The court's ruling in United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 17, 2026), rests on a foundational factual assumption: that Anthropic's consumer-facing AI platform "uses such data to train Claude" and therefore a defendant who used Claude to prepare defense documents lacked a reasonable expectation of confidentiality. This paper demonstrates that the assumption is unproven and may be false as applied to the defendant. Anthropic's privacy policy changed materially during 2025, including the introduction of a mandatory opt-in/opt-out training toggle with a default ON setting. The court never established which version of the privacy policy governed Heppner's use, whether he opted in or out, or what retention and training posture applied to his account. This omission is not doctrinal—it is evidentiary. The ruling rests on a privacy posture that may not have applied to the defendant at all. This paper reconstructs the timeline, identifies the evidentiary void, and proposes a framework for courts evaluating confidentiality in AI-mediated communications.
Narnaiezzsshaa Truong (Fri,) studied this question.