony, no transcript to set beside this one, nothing for opposing counsel to pull and compare. The mechanism that catches experts was structurally blind to this one, because it had nothing to read.

What he did have was a record he created while writing the report. Hundreds of pages of it, produced because somebody asked.

No new rule produced that result. Opposing counsel read the production, recognized what he was looking at, and asked for the prompts.

Asking is one thing. In May a federal magistrate held that counsel can compel it.

In Conservation Law Foundation v. Shell Oil, Magistrate Judge Thomas Farrish ordered production of the generative AI prompts an expert and her research assistant had used in preparing her report. The rationale ran wider than Rule 26(a)(2)(B) and its facts-or-data language. An expert’s methodology is fair ground for discovery, the court held, and the process of culling a document production down to a working subset is part of that methodology.

The work-product objection failed on a detail worth copying into your next stipulation. The parties had a Rule 29 agreement protecting the expert’s notes, drafts and communications. The court found it insufficiently clear to reach AI prompts. That order is stayed pending objection to the district judge, so it is not the last word on the question. It is the first one.

A line is forming around who did the prompting. Prompts an attorney composes have been treated as work product. Prompts an expert composes are being treated as method, which puts them on the discoverable side of the same page.

Set those beside the instruction Bucci gave his jury pool. A juror who consults a chatbot carries the answer into a deliberation no one will ever see. An expert who consults one files it, under oath, through the process built to test the three things a chatbot cannot supply: a source, a method, and a prior record.

The same line runs above the bench

Lawrence Wheeler presides over a family law docket in the District Court of Stephens County, Oklahoma, from chambers in Duncan. He was elected in 2022.

In November 2025 he denied an attorney's request for a psychological evaluation of a parent in a custody matter. His order cited Oklahoma authority supporting the denial. Two of the cited cases do not exist.

Wheeler told an investigator with the Oklahoma State Bureau of Investigation that he had written the order himself and had used ChatGPT only for research, according to an August 17, 2026 letter from Stephens County District Attorney Jason Hicks to the state attorney general's office. Hicks recused himself and asked Attorney General Gentner Drummond to review the OSBI report or reassign the inquiry. Drummond's criminal division reviewed the allegations and declined to prosecute, noting that the Oklahoma Supreme Court and the Court on the Judiciary hold disciplinary power over judicial conduct whether or not that conduct is criminal. Wheeler's counsel declined to comment.

The same misconduct complaint carried a second allegation, still unresolved. It alleged that Wheeler secretly listened to jury deliberations.

Taken together, the complaint describes a judge who handed his legal research to a machine and entered the one deliberation the law seals shut. The first part he admitted. The second remains an allegation.

The pattern extends well past Oklahoma. Two federal judges have attributed errors in their rulings to AI use in chambers. A Georgia prosecutor's AI use put erroneous citations into a ruling in a murder case.

The principle judges articulate for themselves is stated plainly enough. Tasks requiring discretion, the weighing of evidence, assessments of credibility, and the application of law to fact belong to the judge and stay there. Administrative work may be delegated to a machine. The exercise of judgment may not.

That line is one each of them drew for himself. Wheeler wrote the order and used the tool only for research. Anthropic’s counsel said Claude only formatted a citation to an article that turned out to be real. Ranson said Copilot only checked a figure he had already calculated. Autenrieth said the conclusions rested on twenty years of his own experience. Each man kept the judgment and gave away a step he had decided was clerical, and in each case the man drawing the line was the one it protected.

Three courts rejected the division outright, and in the fourth matter 3M moved to disqualify the expert who drew it. What they enforced instead is that a signature covers the whole document, including the part the signer decided was too small to check.

Set Wheeler beside Bucci. One judge is alleged to have collapsed the distance between himself and a proceeding he had no business entering. The other, having issued the order and questioned the man accused of defying it, stepped back from a proceeding he had every authority to decide. The same principle produced both the complaint and the recusal.

The asymmetry nobody is naming

Here is what makes this more than a compliance story.

American courts have become the working regulator of artificial intelligence. Congress has produced no comprehensive statute. Agencies have issued guidance. Meanwhile the wrongful death suits, the product liability claims, the chatbot harm cases and the copyright actions keep landing on dockets, and judges and juries keep having to resolve them. Every ruling on a duty to warn, every verdict on a design defect, every finding on foreseeability is a piece of AI regulation written in the only forum currently able to write any.

The institution doing that work has simultaneously barred the technology from its own deliberative process. No jury answers the broad question in those terms. A jury in a wrongful death case against a chatbot maker answers a narrow one: did this company act reasonably toward this plaintiff. Accumulate enough of those answers and the broad standard exists, assembled out of verdicts nobody drafted as policy.

Which puts twelve citizens in a particular position. They will decide whether an AI company behaved reasonably, under an instruction forbidding them to ask the defendant's own product what it thinks.

Underneath both halves sits a judgment about what kind of decision a verdict is. A verdict is an act of accountability by people who can be held to it. The jury signs its name. A model has no prior testimony to impeach it with. No credibility a court can find shattered. No one to bill when a jury pool is sent home.

