United States v. Heppner, No. 25-CR-00503-JSR (S.D.N.Y.)

In an address to the Australian Bar Association on 29 August 2025, Bell CJ warned about the lack of adequate safeguards to preserve legal professional privilege when information is submitted to open source Gen AI chatbots.

There have not been many (if any) cases squarely addressing this point. Last week, the Southern District of New York confronted it.

In United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y.), defence lawyers claimed privilege over 31 documents characterised as AI-generated analyses created to obtain legal advice. The material had been generated using “Claude”, an open source consumer AI tool developed by Anthropic.

Judge Jed S. Rakoff described the issue as a nationwide matter of first impression: are communications with a public AI platform during a criminal investigation protected by privilege or work product?

Answer: No, because:

(1) “Claude” is not a lawyer;

(2) the platform’s privacy policy permitted use and disclosure of inputs and outputs to third parties (i.e. it was open source) - there was therefore no expectation of confidentiality or privacy;

(3) the material was not created at counsel’s direction for the purpose of legal advice;

(4) Work product failed for the same reason.A link to the written opinion of Judge Rakoff appears below.

For Australian practitioners using open-source AI tools, the message may be equally blunt: if you submit privileged information into an open source AI tool, you may be putting legal professional privilege at risk.

Privilege is not generated by a prompt — but it may be lost with one.