Your Architect Asked ChatGPT About a Code Violation. A Jury Might Read That Conversation.

A project manager at a mid-size architecture firm in Virginia notices a potential egress issue in a residential design after framing is already underway. Before calling the firm's attorney, she opens ChatGPT and types: "What exposure does an architect have for delay damages if shop drawings were late?" Twelve seconds later, she has a response. She also has a timestamped record, stored on OpenAI's servers, that she knew about the problem before anyone else on the project did.

That record is almost certainly discoverable in court.

Four Rulings, One Pattern

Federal courts issued four significant decisions on AI chatbot discoverability in the first half of 2026, and the split they reveal matters enormously for anyone in residential construction. A federal judge in the Southern District of New York ruled on February 17, 2026 in United States v. Heppner (No. 25-cr-00503-JSR) that 31 documents a defendant created using the Claude chatbot were not protected by attorney-client privilege or the work product doctrine. The court's reasoning was surgical: the information was not communicated to an attorney, was not confidential because it was shared with a third-party platform under policies permitting disclosure, and was not generated for the purpose of obtaining legal advice.

Three other courts disagreed with varying reasoning, and the pattern in those rulings reveals a critical distinction. A judge in the Eastern District of Michigan denied a motion to compel a plaintiff's ChatGPT records in Warner v. Gilbarco (Feb. 10, 2026), finding that "ChatGPT and other generative AI programs are tools, not persons" and that work-product waiver requires disclosure to an adversary. A Colorado federal court reached a similar conclusion in Morgan v. V2X (Mar. 30, 2026), holding that AI use "does not eliminate all expectations of privacy or automatically waive protections," and in June a New York state judge in Alpha Tech Lending v. Recchio quashed a ChatGPT subpoena entirely, calling the defendant's AI chats "confidential, strategy-laden iterative work product."

Reassuring, until you read the fine print: every court that protected AI chat logs was ruling on a pro se litigant, someone representing themselves who acts simultaneously as party and legal advocate. Not one of those rulings addressed a corporate employee typing questions into a chatbot on company time.

What Fortis Means for Your Firm

How corporate AI use plays out in court became clear in Delaware in March. In a case called Fortis Advisors v. Krafton (Del. Ch. March 2026), the Court of Chancery examined a CEO's ChatGPT conversations and concluded he had used the platform to "contrive a corporate takeover strategy" designed to eliminate a $250 million earnout obligation. Chat logs provided what the court called damning evidence of pretextual intent: the CEO had asked the chatbot for a step-by-step plan, and then executed it, prompting a sweeping remedial order that reinstated the target company's CEO with full operational control.

Nobody argued those chats were privileged, and nobody could have, because a CEO asking a chatbot how to restructure a deal is not attorney-client communication. Neither is a project manager asking Claude whether a window schedule meets IRC Section R310.1 minimum opening dimensions.

Why Construction Firms Sit on the Wrong Side of This Split

The AIA Trust published a warning in April 2026 that laid out the construction-specific risk with uncomfortable clarity. Jonathan C. Shoemaker of Lee/Shoemaker PLLC, a firm representing design professionals, described scenarios that any PM or firm principal would recognize: asking an AI platform about liability exposure before escalating to counsel, researching whether a missed inspection constitutes negligence, having a junior team member query a chatbot about a workaround instead of flagging the issue to a supervisor.

Shoemaker's conclusion is blunt: "There is no established body of law creating a safe harbor for AI-based legal research conducted by non-lawyers."

A Reuters analysis published June 8 reinforced the point, noting that "the dangers of using publicly available AI tools have far greater impact on corporate litigants" than on individuals representing themselves, and that the Department of Homeland Security obtained what appears to be the first federal search warrant for ChatGPT user data as early as 2025. Consumer AI platforms log every prompt, those logs are retrievable through legal process, and under broad civil discovery rules they are almost certainly producible.

110,000 Homes in the Discovery Window

According to the NAHB/Wells Fargo Housing Market Index survey from July 2025, 49% of single-family homebuilders use AI in some capacity. Applied to the Census Bureau's roughly 900,000 annual single-family starts, that means approximately 441,000 new homes per year involve an AI-using builder, architect, or subcontractor generating chatbot logs that a plaintiff's attorney could theoretically subpoena.

