Attorney admitted in NY here. It's fascinating that Judge Rakoff likely would have come to the opposite conclusion if the Claude chat was at the attorney's request or suggestion. I am surprised the court placed so much reliance on the Terms of Service, which are probably not so different than those of Outlook, Gmail, etc., say, yet nobody disputes that attorney-client emails remain privileged notwithstanding the Terms of Service of those providers. At least I have never seen anyone argue in NY that privilege is waived by emailing. And unlike sending an email to another person, chatting with Claude is a solo conversation more like organizing one's notes, which if in contemplation of obtaining legal advice seems privileged to me. I think this is a very close question and am not sure it would come out the same way in other courts or on even slightly different facts. Very interesting legal question.
Here's my question:
If the attorney-client privilege, and more importantly, the work product doctrine don't apply here, would they also not apply to direct conversations between an attorney and an AI?
It seems to me that the court would need to apply some twisted logic to claim that those protections apply to an attorney, but not to a petitioner or respondent.
Obviously this (along with the original unwritten order a few weeks ago) is causing a stir, but this decision isn't as weird as it sounds. The defendant's assertion was essentially a retroactive application of privilege: he didn't use Claude to draft documents at his attorney's request but instead used Claude effectively in lieu of an attorney and later provided the Claude-drafted materials to his attorney (heavily paraphrasing here). Privilege is not a bandage that closes self-inflicted wounds.
I have some concerns about some of the reasoning, namely the practical implications of referencing Claude's TOS in a world where public AI features are creeping into everything, but I expect some of the reasoning is based on this particular defendant likely being more sophisticated than an average person.
The overruling of both Shih and the standards laid out in NYSBA ethics opinions 820/842 (and various other state bar associations, and the fact that apparently no one tried to challenge those in court until AI) without real discussion of implications seems rather unusual; and that's a rather charitable reading to avoid the crazier "Claude is a person" framing
also, he quotes Gould v Mitsui: documents do not "acquire protection merely because they were transferred" to counsel; but that same case says they do acquire protection if communicated "for the purpose of obtaining or rendering legal advice"
How is this not effectively a ban on representing yourself in court? The lawyers and judge are going to be using AI. But the layman isn't allowed to use it?
people point out in sibling comments that is phone call then be out of client-attorney privileges? since it goes through a "3rd party"? maybe not the call itself but the voicemail for example. can it be "extracted" for the same purpose?
another point to make it safer would be sharing the "chat" with the lawyer, this way it becomes media of communication.
Communicating with an attorney is protected by privilege. In this case it seems they ruled that researching your case is not protected by privilege.
What about drafting communications with an attorney? Is a draft email that has not yet been sent protected? What about a Word doc containing a draft of an email? What about a Google search for “how do I spell amfeetamine?” that is part of your process of drafting your communication with your lawyer?
There is no way that this state of things survives long-term. Rationally, it's really no different than any other tool involved in production of your work product.
Not surprising that the legal profession decided that their special privilege doesn’t apply to you if you’re not paying them hundreds of dollars per hour
Heppner's argument was dumb but it opens a field of interesting questions. If I use a document processor (like Google Docs) to compose a message to my attorney, which message itself would be privileged, but I use some sidebar feature of Google Docs/Gemini to clean up a sentence that I thought was clunky, and elsewhere I have, for whatever reason, enabled features that permit Google to use inputs and outputs to train or refine their models, has that destroyed the privilege?
Letting this being decided by courts is like letting tobacco companies decide if smoking is addictive or not. It is obvious that they will always rule in their own interest.
I might have watched too many episodes of Better Call Saul but this makes a lot of sense for me, the logic of attorney client privilege is that the lawyer would not allow you to use him for performing crime. This fact gives a clear separation, if you talk to a friend or a chatbot about defence in court, there is not guarantee that you also talked about actively performing crime.
I highly recommend everyone actually read the opinion. It's such a thorough legal takedown of Heppner, you'll learn how the law works and why it doesn't apply to a lot of the made up cases in this thread:
TLDR:
- Claude told him IANAL
- Claude privacy policies say they "may disclose personal data to third parties in connection with claims, disputes, or litigation"
- Work product doctrine, does not apply in the same way to plaintiffs
- Lawyers did not direct him to use Claude (i.e. the laywers did not direct him to do research for the case using a specific tool)
My takeaway is that, as is, I should not do any work without a VPN or in plaintext. Everything else was up for grabs even before this case.
Questions this raises for me (making a note here to maybe research a bit later):
Does this analysis change if using on-site AI? What if the ToS is different? Is it possible to stand up a service that does get the protections required? This might also be interesting when dealing with trans-atlantic work.
