[Editor’s Note: This article has been republished with permission. It was originally published October 14, 2026 on the Minerva26 Blog]
A federal court in Michigan read a former sales rep’s ChatGPT questions about whether Yahoo could recover his deleted emails under a subpoena, found that he intended to deprive his former employer of evidence under Rule 37(e)(2), and imposed an adverse inference that then helped carry a preliminary injunction against him. Every AI ruling in this series so far has asked whether a party’s AI use is protected; this one asks what it proves. The lesson for litigators: your clients’ AI chat history is ESI, it is evidence of state of mind, and you need to talk to them about it before they use AI on their case. Reading time approximately 10 minute
Podcast | Transcript
Your Clients Are Asking AI About Their Cases
Your clients are asking AI about their cases, and not only your corporate clients with enterprise tools. The sales rep in a non-compete fight, the employee in a discrimination case, and the business owner in a contract dispute all ask before they call you, after they hang up with you, and in the middle of discovery. Every one of those conversations is stored on a phone, a laptop, or an account, with a timestamp, in the client’s own words.
This came up in more than several conversations I had at Relativity Fest this year. Lawyers are frustrated by clients who arrive with “I consulted AI, and here’s what it told me,” and I’ve had clients whose AI answers were either part of the story or completely wrong. That trend is happening now, and you need to be prepared to respond to it.
In Conlan Tire Co. v. Gonzales, a federal court read a former sales rep’s ChatGPT questions about whether his deleted emails could be recovered under a subpoena and used them to find that he intended to deprive his former employer of evidence. The resulting adverse inference then helped carry a preliminary injunction against him. AI chat history now belongs on every preservation letter and forensic collection protocol, and if you are not asking for it, opposing counsel may be.
What This Case Is About
Conlan Tire hired Alec Gonzales as an account executive in July 2022 and has a copy of his Confidentiality and Non-Solicitation Agreement, signed July 13, 2022. Conlan terminated him in December 2024, he went to a direct competitor, and Conlan sued in March 2025, alleging he was soliciting its customers.
Gonzales had exactly one defense: he never signed the Agreement. His first declaration said he did not recall signing it. His second, filed June 16, 2025, said he “never agreed to nor personally signed” it, and swore that he still had the MacBook and cell phone he used in 2022 and that neither could apply an electronic signature.
Because the case turned on whether the signed Agreement went back and forth by email, the parties stipulated to an ESI order on June 24, 2025, requiring Gonzales to give Conlan’s forensic examiner access to his Yahoo account by the end of June and to his 2022 devices and iCloud account by July 10. He did not provide Yahoo access until July 7, a week late, and he produced a laptop but not his 2022 iPhone, which he now said he no longer owned.
Three Questions to ChatGPT
The Yahoo account showed no emails exchanging the signed Agreement, but the examination of his devices found something else. On June 10, 2025, Gonzales asked ChatGPT whether Yahoo could retrieve an email deleted a year ago, even under a subpoena (no); how long a permanently deleted email stays retrievable by Yahoo (at most seven days); and whether Yahoo logs permanently deleted emails (no). Six days later, he filed the declaration saying he never signed the Agreement. Conlan moved for sanctions under Rule 37(e).
Intent to deprive under Rule 37(e)(2) is the hardest thing to prove under that rule. Parties almost never say out loud that they deleted something so the other side would not get it, so courts build intent from timing, inconsistent explanations, and conduct around the loss, as in Hunters Capital, where City of Seattle officials lost their phones one after another. In Conlan, the party wrote down what he was thinking.
Ruling One: The ChatGPT Questions Established Intent to Deprive (ECF No. 38)
Gonzales said he asked the questions to test whether Conlan’s proposed iCloud search was a fishing expedition, and argued that ChatGPT’s answers would have discouraged deletion because deleted emails stay retrievable for 30 days. One has to wonder whether he asked ChatGPT for those arguments.
