On 20 April 2026, a federal judge in Alabama ordered a lawyer to hand over screenshots of his ChatGPT history. The judge wanted to know whether four quotations in the lawyer’s brief, attributed to real court decisions but found nowhere in them, had come from the chatbot. Three days later, the lawyer cancelled his ChatGPT account. He also asked OpenAI for a prorated refund for the rest of the month, which came to five dollars and twenty-seven cents and cut off his access a day before the court’s deadline (Miller v. Regions Bank, 2026, p. 7).
The court’s response, in a 34-page order on 21 May, was severe: a public reprimand, a six-month suspension from practising in that court, removal from the case and a referral to the state bar (Miller v. Regions Bank, 2026, pp. 33–34). The judge was careful to say why. The lawyer was not being punished “merely for using generative AI” (p. 2). He was being punished for what he did to the record.
In New York, a much bigger fight about missing chat records is waiting for a decision. The New York Times and other publishers suing OpenAI for copyright infringement have asked the court to sanction the company, arguing that it let ChatGPT conversations be deleted despite a court order to preserve them. OpenAI says it complied in good faith. As of mid-October, we could find no ruling.
Put together, the two cases teach something every organization using AI needs to understand: the conversations people have with AI systems are now evidence. Whether they are kept, for how long, and who can delete them is not a technical default. It is a decision, and someone will eventually be asked to explain it.
Case one: the account that was deleted after the order
The facts in Miller v. Regions Bank come from the court’s own order. On 30 November 2025, the plaintiff’s lawyer filed a brief opposing summary judgment in an employment case. It contained four quotations that the court could not find in the cases they were attributed to, or anywhere else (pp. 2–3). At a hearing in April, the lawyer said he had used ChatGPT “to do a general search” but that it had not written his brief. He admitted he had not checked his ChatGPT history, and offered to “pull a log” (pp. 4–5, 18). The court ordered him to do exactly that.
What came back nine days later was a one-page unsigned note saying his account “has been deleted” and could not be recovered (p. 5). The court then ordered him to produce his emails from OpenAI. Those showed that he had bought a paid ChatGPT subscription four days before filing the brief, and that he cancelled the account on 23 April, three days after the production order, requesting the refund that ended his access immediately (pp. 6–7).
From that sequence the court drew two conclusions. First, once he was ordered to produce his chat history, he had “a clear and specific duty to preserve that evidence” (p. 6). Second, by deleting it he acted in bad faith, which allowed the court to draw an adverse inference: to conclude that the missing history would have shown ChatGPT generated the false quotations (pp. 13–14, 22–23). The court also found him in civil contempt (p. 27). Its summary was blunt: when confronted with a mistake, he “chose dishonesty over candor and destruction over disclosure” (p. 32).
Key idea: The duty to preserve, and spoliation
Once a dispute is under way or reasonably expected, the people involved must keep the records that bear on it, even records they would normally delete. Destroying or losing that material is called spoliation. Courts can respond by assuming the missing evidence would have hurt the side that lost it. For ordinary people this matters because chat logs, emails and system records are often the only proof of what an organization or its AI actually said or did.
The relevant federal rule for lost electronic records, Rule 37(e) of the Federal Rules of Civil Procedure, sets out the stakes. If information that should have been preserved is lost because a party did not take reasonable steps, a court can order measures to cure the harm. If the party acted “with the intent to deprive” another party of the information, the court can presume the information was unfavourable, tell a jury to do the same, or even end the case (Fed. R. Civ. P. 37(e)). The Alabama court noted it was unclear whether that rule fits a lawyer hiding evidence from the court itself, rather than from an opponent, and relied on its inherent power instead, while finding the result would be the same either way (Miller v. Regions Bank, 2026, p. 19 n.5).
Case two: the logs a company says it could not stop deleting
The New York case is different in almost every way, and that is what makes it useful. Nobody has found that OpenAI acted in bad faith. The questions are what a preservation order required, what a company with tens of billions of conversations could reasonably do, and whether the gap matters. Everything below about the dispute comes from the parties’ own court filings. These are allegations and arguments, not findings.