It's not "no attorney-client privilege for AI chats" in general.
But a situation where the same would also apply if, instead of going to an chat bot, the person had gone to a random 3rd party non-attorney related person.
As in:
- the documents where not communication between the defendant and their attorney, but the defendant and the AI
- the AI is no attorney
- the attorney didn't instruct the defendant to use the AI / the court found the defendant did not communicate with the AI with the purpose of finding legal consule
- the communications with the AI (provider) where not confidential as a) it's a arbitrary 3rd party and b) they explicitly exclude usage for legal cases in their TOS
Still this isn't a nothing burger as some of the things the court pointed out can become highly problematic in other context. Like the insistence that attorney privilege is fundamentally build on a trusting human relationship, instead of a trusting relationship. Or that AI isn't just part of facilitating communication, like a spell checker, word program or voice mail box, legal book you look things up. All potentially 3rd parties all not by themself communication with a human but all part of facilitating the communication.
For criminal matters, the Govt will be so far into your phone and accounts, that unless you meet your lawyer in a faraday cage, you have no real attorney client privilege. Zoom attorney calls, cell phone calls, in person meetings with devices in the room... Using cloud connected systems to draft your legal defense... All of it is monitored and a form of parallel construction is done to use the information learned against you in ways that shape the case in the Govts favor before you even get to court. Just another use of the surveillance state.
"Judge Rakoff issued an oral ruling that neither the attorney-client privilege nor the work product doctrine protected the AI-generated documents.12 The decision rests on traditional principles of privilege.
The attorney-client privilege protects (1) communications, (2) among only privileged parties, (3) made for the purpose of providing or obtaining legal advice.13 Importantly, the protection of the attorney-client privilege is lost if the communication is shared outside of the privileged parties.14 The party claiming privilege has the burden of showing that confidentiality was maintained.15 Judge Rakoff stated that the attorneyclient privilege did not apply because the communications were shared with a thirdparty tool that did not maintain confidentiality.16
Second, Judge Rakoff held that the work product doctrine did not protect the documents.17 The work product doctrine protects (1) legal work product, (2) discussing legal strategy, (3) prepared by or at the direction of legal counsel, (4) in anticipation of litigation.18 Judge Rakoff rejected Heppners arguments that the work product doctrine could apply because the AI-generated reports did not reflect the legal strategy of Heppners legal counsel, although they contained theories generated by the client and Claude.19 Since neither Heppner nor the AI tool are legal counsel, and Heppner was not working at the direction of Heppners legal counsel, the materials were not protected by the work product doctrine. Judge Rakoff noted that the AI tools disclaimer that users have no expectation of confidentiality also undermined the work product doctrine claim.20
12 Transcript of Pretrial Conference at 6, United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb 10, 2026).
13 See United States v. Mejia, 655 F.3d 126, 132 (2d Cir. 2011).
14 See In re Six Grand Jury Witnesses, 979 F.2d 939, 943 (2d Cir. 1992).
15 See In re Grand Jury Subpoenas Dated Mar. 19, 2002 and Aug. 2, 2002, 318 F.3d 379, 384 (2d Cir. 2003).
16 Tr. at 3, Heppner, No. 25-cr-00503-JSR.
17 Id. at 6.
18 See In re Grand Jury Subpoenas, 318 F.3d at 383.
No attorney was involved. An AI tool is not a lawyer. It has no law license, owes no duty of loyalty, cannot form an attorney-client relationship, and is not bound by confidentiality obligations or professional responsibility rules. Discussing legal matters with an AI platform is legally no different from talking through your case with a friend.
2
Not for the purpose of obtaining legal advice. Anthropic's own public materials state that Claude follows the principle of choosing the "response that least gives the impression of giving specific legal advice." The tool explicitly disclaims providing legal services. You cannot claim you used a tool for legal advice when the tool itself says it does not provide it. Claude's terms were specifically highlighted by the government, which directly undermined the claim that Heppner was seeking legal advice from the tool.
3
Not confidential. This is the finding with the broadest implications. Anthropic's policy expressly states that user prompts and outputs may be disclosed to "governmental regulatory authorities" and used to train the AI model. Judge Rakoff found there was simply no reasonable expectation of confidentiality. As he put it, the tool "contains a provision that any information inputted is not confidential." This is not unique to Claude. OpenAI's privacy policy contains comparable provisions permitting data use for model training and disclosure in response to legal process.
