Raffaella Aghemo
The use of generative artificial intelligence tools is always a subject of debate, even in the US Federal Courts, where, curiously enough on the very same date, two rulings took diametrically opposed directions.
I do not know if you recall the case United States v. Heppner, No. 25-cr-00503-JSR (SDNY 28 October 2025), a federal securities fraud case against Bradley Heppner, a former financial sector executive, who had used a third-party generative AI tool, Anthropic’s Claude, to generate requests for information regarding government investigations and his potential legal exposure, without any prompting from his defense counsel, and which had produced some thirty-one documents containing the defendant’s prompts and outputs.
The ruling, handed down at first instance by US District Judge Jed Rakoff of the Southern District of New York, established that a defendant’s use of generative artificial intelligence to analyse their legal exposure is not protected by legal professional privilege or the work product doctrine.
For the sake of completeness, it should also be noted that, in this same ruling, attorney-client privilege was not granted, as the defendant was not a lawyer, nor was the so-called ‘work-product’ privilege granted, as the ruling states: ‘The work-product rule provides qualified protection for materials prepared by or at the request of a lawyer in anticipation of litigation or for a trial.’ Furthermore, as admitted by the defense counsel in discussions with the Government, the defendant’s research on artificial intelligence was not prepared at the request of anyone, including the lawyer, but on his own initiative, without any guidance from the defense counsel, although he did subsequently share the results once it had been created.
However, in another case, Warner v. Gilbarco Inc., No. 2:24-cv-12333, the Michigan District Court judge ruled that documents relating to a self-represented claimant’s use of ChatGPT to draft legal documents were protected by the work-product doctrine. You can find the documents for this case at this link: https://dockets.justia.com/docket/michigan/miedce/2:2024cv12333/379552.
The claimant, a former employee of the defendant, brought a claim for discrimination and other legal actions against her employer, Gilbarco, Inc., and represented herself in court, admitting during the discovery phase that she had used ChatGPT to draft her court documents.
Although the defense had requested a broad discovery of ‘all documents and information’ concerning the claimant’s use of third-party artificial intelligence tools in relation to the litigation, the claimant objected, arguing that such materials reflected her internal analysis and mental impressions, and were therefore protected from disclosure.
The court ruled in favour of the claimant and, rejecting the defendant’s arguments, concluded that Warner’s interactions with ChatGPT were protected by the work-product doctrine. The document states: ‘Insofar as the defendants seek the production of “all documents and information regarding the use of third-party artificial intelligence tools in connection with this case, as requested by the defendants”, this request is DENIED, as the information sought is not available, and as a general rule, a party may not discover documents and tangible things prepared in anticipation of litigation or for a trial by a claimant or their representative.” It adds: “ChatGPT (and other generative AI programmes) are tools, not people, even though there may be administrators in the background. As noted by the claimant in her response, the defendants’ motion ‘to ask the Court to compel the production of the claimant’s internal analysis and mental impressions, that is, her thought process, rather than any existing document or evidence, is not discoverable as a matter of law.’
In essence, on the same day, 10 February 2026, the two judges reached conclusions that appear to diverge, although these decisions highlight the specific nature of the facts regarding privilege and the analysis of work-product in the context of AI, and emphasise the importance of: 1) how AI tools are used in relation to legal matters; and 2) the importance of matching the right tool to each task.
In this specific case, one consideration made by the judges of the Michigan Court appears central, namely that the work-product doctrine can safely be applied to materials created with the aid of AI. Even if the information relating to the use of AI were otherwise available, in these circumstances it is nonetheless subject to the protection afforded by the work-product doctrine, which expressly protects ‘documents and tangible things prepared in anticipation of litigation or a proceeding by another party or its representative’. As the claimant was acting on her own behalf, she was entitled to invoke work-product protection in respect of that material.
What does appear to be at odds, however, is the argument regarding the rejection of the defendants’ request, who contended that the claimant had waived work-product protection by using ChatGPT (presumably the free version). The court rejected this argument, explaining that waiving work-product protection requires disclosure “to an opposing party or in a manner that could end up in the hands of an opposing party”, and did not attach significance to the phrase “even if there might be administrators in the background”.
In conclusion, it can be inferred from a comparative analysis of the two concurrent federal decisions that, in the Warner case, the decisive factor was that the claimant prevailed because she was acting on her own behalf, effectively as her own lawyer, and was preparing the materials in anticipation of her own litigation.
Certainly, reading the platforms’ “Terms of Use” and using versions that protect outputs with greater confidentiality, even from administrators, can be excellent advice should one decide to use them in areas as sensitive as legal matters.
All Rights Reserved
Raffaella Aghemo, Lawyer