Not all of those homes will face litigation, but construction defect claims are rising fast. A University of Hawaii Economic Research Organization study found that 27% of new homes built in Hawaii between 2013 and 2023 were involved in construction defect litigation when class actions are included, up from 16% in the previous decade, and that 15% of multifamily units alone were encumbered by claims, up from 4%. Hawaii is an outlier, but the trend tracks nationally. Apply even a conservative 25% litigation rate to that 441,000-home pool, and roughly 110,000 homes per year sit in a zone where AI discovery requests could arise, each one representing a potential subpoena to OpenAI, Anthropic, Google, or Microsoft for employee chat logs.

Rough math, admittedly, since it multiplies a national adoption rate by a high-litigation-market claim rate, and most markets have lower litigation exposure than Hawaii. But the directional signal is hard to ignore: hundreds of thousands of construction professionals are generating discoverable records every day, and almost none of them know it.

When Knowledge Becomes Liability

Discovery risk alone is concerning, but what makes AI chat logs uniquely dangerous in construction defect cases is what they prove: knowledge. Most defect claims turn not on whether a problem existed but on when the builder or architect knew about it and what they did afterward, and AI chat logs answer that question with a precision that emails and meeting notes rarely achieve because people ask chatbots things they would never commit to an email, and they do it at the exact moment concern first arises.

"Your project manager asked ChatGPT about waterproofing code requirements on March 5, and the inspection failed on June 12, so what happened in the ninety-nine days between?"

Posed to a jury, that question is devastating. Cybersecurity researchers quoted in an NBC investigation of AI chat forensics described AI logs as a "treasure trove" for investigators precisely because users treat chatbots with a candor they would never show in formal communications, and construction professionals are no exception. A PM who would never email "I think we might have a code problem" will absolutely type it into ChatGPT at 11 p.m. looking for reassurance.

A Strong Counterargument

Most courts have leaned protective so far, and the legal framework remains unsettled. Three of four 2026 rulings shielded AI chats from discovery. Enterprise AI platforms with strict data retention policies and contractual confidentiality provisions offer a meaningfully different risk profile than consumer ChatGPT, and a firm using a private, on-premises AI deployment with no third-party data sharing has a much stronger argument against discoverability than one whose employees are typing sensitive questions into a free-tier chatbot. Heppner rested heavily on Claude's user policies permitting disclosure to third parties; a platform with different terms might yield a different result.

But three of those four protective rulings involved pro se litigants, and construction firms do not represent themselves. Under Heppner's reasoning and the evidence framework established in Fortis, a corporate employee's chatbot history sits squarely in the discoverable category until a court says otherwise.

What This Analysis Does Not Cover

No published construction defect case has yet involved a discovery dispute over AI chatbot logs, so every scenario described here extrapolates from adjacent case law in commercial litigation, employment discrimination, and criminal proceedings. Hawaii's litigation data represents one state with unusually high defect claim rates, and national figures would be lower. Enterprise AI platforms may have contractual protections that consumer tools lack, though those protections remain untested in court, and the AIA Trust's warning is a trade organization publication rather than a judicial opinion.

Four Things to Do Before Your Next Project

Adopt a written AI governance policy that defines permissible uses and prohibits employees from seeking legal opinions through consumer AI platforms, which the AIA Trust recommends as the minimum floor. Route all legal questions to actual attorneys to preserve privilege, because if an employee has a question about liability, code compliance, or exposure, the answer needs to come from a lawyer whose communications are privileged rather than a chatbot whose logs sit on a server in San Francisco.

Evaluate your AI tools carefully, because enterprise platforms with opt-out provisions for model training, limited data retention, and contractual confidentiality provisions carry materially lower discovery risk than consumer products, and if your firm uses a private deployment that does not transmit prompts to a third-party server, document that architecture because it will matter in litigation.

Most critically, train your team on the risk, because every construction litigation attorney who has examined this issue agrees on one point: most professionals who use AI for legal research believe those conversations are private, but every prompt is logged and every log is potentially producible. That project manager in Virginia who typed her question about delay damages at 9 a.m. created a piece of evidence that may outlast the building she was worried about.