And the distinction between free and paid plans matters less than many assume. Both Anthropic and OpenAI use conversations from free and individual paid plans (Claude Free, Pro, and Max; ChatGPT Free, Plus, and Pro) for model training by default. Users can opt out, but opting out of training does not eliminate the platforms' rights to disclose data to government authorities or in response to legal process. Only enterprise-tier agreements (ChatGPT Enterprise and Business; Claude's commercial and government plans) exclude user data from training by default and offer contractual confidentiality protections. A $20-per-month subscription does not buy you privilege.
4
Pre-existing documents cannot be retroactively cloaked in privilege. The AI-generated documents were created by Heppner before he transmitted them to counsel. Sending these unprivileged materials to his lawyers after the fact did not retroactively make them privileged.
Implications for waiver of privilege
Heppner fed information he had received from his attorneys into Claude. The government argued, and Judge Rakoff agreed, that sharing privileged communications with a third-party AI platform may constitute a waiver of the privilege over the original attorney-client communications themselves. The privilege belongs to the client, but so does the responsibility to maintain it."
"Privacy policies, including the one on Claude's website, openly inform users how their data is used. However, very few users actually read the fine print on these privacy policies, or even know these policies exist in the first place. It would probably surprise most people to learn that Claude's privacy policy explicitly gives its parent company, Anthropic, the right to disclose a user's data to third parties in connection with legal disputes and litigation."
This is a pretty terrible decision and inconsistent with all sorts of all other standards. If I did legal research in Google docs, it'd be covered. If I went to a legal library and took notes, it'd be covered, etc
139 comments
It seems to me that the court would need to apply some twisted logic to claim that those protections apply to an attorney, but not to a petitioner or respondent.
I have some concerns about some of the reasoning, namely the practical implications of referencing Claude's TOS in a world where public AI features are creeping into everything, but I expect some of the reasoning is based on this particular defendant likely being more sophisticated than an average person.
also, he quotes Gould v Mitsui: documents do not "acquire protection merely because they were transferred" to counsel; but that same case says they do acquire protection if communicated "for the purpose of obtaining or rendering legal advice"
another point to make it safer would be sharing the "chat" with the lawyer, this way it becomes media of communication.
What about drafting communications with an attorney? Is a draft email that has not yet been sent protected? What about a Word doc containing a draft of an email? What about a Google search for “how do I spell amfeetamine?” that is part of your process of drafting your communication with your lawyer?
FWIW not all cases have gone the same way, so there is likely to be a higher reckoning on this in multiple countries: https://fingfx.thomsonreuters.com/gfx/legaldocs/mypmyjwdzpr/...
https://news.ycombinator.com/item?id=47778308 AI ruling prompts warnings from US lawyers: Your chats could be used against you (reuters.com)
~3 hours ago, 43+ comments
https://news.ycombinator.com/item?id=47555642 Be careful: chatting with AI about your case is discoverable (harvardlawreview.org)
~18 days ago, 13 comments
Running your own LLM on your own hardware is how you can do this without getting hit with discovery.
And also, you want to run a LLM thats abliterated and larger. And if you connect to the internet, USE A VPN.
I wonder if anybody has gone all the way and made a darknet LLM service with no logs served only over TOR with XMR payments.
TLDR:
- Claude told him IANAL
- Claude privacy policies say they "may disclose personal data to third parties in connection with claims, disputes, or litigation"
- Work product doctrine, does not apply in the same way to plaintiffs
- Lawyers did not direct him to use Claude (i.e. the laywers did not direct him to do research for the case using a specific tool)
My takeaway is that, as is, I should not do any work without a VPN or in plaintext. Everything else was up for grabs even before this case.
Does this analysis change if using on-site AI? What if the ToS is different? Is it possible to stand up a service that does get the protections required? This might also be interesting when dealing with trans-atlantic work.
https://law.resource.org/pub/us/case/reporter/F2/296/296.F2d...
It's not "no attorney-client privilege for AI chats" in general.
But a situation where the same would also apply if, instead of going to an chat bot, the person had gone to a random 3rd party non-attorney related person.
As in:
- the documents where not communication between the defendant and their attorney, but the defendant and the AI
- the AI is no attorney
- the attorney didn't instruct the defendant to use the AI / the court found the defendant did not communicate with the AI with the purpose of finding legal consule
- the communications with the AI (provider) where not confidential as a) it's a arbitrary 3rd party and b) they explicitly exclude usage for legal cases in their TOS
Still this isn't a nothing burger as some of the things the court pointed out can become highly problematic in other context. Like the insistence that attorney privilege is fundamentally build on a trusting human relationship, instead of a trusting relationship. Or that AI isn't just part of facilitating communication, like a spell checker, word program or voice mail box, legal book you look things up. All potentially 3rd parties all not by themself communication with a human but all part of facilitating the communication.