The Court did not agree with either explanation. It did not find his justification for his “AI-assisted research into the mechanics of deleted Yahoo email retrieval under a subpoena convincing or even credible,” because the questions had no rational connection to the iCloud request. And it read the ChatGPT answers more carefully than he did: the 30-day window applied to emails in the trash folder, while emails removed from the trash were gone within seven days. He gave the examiner access on July 7, 28 days after ChatGPT told him that. His shift from “I don’t recall” to “I never signed” came six days after the questions, and the Court found that coincidence “strains credulity.”
The Court called his explanations “contrived, implausible, and deceptive” and found “compelling evidence of intentional deprivation.” But it held the adverse inference in abeyance because it was “not yet convinced” the emails could not be replaced through additional discovery, which is the threshold element of Rule 37(e). It awarded Conlan its fees for litigating the existence of the emails and ordered targeted discovery: an expanded Yahoo search, every device and cloud account Gonzales used in 2022, a chance for him to inspect Conlan’s ESI, and a deposition of Conlan’s HR coordinator.
Ruling Two: Replacement Discovery Stays Narrow (ECF No. 45)
Gonzales then tried to use that order to inspect Conlan’s 2024 and 2025 ESI and question the HR coordinator about Conlan’s IT systems and retention practices. The Court denied his motion to clarify, holding that the discovery’s sole purpose was recovering his deleted July and August 2022 emails and calling his broader reading baseless. On Case of the Week we stay within the four corners of the decision, but it appears Conlan’s counsel did a very good job of keeping the Court focused on exactly when things happened and what was relevant.
Ruling Three: A Permissive Adverse Inference (ECF No. 48)
The additional discovery made things worse for Gonzales. The expanded Yahoo search found emails exchanging other signed pre-employment documents with Conlan’s recruiter, contradicting his sworn claim that he signed all onboarding documents in person. The laptop he produced was manufactured in 2024. He had concealed a second iCloud account that he accessed in 2022 and during the litigation, and he refused to produce his current iPhone. Conlan no longer had any 2022 emails, and the HR coordinator had no direct knowledge of whether he returned the Agreement.
The Court held that the emails “cannot be restored or replaced through additional discovery,” found prejudice because they were the most relevant evidence on the only disputed issue, and imposed the remedy. Pay attention to its exact terms: the Court “will presume the deleted emails were unfavorable to Gonzales and will instruct a jury that it may also presume” the same. That is a permissive instruction under Rule 37(e)(2)(B), not a mandatory one.
Ruling Four: What the Adverse Inference Was Worth (ECF No. 49)
The same day, the Court granted in part Conlan’s motion for a preliminary injunction. It relied on the HR testimony, Gonzales’s evolving and partly false declarations, and the presumption that the missing emails would show he returned the signed Agreement, concluding that “there is much evidence that Gonzales signed the Agreement.” The injunction runs until April 29, 2027. Three ChatGPT questions led to a finding of intent, his handling of replacement discovery made the emails irreplaceable, and the adverse inference helped carry the injunction.
Two Questions I Would Have Looked at for the Defense
The Court noted that nothing in the record ruled out deletion before June 10, and emails deleted long before any dispute may not be ESI that should have been preserved. But that argument was not available to Gonzales, because it would have meant admitting the emails existed, and his entire defense was that he never signed or returned the Agreement. His defense rested on his own credibility, the emails were the only thing that could have backed him up, and every step he took in discovery made him less believable. When your client’s case is his word, the forensic record has to support it. Here, it did the exact opposite.
Rule 37(e) makes “cannot be restored or replaced” a threshold for both subsections, yet in August 2025 the Court awarded fees while saying it was not yet convinced the information could not be replaced, without explaining the authority for doing so. That’s my read, not a holding, and courts have broad discretion, but it is a question worth raising.
