# Your Opponent Can Subpoena OpenAI. Your Business Chats May Not Be Protected.
If you are running a company on ChatGPT or Claude, a dispute you hope never happens may already have a discovery path running through your AI account. On June 4, 2026, a New York court made that concrete. In a fight between members of an LLC, one side served a subpoena on OpenAI directly, demanding the other side's entire ChatGPT account: every prompt, every uploaded document, every output.
The court quashed the subpoena. That sounds like good news for anyone who uses AI. Read more carefully, it is not the reassurance it appears to be. The subpoena was quashed for a specific reason that has little to do with how founders use AI day to day. The more durable lesson is the one the result obscures: AI vendors are now a named stop on the discovery map, and an adversary can go to them directly.
In April I covered the federal split on whether talking to a chatbot waives privilege or work-product protection. You can read that analysis at /sparkpoint/ai-privilege-heppner-warner/. This piece answers a different question: which of your AI records can an adversary actually reach, including through the vendor, and how should you govern AI use before a dispute starts?
What the court actually held
The case is Assini v. Hayward, 2026 NY Slip Op 26086 (Sup. Ct. Nassau County June 4, 2026, entered June 10, 2026), a dispute among members of Alpha Tech Lending LLC. One of the parties, representing himself, had used ChatGPT to help prepare his litigation filings. The opposing side served a non-party subpoena under CPLR 2304 on OpenAI OpCo, LLC, asking for his entire account: all prompts, inputs, uploaded documents, and outputs.
Judge Rhonda E. Fischer quashed it on work-product grounds. Under CPLR 3101(d), materials prepared in anticipation of litigation are protected work product, and the court concluded that a self-represented litigant's AI-assisted litigation preparation can qualify, the same way an attorney's preparation would. Because the materials were protected, a subpoena aimed at OpenAI as their third-party custodian could not reach them either.
In getting there, the court followed two recent federal civil decisions, Morgan v. V2X, Inc. (D. Colo. Mar. 30, 2026) and Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026), both of which extended work-product protection to a self-represented party's AI use. It declined to follow the criminal case at the other end of the spectrum, United States v. Heppner, No. 1:25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026), where a defendant's consumer Claude use, undertaken on his own and not at counsel's direction, failed both privilege and work-product analysis.
The court also pointed the litigant to 22 NYCRR Part 161, New York's new rule on AI use in court filings, effective June 1, 2026. New York does not prohibit using AI to prepare submissions, and it generally does not require disclosing that AI was used. What the rule requires is that the filing party review the work and confirm it does not contain fabricated or fictitious cases, statutes, or other material. The court warned, citing Augustin v. Formula 3 Brooklyn Inc., 2025 NY Slip Op 51113(U) (Sup. Ct. Kings County 2025), that careless AI use can frustrate a case. There is a quiet irony in the sequence: the ruling that protected this litigant's AI use came under a rule whose central command is to make sure AI has not invented your citations.
The mistake founders will make
It would be easy to read Assini as a green light, proof that what you type into ChatGPT is protected. That reading is wrong, and the gap between what the case actually protects and what founders assume it protects is where the risk lives.
Work-product protection, under CPLR 3101(d) and Federal Rule of Civil Procedure 26(b)(3), turns on a single threshold: the material has to have been prepared in anticipation of litigation. A self-represented litigant drafting court filings clears that bar. So does an attorney preparing for trial. A founder drafting a partnership contract, writing a board memo about a strategic pivot, working through a difficult termination, or shaping a fundraising narrative does not. None of that is litigation preparation, however sensitive it is, and however much you would prefer a court never read it.
So the materials that Assini, Morgan, and Warner protected and the materials a founder generates every day sit in two different legal lanes. Those cases all involved people who were already in litigation and using AI specifically to prepare for it. Ordinary business use is a different category, and it does not inherit their protection.
The three-lane discoverability map
Most founder AI use falls into one of three lanes. Knowing which one you are actually in is the practical starting point.
Lane A: Ordinary business use. Contracts, strategy documents, board memos, HR decisions, fundraising materials, vendor negotiations. This is where founders spend the large majority of their AI time. It is not work product, because it was not created for litigation. In a dispute, it is presumptively discoverable if relevant and proportional to the case, and it may be sought directly from the AI vendor through a non-party subpoena. The fact that you typed it into a chat window rather than a document does not change the analysis.
Lane B: Litigation preparation you do on your own. A dispute is brewing, and you start using ChatGPT to think through your position, draft arguments, or organize documents about the conflict, without a lawyer directing the work. This is the contested zone. Assini, Morgan, and Warner suggest civil work-product protection can apply here, but the question is genuinely unsettled, and the criminal court in Heppner came out the other way. Treat Lane B as a real possibility, not a guarantee.
