A New York federal court just issued the first published ruling to tackle head-on whether conversations with a public AI chatbot are protected by attorney–client privilege or the work product doctrine.
To use a legal term: Nope. The decision gives lenders a new weapon when chasing borrowers who engage in behavior to hinder, delay, or defraud creditors.
The case involved the target of a federal investigation who used a public AI platform to create strategy-focused “reports” about the facts and law of his case. Federal agents later seized electronic devices containing those AI exchanges during a search. The defendant claimed privilege and work product protection, arguing that because he eventually shared the AI outputs with his lawyers (and had fed in information he’d gotten from his lawyers), the materials should be shielded. The court said no.
First, the software engineers at OpenAI and Anthropic aren’t sure what they have created, but everyone can agree that it’s not a licensed attorney. Attorney-client privilege only protects confidential communications with your attorney. Until Claude or ChatGPT hangs a shingle and sends you an engagement agreement, it doesn’t qualify. Second, the defendant had no real expectation of confidentiality because the AI provider’s terms of service allowed the company to collect user inputs and outputs, use them to train its models, and even disclose them to third parties, including regulators.
Third, the chats weren’t about getting legal advice from counsel. The defendant started these conversations on his own, and the AI tool itself disclaimed giving legal advice. Finally, the court rejected the work product argument. Work product protection is meant to shield a lawyer’s thinking and strategy. These chat transcripts were created by the client, on his own, using a public tool. The court added that even information that starts privileged loses its protection once it is pasted into a public chatbot. If courts follow this decision, then every AI chat outside the attorney-client relationship is a potential exhibit waiting to be produced.
That got us thinking about the many ways that desperate and nefarious debtors behave when they are insolvent, in the shadow of bankruptcy, or looking to hide assets or hinder the collection efforts of their creditors. The schemes can be as complex as Rube Goldberg inventions, with layer-upon-layer of shell companies and special-purpose entities. The United States continues to produce people engaged in the American hustle, and every day brings a new Ponzi scheme, Ponzi-adjacent scheme, or garden-variety fraudulent transfer. These people don’t want to get caught and will go to great lengths to get away with their schemes.
While Google behaves like a librarian, an AI can behave like a cyber confederate. Google will find you books and articles on money laundering, but an AI can structure a bespoke money-laundering scheme tailored to your unique situation. If the debtor can’t afford or doesn’t want to fly to an offshore safe haven to discuss moving assets in a setting befitting a James Bond movie, then he or she can simply have AI provide the best locales for the job.
Under state or federal law, certain transfers of assets can be undone if the debtor made the transfers intending to hinder, delay, or defraud creditors. The courts have long recognized that debtors are not likely to testify that they intended to defraud creditors. And rare is the case with a “smoking gun” email where the debtor lays out their fraudulent scheme or admits to it. But a court can infer fraudulent intent from the presence of “badges of fraud” and extrinsic evidence. Badges of fraud include:
(1) The transfer or obligation was to an insider;
(2) The debtor retained possession or control of the property transferred after the transfer;
(3) The transfer or obligation was disclosed or concealed;
(4) Before the transfer was made or obligation was incurred, the debtor had been sued or threatened with suit;
(5) The transfer was of substantially all the debtor’s assets;
(6) The debtor absconded;
(7) The debtor removed or concealed assets;
(8) The value of the consideration received by the debtor was reasonably equivalent to the value of the asset transferred or the amount of the obligation incurred;
(9) The debtor was insolvent or became insolvent shortly after the transfer was made or the obligation was incurred;
(10) The transfer occurred shortly before or shortly after a substantial debt was incurred;
(11) The debtor transferred the essential assets of the business to a lienor that transferred the assets to an insider of the debtor;
(12) The debtor made the transfer or incurred the obligation without receiving a reasonably equivalent value in exchange for the transfer or obligation, and the debtor reasonably should have believed that the debtor would incur debts beyond the debtor’s ability to pay as they became due; and
(13) The debtor transferred the assets in the course of legitimate estate or tax planning.
As for the “extrinsic evidence,” that is everything else with bearing on the credibility of the debtor and whether there was a legitimate or fraudulent motive behind the transfer. Wielding this decision, a lender could send discovery to a debtor to produce “all communications with AI-based tools, including prompts, inputs, and outputs.” And a debtor could be asked during a deposition, “Did you use any AI tools when you did X, Y, or Z?” Lenders are already asking for email, text, and direct messaging communications to find evidence of fraudulent intent, and this is the natural next step. If a debtor enlists an AI in a fraudulent scheme, this decision suggests that the courts can make it snitch.