What this does to the doom narrative

The public conversation about artificial intelligence has settled into a familiar register. Bot swarms overwhelm the open web. Synthetic content displaces human work. The machines get loose and nobody is in a position to stop them. The premise underneath all of it is that the technology is arriving faster than any human institution can respond.

The premise is wrong, and the evidence is sitting in Paulding County.

A deputy clerk noticed and told the judge. The judge brought the juror back in and questioned him. A prosecutor and a defense lawyer who had spent six years disagreeing about everything moved jointly, on the spot, for the same remedy. The judge granted it that afternoon, then stepped aside from judging a contempt of his own instruction.

No new statute was required. No agency rulemaking was consulted. Procedure that predates the republic ran through a county courthouse in a single afternoon and produced an enforceable result, a contempt docket, and a hearing date.

Courts are slow in the aggregate and fast in the particular. They regulate one case at a time. They assign responsibility and enter judgment. Repeat that several thousand times and a body of law exists.

The question of who governs artificial intelligence is being answered right now. It is being answered by verdicts.

For trial counsel

Five practical consequences follow, and the last one runs the other way.

Ask about AI in voir dire. Prospective jurors who use these tools daily may sincerely believe consulting one differs in kind from research. Ask directly whether they use generative AI, and ask whether they understand that it produces false information with total confidence. Georgia practitioners should note that voir dire scope is bounded by O.C.G.A. § 15-12-133 and Uniform Superior Court Rule 10.1, and that a question about a juror's habits and credulity fits comfortably inside those bounds.

Request an instruction that names the tools. Generic prohibitions on internet research leave an argument available to the other side. Naming ChatGPT, Claude, Gemini and Copilot closes it.

Move quickly on suspected juror misconduct. Investigation is time sensitive, and juror secrecy rules vary sharply by jurisdiction. Note also that the Paulding court sealed its own record of the exchange as to recording, which leaves the only account of what was said with the court reporter and the participants.

Examine experts on their own AI use, and ask for the prompts. Weber supplies the demand: counsel carries an affirmative duty to disclose AI use, and AI-generated evidence belongs in a Frye hearing before it reaches a jury. An expert who cannot produce his prompts or explain how a tool reached its output has handed you the cross-examination. Kohls shows what the record looks like when nobody asks, and the Watson Grinding production shows what turns up when somebody does. Conservation Law Foundation supplies the authority for compelling it, and the warning that goes with it: check that your own protective stipulation names AI prompts, because one drafted to cover notes, drafts and communications did not.

Then turn the same questions on your own expert. Everything above works against you with the parties reversed. Your expert’s prompts are discoverable on the identical methodology theory, and the stipulation meant to protect them is the one Conservation Law Foundation found wanting. Ask at the engagement stage rather than at the deposition, and ask the expert to state in his own words which conclusions are his. Keep that answer. Where an expert has never testified, it is the only record of his method that exists, and the 3M production is what the other side does with the gap.

A live test

Three dates now sit in sequence on the Paulding County calendar.

• Mid-October A former prospective juror's forty-five days run out. He either pays the county what the wasted pool cost or faces incarceration.
• Nov. 4 · 5:00 PM Proposed voir dire questions are due to the court in writing. Bucci has decided against using jury questionnaires this time, so whatever counsel want asked, they have to ask themselves.
• Nov. 9 · 8:30 AM A new panel files into Courtroom 1.

Two lawyers in Dallas, Georgia will sit down before that deadline and decide whether to ask a room of strangers what they do with a chatbot. Their answer, in the one case that has already lost a jury to exactly that question, is the whole argument in miniature.

SourcesOrder on show cause hearing (Lyles, J.), filed August 31, 2026; Show Cause Order, filed June 29, 2026; Order Rescheduling Specially Set Trial, filed June 24, 2026; Second Order on Request to Use a Recording Device Pursuant to Rule 22, filed June 10, 2026; and Order on Motion to Quash, filed August 4, 2026, all in State v. Lacy Jo Lee Boles, No. 20-CR-001378-P2, Superior Court of Paulding County, Georgia. Court TV, “Murder trial postponed after juror asks ChatGPT about the case,” June 17, 2026. Mike Scarcella, “Oklahoma judge used AI in ruling that contained false citations, prosecutor says,” Reuters, September 9, 2026; the Hicks letter was first reported by The Oklahoman. Kohls v. Ellison, No. 24-cv-3754 (D. Minn. Jan. 10, 2025); Concord Music Group, Inc. v. Anthropic PBC (N.D. Cal. May 23, 2025); Matter of Weber, 2024 NY Slip Op 24258 (Sur. Ct. Saratoga Cty. Oct. 10, 2024). Watson Grinding explosion litigation, Harris County, Texas, verdict of August 10, 2026, as reported by CBS News, 404 Media and Above the Law. Conservation Law Foundation, Inc. v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn. May 18, 2026) (Farrish, M.J.), order stayed pending objection. Administrative Office of the U.S. Courts, proposed model jury instructions on the use of electronic technology.