AINEWS
Talking to ChatGPT Isn't Talking to a Lawyer, Federal Judge Rules
A federal court ordered OpenAI to turn over Richard Kim's ChatGPT prompts, ruling they are discoverable records, not confidential communication with a lawyer.
A federal court ordered OpenAI to turn over Richard Kim's ChatGPT prompts, ruling they are discoverable records, not confidential communication with a lawyer.
AI assistants are legally what they have always been technically: third-party record-keepers. A federal judge in Manhattan confirmed that on Monday, ruling that prosecutors can pull the ChatGPT logs of Richard Kim, the former chief executive of the crypto startup Zero Edge, from OpenAI, and rejecting the argument that the prompts deserve the same legal shield as a conversation with a lawyer.
The warrant covers every prompt, response, and account record tied to Kim from October 2023 through May 2026, a window that includes both the alleged fraud and Kim's post-arrest use of ChatGPT to research his own defense (CNET). Federal prosecutors allege that Kim, a former Goldman Sachs and JPMorgan alum who later held a senior role at Galaxy's crypto business, diverted roughly $3.8 million of a $4.3 million seed round raised by Zero Edge into personal crypto trades and online gambling (Banking Dive. Kim has pleaded not guilty to securities and wire fraud charges, and his defense team has asked the court to quash the OpenAI subpoena, citing attorney-client privilege and the risk of exposing his trial strategy (InnerCity Press).
The legal move that will outlast this fraud case is structural. A warrant against a third-party records custodian, the same kind federal agents have used for years against email providers, telecom companies, and cloud-storage services, treats OpenAI as a holder of someone else's communications, not as a confidential intermediary. Schofield's order does not adjudicate privilege in the abstract; it finds that ChatGPT logs sit in the same legal bucket as search history, text messages, or a Gmail inbox, and that they can be reached by the same warrant a prosecutor would use to subpoena Google or Verizon.
The decision also fits a longer Southern District of New York pattern. Earlier this year, a different SDNY judge held that documents generated through unsecured public AI tools are not entitled to attorney-client or work-product protection, a warning shot to lawyers and executives who have been pasting sensitive material into ChatGPT for "research" (O'Melveny). The Proskauer alert covering both rulings frames the practical takeaway in starker terms: unsecured public AI tools are not a confidential channel, and any work product, witness preparation, or strategic analysis that runs through them can be reached by a warrant (Proskauer). The Harvard Law Review's write-up of United States v. Heppner, a March 2026 decision on the treatment of digital evidence, points in the same direction: courts have been steadily expanding what counts as a discoverable third-party record, and AI chat logs are catching up to email and search history (Harvard Law Review).
For anyone who has used a public chatbot to think through a medical question, draft a sensitive email to an employer, plan a divorce, or ask whether a financial move might be illegal, the practical lesson is the same one a defense attorney would give: an AI is a service, not a confessor, and the prompts you send it sit on a server controlled by a company that can be subpoenaed. The legal protections that attach to a call with a doctor, a therapist, or a lawyer have not been extended to those prompts, and there is no near-term sign that they will be.
Two things will matter next. The first is whether Kim's motion to quash survives appellate review, since a reversal would narrow the warrant's reach in this case but leave the broader third-party records doctrine intact. The second is whether other judges, inside and outside the Southern District, follow Schofield's lead in treating public AI tools as discoverable intermediaries, the way they already treat Google, Verizon, or Dropbox. If they do, the era of treating a chatbot conversation as a private think-aloud ends by default, not by legislation.