A dozen US cases turn chatbot prompts into ordinary discoverable exhibits

Twelve cases in twenty-four months is one every eight weeks or so ^[3], a base rate low enough to invite the wrong conclusion in either direction. Discovery exposure has never scaled with how many records a firm generates; it scales with whether the one record that decides the question exists in a form opposing counsel can name and request ^[3].

There are two channels here and they are not the same size. The government-request channel is narrow: 75 content requests to OpenAI in the second half of 2025, compliance in 62 of them, 84 accounts involved ^[9], which is a compliance rate of about 83 percent ^[1] and roughly 1.35 accounts per complied request ^[2], meaning subpoena-by-subpoena work. The civil channel is industrial. A federal magistrate ordered OpenAI in May 2025 to preserve user logs including deleted chats ^[7], and by early 2026 that line of litigation had produced 20 million de-identified conversation logs in a copyright matter ^[8]. The units do not match (conversations against accounts), but a ratio of about 238,000 to one ^[4] tells an operator which pipe to model.

The preservation order is the part that turns this from a privacy story into an internal controls question. A consumer product's delete button is simply an interface affordance, and it carries no retention schedule of its own; once a court orders preservation, the vendor's obligation outranks the user's housekeeping ^[7]. In the Missouri matter, investigators did not need the vendor at all: the query about detecting responsibility after vandalism came off the suspect's own device ^[5].

The corporate shape of this is already legible in the R.K.C. filings, where defence lawyers put a teenager's October 2024 ChatGPT conversations, including discussion of a potential $1 million settlement, into the public record ^[4][2]. Swap an employee for the teenager and a warranty reserve for the settlement number, and the exhibit writes itself. A firm with no approved-tools list and no retention period has effectively chosen the indefinite retention posture by default, and it is funding the consequences out of a legal budget it never sized for the purpose.

The exposure described here has a real limit: the 20 million logs went across de-identified ^[8], and some courts have grappled with work-product protection where attorneys used chatbots during litigation preparation, with mixed results and a patchwork by jurisdiction ^[10]. If a higher court extends something like work-product to AI sessions, the no-privilege reading narrows to everyone who is not a lawyer, and the exposure I am describing gets smaller for the firms with counsel in the loop. If instead the count stays at a dozen for another two years while usage grows, the honest verdict is that this was a tail risk that never arrived. Neither branch changes what the courts have said so far, which is that a prompt sits in evidence the way an email does ^[3].