Lane C: Attorney-directed litigation preparation. Your lawyer instructs you to use an AI tool in a specific way, or uses it as part of their own preparation on your behalf. This is the strongest position. Judge Rakoff expressly left this path open in Heppner, invoking the older Kovel doctrine (United States v. Kovel, 296 F.2d 918 (2d Cir. 1961)): when counsel directs the use and treats the output as part of her own work product, the protection is considerably stronger. No court has squarely applied that framework to AI yet, but the case that tests it is coming.
The realistic picture is that nearly all of a founder's AI use is Lane A, a sliver might reach Lane B once a dispute is already visible on the horizon, and Lane C exists only when you have deliberately set it up with counsel.
What founders should do now
None of this is a reason to stop using AI. It is a reason to use it with a clear view of what a dispute can reach. Five steps matter most.
1. Separate ordinary business AI use from litigation-related AI use. Run the three-lane map as a quick internal inventory. If most of your AI use is Lane A, which it almost certainly is, work from the assumption that it is reachable in a dispute. That does not mean writing less. It means being deliberate about what goes into a chat window, and on which account, when the subject is genuinely sensitive.
2. Move sensitive work into enterprise tools with appropriate confidentiality and disclosure terms. Consumer terms for most major platforms reserve the provider's right to disclose user content in response to court and law enforcement requests, without notice to you. Enterprise plans (ChatGPT Enterprise, Claude for Work, Claude Enterprise) run on negotiated contracts that can add confidentiality obligations, restrictions on training and reuse, and notice before any disclosure, giving you a chance to object or seek a protective order first. This is not privilege, and it does not put the records beyond a court's reach. It improves your practical position before a dispute exists.
3. Prohibit company-sensitive work in personal AI accounts. A subpoena to OpenAI for an employee's personal account is harder for the company to control or contest, because the company is not the account holder and may not receive notice before production. A written acceptable-use policy that bars personal consumer accounts for company work, paired with one approved enterprise tool per use case, closes that gap directly.
4. Add AI chat history to litigation-hold planning. Preservation duties attach once litigation is reasonably anticipated. If your AI tools auto-delete conversation history and a dispute triggers a hold you cannot honor, that is a preservation problem independent of discoverability. Check retention settings now, and know how a hold would actually execute across your AI accounts.
5. Route dispute-related AI use through counsel before the material is created. Work-product protection is strongest when litigation preparation is directed by counsel, the Lane C posture. That structure has to exist before the material is made; it cannot be added afterward. When a dispute starts to look real, that is the moment to involve counsel and let them shape how AI is used.
For investors and boards, the same facts read as a governance question. A portfolio company running sensitive compensation, governance, and strategy discussions through personal consumer AI accounts is building a record no one can easily preserve or protect. Asking whether a company has an AI acceptable-use policy and enterprise terms with meaningful confidentiality provisions is a fair diligence question, and in a dispute involving the company, that record can matter to you too.
Frequently Asked Questions
Can my adversary really subpoena OpenAI or Anthropic directly?
Yes. The mechanism used in Assini, a non-party document subpoena under CPLR 2304 in New York, with close equivalents in most other jurisdictions, lets a party serve a subpoena on a third-party custodian rather than on the opposing party. An AI platform can be that custodian. Whether a court enforces it depends on the facts and on whatever protective doctrines apply. In Assini, a work-product defense was available because the materials were litigation preparation. For ordinary business use, that defense generally would not be.
If my AI use was for business, not litigation, is there any protection?
The main protection is operational and contractual, not privilege-based. Work-product and privilege doctrines generally do not reach ordinary business use, so what you are left with is how you set things up: which account the work lives on, and what your vendor terms say. Enterprise terms with confidentiality obligations, training and reuse restrictions, and a notice-before-disclosure provision can improve your practical position and give you a chance to object or seek a protective order. They do not automatically bar a court from ordering production. The protection is real but limited, and it is strongest when it is in place before a dispute begins.
What does 22 NYCRR Part 161 require?
It does not prohibit AI use in New York court submissions, and it generally does not require disclosing that AI was used. What it requires is that the filing party review the AI-assisted work and confirm it does not contain fabricated or fictitious cases, statutes, or other material. That duty is grounded in the existing certification rule (22 NYCRR 130-1.1) and the duty of candor to the tribunal. The obligation runs to the accuracy of what you file, not to the fact that a tool helped you prepare it.
Closing Perspective
Assini is being read in some quarters as a win for people who use AI in litigation. In a narrow sense it is. The more useful read is that a court has now put on the record the discovery path an opposing party may try: serving the AI vendor directly, without going through you first.
The defense that worked in Assini worked because the litigant was already in a case and using AI for that case. Most founders are not doing that. They are building companies, and their AI is a business tool. That single distinction, between preparing for litigation and running a business, is the line between the records a court will protect and the records a court can order produced.
The practical point is this: by the time a dispute arrives, the relevant chat history already exists, and you cannot un-write it. The decisions you can still make are the ones in front of you now, about which tools you use, on which accounts, for the work that would be hardest to explain later.
This article is for informational purposes only and does not constitute legal advice. Every company's situation is different, and you should consult with qualified legal counsel before making decisions based on the developments discussed here.