The order. On 13 May 2025, Magistrate Judge Ona T. Wang ordered OpenAI to “preserve and segregate all output log data that would otherwise be deleted on a going-forward basis”, as the publishers quote it (News Plaintiffs, 2026, p. 3). Output logs are the records of what ChatGPT said to users.
What OpenAI says it did. In its opposition, filed on 14 August 2026, OpenAI says that it placed in a segregated legal hold all consumer ChatGPT logs generated from 14 April 2025 until the order ended on 26 September 2025, including chats users had asked to delete (OpenAI, 2026, p. 2). It says that stopping deletions from its compressed warehouse of older chats was “technically infeasible”, and that at a hearing on 27 May 2025 the judge clarified the order did not require “wholesale” preservation, saying: “I’m not saying stop all your deletion processes” (pp. 2, 10). Users could therefore still delete chats more than 30 days old (p. 11). When the parties later drew a sample of 20 million conversations for analysis, OpenAI says it had to replace about 5 percent of them because users had deleted them in the meantime, and that a court-ordered test showed those exclusions did not bias the sample (pp. 3–4, 13).
What the publishers say. The publishers’ motion was deemed filed on 9 July 2026 (Stipulated Order, 2026). In their reply of 10 September, they argue the order was “unambiguous and unconditional”, that OpenAI failed to preserve data it was told to keep, and that it was slow to disclose internal datasets and analyses, which they say were relevant to whether ChatGPT reproduced their articles (News Plaintiffs, 2026, pp. 1–3). They ask the court to bar OpenAI from relying on the 20-million-log sample “for any purpose”, to make findings of fact against it, and to award attorneys’ fees (pp. ii, 14–17). OpenAI calls those requests punitive and says discovery revealed “no evidence whatsoever” for the publishers’ central claim (OpenAI, 2026, p. 1).
Where it stands. Our check of the public docket copy on CourtListener on 15 October 2026 found briefing complete and no ruling on the motion. The court could grant it, deny it, or something in between.
Why AI logs are evidence
It is tempting to think of chat logs as clutter. They are not. When an AI system answers a customer, drafts a document, recommends a decision or takes an action, its log is often the only record of what it was asked, what it said and what happened next. In Miller, the log was the best evidence of where false legal authority came from. In the New York case, the logs are the evidence of what the product actually produces for users. In a workplace, the same kind of record could show what an AI tool told an employee about a policy, what a chatbot promised a customer, or why an automated system flagged a person.
That is why Operational AI Readiness treats logging not as a technical nicety but as an “evidence spine”: the minimum set of records that lets an organization answer serious questions after something happens. It names system identity, version information, where the context came from, reviewer actions and what happened downstream (Bakos, 2026, Ch. 6). The book is equally clear that this is not endless retention: retention decisions should be “deliberate and documented”, not side effects of whatever a logging tool keeps by default (Bakos, 2026, Ch. 6).
Retention is a decision, not a default
The hard part is that there are good reasons to delete. AI conversations often contain deeply personal information, and privacy law and plain decency push towards keeping less. OpenAI’s filing makes this point at scale: it describes a warehouse of more than 50 billion conversations and the privacy interests of users in over 160 countries (OpenAI, 2026, pp. 1, 4). A blanket “keep everything forever” rule would create its own harm.