Where Conlan Fits in the AI-in-Discovery Line
Every AI decision in this series so far has asked whether a party’s AI use is protected. Warner v. Gilbarco and Morgan v. V2X held that a party’s own AI use can be work product under Rule 26(b)(3); United States v. Heppner found no protection when a client used AI outside the relationship with counsel; and Tate Group Automotive v. Legacy Automotive Capital protected an executive’s ChatGPT conversations because Texas Rule 192.5 covers work product prepared by or for a party, which is broader than the federal rule.
Conlan asks what AI use proves. Nobody claimed protection, because the conversations surfaced through a court-ordered forensic exam, and the queries were not research into the strengths of a claim, as in Morgan, Warner, and Heppner, but research into whether deleted evidence could be recovered. That is why litigants’ own AI use is going to show up in discovery consistently, and not just in big, complex matters. This is a single-defendant non-solicitation case about a tire salesman.
What to Do This Week
Treat AI chat history as ESI you identify, collect, preserve, and produce. These conversations came out of a forensic exam of the defendant’s own devices; nobody subpoenaed OpenAI. If your forensic protocol, preservation letter, or requests for production don’t name AI chat history, including ChatGPT, Claude, Gemini, Copilot, or whatever is on the phone, add it.
Talk to your clients about AI before they use it on their case. After Heppner came out, I and a number of colleagues sent emails to our clients that same day. There are fair questions about whether Heppner holds up in a civil context, but Conlan shows the other half of the risk: unprotected AI use is evidence of state of mind. Gonzales had a lawyer and still went to ChatGPT to research recovering deleted emails. Anything your client asks an AI tool about the case is a written, timestamped record the other side and the judge may end up reading. If you don’t want to see it on the cover of the New York Times, don’t write it down.
Verify the device facts before your client signs a declaration. Gonzales swore he still had his 2022 phone and laptop, and the Court relied on those false statements twice. Counsel was not sanctioned here, but judges in other cases have held counsel responsible for exactly that.
If you’re the responding party, treat replacement discovery as your best chance to avoid Rule 37(e)(2). As in U.S. ex rel. Staggers v. Medtronic, restoration and replacement is its own evaluation, and full compliance is the only way to use it. Gonzales produced the wrong laptop, concealed an iCloud account, and withheld his phone, and that conduct is what let the Court find the emails irreplaceable.
Make sure your client’s own records can answer the central question. Conlan had no 2022 emails, so whether a signed agreement came back depended on the former employee’s personal Yahoo account. For any agreement your client will need to enforce, the record of transmission and return matters as much as the signed document. Document, document, document the processes.
What to Watch Next
I expect more cases like this one, on both sides of the v., in employment, trade secret, and family law matters, using AI conversations to prove knowledge, intent, and timing. What made Conlan work is that the conversation includes the answer the user got back, so the Court could measure his conduct against what ChatGPT told him rather than how Yahoo actually works.
Watch for requests aimed at the accounts themselves. Many of these tools keep conversation history in the user’s account, and I expect parties to ask for account access and exports directly, the way Conlan required iCloud credentials.
And watch for the collision with the protection cases. At some point, a represented litigant will argue that his own AI research is work product under Warner and Morgan, and the other side will point to Heppner. My read is that research into making evidence unrecoverable will not get protection in any court, but research into the strengths of a claim or defense is a much harder question.
She is a co-author of Electronic Discovery and Records and Information Management, and host of Case of the Week at Minerva26. As Principal at ESI Attorneys, Kelly manages the boutique eDiscovery and information law firm that acts as operational business partners with its clients to advise law firms, corporations, and municipalities on all areas of electronic information including eDiscovery, privacy, cybersecurity, and information governance.
Kelly is also the CEO of Minerva26. — a SaaS-based practical resource for litigators handling eDiscovery — that curates discovery decisions, rules, and additional content. She is developing an online academy to provide on-demand education for lawyers and legal support professionals to stay abreast of changes in the law and technology that affect litigation and clients’ obligations to respond.
You can reach Kelly at [email protected], join her Facebook community group at Let’s Talk eDiscovery, or connect with her on Twitter @kellytwigger.