The attorney-client privilege protects (1) communications, (2) among only privileged parties, (3) made for the purpose of providing or obtaining legal advice.13 Importantly, the protection of the attorney-client privilege is lost if the communication is shared outside of the privileged parties.14 The party claiming privilege has the burden of showing that confidentiality was maintained.15 Judge Rakoff stated that the attorneyclient privilege did not apply because the communications were shared with a thirdparty tool that did not maintain confidentiality.16
Second, Judge Rakoff held that the work product doctrine did not protect the documents.17 The work product doctrine protects (1) legal work product, (2) discussing legal strategy, (3) prepared by or at the direction of legal counsel, (4) in anticipation of litigation.18 Judge Rakoff rejected Heppners arguments that the work product doctrine could apply because the AI-generated reports did not reflect the legal strategy of Heppners legal counsel, although they contained theories generated by the client and Claude.19 Since neither Heppner nor the AI tool are legal counsel, and Heppner was not working at the direction of Heppners legal counsel, the materials were not protected by the work product doctrine. Judge Rakoff noted that the AI tools disclaimer that users have no expectation of confidentiality also undermined the work product doctrine claim.20
12 Transcript of Pretrial Conference at 6, United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb 10, 2026).
13 See United States v. Mejia, 655 F.3d 126, 132 (2d Cir. 2011).
14 See In re Six Grand Jury Witnesses, 979 F.2d 939, 943 (2d Cir. 1992).
15 See In re Grand Jury Subpoenas Dated Mar. 19, 2002 and Aug. 2, 2002, 318 F.3d 379, 384 (2d Cir. 2003).
16 Tr. at 3, Heppner, No. 25-cr-00503-JSR.
17 Id. at 6.
18 See In re Grand Jury Subpoenas, 318 F.3d at 383.
19 Tr. at 5, Heppner, No. 25-cr-00503-JSR.
20 Id. at 6."
https://www.debevoise.com/-/media/files/insights/publication...
"Reasons Privilege Failed
1
No attorney was involved. An AI tool is not a lawyer. It has no law license, owes no duty of loyalty, cannot form an attorney-client relationship, and is not bound by confidentiality obligations or professional responsibility rules. Discussing legal matters with an AI platform is legally no different from talking through your case with a friend.
2
Not for the purpose of obtaining legal advice. Anthropic's own public materials state that Claude follows the principle of choosing the "response that least gives the impression of giving specific legal advice." The tool explicitly disclaims providing legal services. You cannot claim you used a tool for legal advice when the tool itself says it does not provide it. Claude's terms were specifically highlighted by the government, which directly undermined the claim that Heppner was seeking legal advice from the tool.
3
Not confidential. This is the finding with the broadest implications. Anthropic's policy expressly states that user prompts and outputs may be disclosed to "governmental regulatory authorities" and used to train the AI model. Judge Rakoff found there was simply no reasonable expectation of confidentiality. As he put it, the tool "contains a provision that any information inputted is not confidential." This is not unique to Claude. OpenAI's privacy policy contains comparable provisions permitting data use for model training and disclosure in response to legal process.
And the distinction between free and paid plans matters less than many assume. Both Anthropic and OpenAI use conversations from free and individual paid plans (Claude Free, Pro, and Max; ChatGPT Free, Plus, and Pro) for model training by default. Users can opt out, but opting out of training does not eliminate the platforms' rights to disclose data to government authorities or in response to legal process. Only enterprise-tier agreements (ChatGPT Enterprise and Business; Claude's commercial and government plans) exclude user data from training by default and offer contractual confidentiality protections. A $20-per-month subscription does not buy you privilege.
4
Pre-existing documents cannot be retroactively cloaked in privilege. The AI-generated documents were created by Heppner before he transmitted them to counsel. Sending these unprivileged materials to his lawyers after the fact did not retroactively make them privileged.
Implications for waiver of privilege
Heppner fed information he had received from his attorneys into Claude. The government argued, and Judge Rakoff agreed, that sharing privileged communications with a third-party AI platform may constitute a waiver of the privilege over the original attorney-client communications themselves. The privilege belongs to the client, but so does the responsibility to maintain it."
https://natlawreview.com/article/your-ai-conversations-are-n...
"Privacy policies, including the one on Claude's website, openly inform users how their data is used. However, very few users actually read the fine print on these privacy policies, or even know these policies exist in the first place. It would probably surprise most people to learn that Claude's privacy policy explicitly gives its parent company, Anthropic, the right to disclose a user's data to third parties in connection with legal disputes and litigation."
https://nysba.org/loose-ai-prompts-sink-ships-how-heppner-sh...