So the lesson is not “never delete”. It is that deletion has to be designed, and so does the ability to stop it. The New York dispute turns partly on whether a deletion pipeline could be paused when a court asked. The Alabama case turns on a person pressing delete after being told not to. In both, the question a court asks is the same: did you take reasonable steps to keep what you were required to keep? Organizations that answer that well have usually made four decisions in advance.
| Decision | What a good answer looks like |
|---|---|
| 1. What do we record? | For each AI system that influences decisions: inputs, outputs, model and prompt version, tools called, human reviewer and outcome, with timestamps. |
| 2. How long do we keep it, and why? | A written retention period per system, with a stated reason (legal, operational, privacy) and a named owner who approved it. |
| 3. Can we pause deletion? | A tested legal-hold switch that can suspend deletion for a specific person, account, system or date range within days, including in vendor tools and staff AI accounts. |
| 4. Who is told when a dispute starts? | A trigger: when a claim, complaint or investigation is expected, legal notifies system owners, and staff are told in writing not to delete relevant AI chats or accounts. |
Notice the last row. In Miller the deleted record sat in one person’s private chatbot account, bought with a personal subscription. Many organizations have no idea which AI accounts their staff use for work, so they cannot preserve what they cannot find.
What the record does not show
In the New York case there is no ruling, and nothing here should be read as a finding that OpenAI breached the preservation order or concealed evidence. We read the publishers’ reply and OpenAI’s public, partly redacted opposition; we did not read the sealed versions, the publishers’ original motion, or the 2025 preservation order itself, which we quote as the publishers quote it. We cannot say how the court will weigh OpenAI’s technical explanations or the sampling test.
In Miller, the court’s conclusion that ChatGPT generated the false quotations is an adverse inference drawn because the evidence was destroyed, not a direct finding from the chat history, which no one has seen. The lawyer did not admit using AI to create the quotations (p. 13). The order also does not tell us what happened to the client’s underlying employment claim, and we have not checked whether the lawyer sought to challenge the sanctions.
Questions you can ask
- If you are in a dispute with an organization (an employer, an insurer, a landlord, a service provider) and an AI chatbot or tool was involved, ask in writing that they preserve the relevant chat logs and system records, and keep your own screenshots with dates.
- If you use AI tools at work: ask whether your chats are covered by your employer’s retention and legal-hold rules, and never delete AI history connected to a complaint, audit or case.
- For organizations: can you name every AI system and staff AI account that might hold relevant records, and could you pause deletion in each within a week?
- For leaders signing off on AI vendors: what are the vendor’s default retention periods, and does the contract let you impose a legal hold on your data?
Go further. Operational AI Readiness is FairByDesign’s free field manual for AI that is already live. Chapter 6, Runtime Visibility and the Evidence Spine, explains what to record so you can reconstruct what a system did, and Appendix B gives an evidence logging schema you can adapt. Read it free, or check your own evidence spine with the free readiness assessment.
Next step: we will report the ruling in the New York case when it comes, along with what it means for anyone keeping AI records. To follow the evidence closely, including our analysis of court filings as they land, join Mercury Intelligence for $10 a month (or $100 a year; see what is included).
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Fed. R. Civ. P. 37. Failure to make disclosures or to cooperate in discovery; sanctions. https://www.law.cornell.edu/rules/frcp/rule_37
In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143-SHS-OTW (S.D.N.Y. July 14, 2026) (order re motion for sanctions, ECF No. 1633). https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1633.0.pdf
Miller v. Regions Bank, No. 2:24-cv-1324-HDM (N.D. Ala. May 21, 2026) (order imposing sanctions, ECF No. 39). https://storage.courtlistener.com/recap/gov.uscourts.alnd.191468/gov.uscourts.alnd.191468.39.0.pdf
News Plaintiffs. (2026, September 10). News Plaintiffs’ reply to OpenAI Defendants’ brief in opposition to News Plaintiffs’ motion for sanctions (ECF No. 1933), In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.). https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1933.0.pdf
OpenAI. (2026, August 14). OpenAI Defendants’ brief in opposition to News Plaintiffs’ motion for sanctions (ECF No. 1673, public redacted version), In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.). https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1673.0.pdf
Stipulated Order. (2026, July 15). Order regarding briefing schedule for News Plaintiffs’ motion for sanctions against OpenAI (ECF No. 1634), In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.). https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1634.0.pdf
Leave a Reply
You must be logged in to post a comment.