Privacy & security

What Heppner, Warner, and Morgan Held About AI, Privilege, and Work Product

Bells Up AI gives solo and small-firm lawyers access to the latest AI models under provider agreements that prohibit training on their input and restrict disclosure. Three recent federal rulings show why those terms matter.

In Heppner, a criminal defendant used Claude without his lawyers’ direction. The court held that Anthropic’s consumer terms, which allowed training, collection, and third-party disclosure, destroyed any claim of privilege. In Warner and Morgan, two courts held that AI could be used to handle litigation documents without waiving work product protection. The Morgan court also specified the terms a lawyer’s contract with an AI provider must contain before confidential material covered by a protective order may be entered into the tool.

These rulings demonstrate why legal professionals need to understand how using AI without specific contractual protections can result in waivers of privilege or other confidentiality breaches.

What Heppner held

United States v. Heppner, No. 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 17, 2026).

Bradley Heppner, a former chief executive under federal investigation for securities fraud, used Claude to prepare about thirty-one documents outlining defense strategy. The court noted that his lawyers “did not direct [Heppner] to run Claude searches.” He shared the documents with his lawyers afterward. FBI agents seized copies of these documents, and Judge Rakoff rejected the defendant’s claims that the chats were privileged or subject to work product protection.

Specifically, Judge Rakoff held that because Claude is not an attorney and disclaims giving legal advice, Heppner’s conversations with Claude could not be communications with counsel made to obtain legal advice. Judge Rakoff further held that handing the documents to his lawyers afterward did not change that analysis, because “non-privileged communications are not somehow alchemically changed into privileged ones upon being shared with counsel.” On work product, Judge Rakoff held that the documents were not protected because they were “prepared by the defendant on his own volition” and not at counsel’s direction.

Regarding confidentiality, Judge Rakoff quoted provisions from Anthropic’s consumer privacy policy, which governed Heppner’s account, writing:

“[T]he written privacy policy to which users of Claude consent provides that Anthropic collects data on both users’ ‘inputs’ and Claude’s ‘outputs,’ that it uses such data to ‘train’ Claude, and that Anthropic reserves the right to disclose such data to a host of ‘third parties,’ including ‘governmental regulatory authorities.’”

Judge Rakoff added that the policy put users on notice that Anthropic, “even in the absence of a subpoena compelling it to do so, may ‘disclose personal data to third parties in connection with claims, disputes[,] or litigation.’” In a footnote, he said that even if some of what Heppner typed had been privileged, he waived the privilege by sharing it with Claude and Anthropic, and in light of the policy he “had no reasonable expectation that the inputs would not be shared with other third parties.”

Judge Rakoff also said what might have changed the result. “Had counsel directed Heppner to use Claude, Claude might arguably be said to have functioned in a manner akin to a highly trained professional who may act as a lawyer’s agent within the protection of the attorney-client privilege.”

Warner and Morgan reached a different result on work product

Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich. Feb. 10, 2026).

A pro se plaintiff in an employment case had used ChatGPT to prepare her case, and her former employer moved to compel “all documents and information concerning her use of third-party AI tools in connection with this lawsuit.” Magistrate Judge Anthony Patti held that the material was work product. He reasoned that generative AI programs are “tools, not persons,” and that work product protection is waived only by disclosure to an adversary or in a way likely to reach one. The court did not examine ChatGPT’s terms.

Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB (D. Colo. Mar. 30, 2026).

Magistrate Judge Maritza Dominguez Braswell followed Warner in another pro se employment case. She held that a litigant’s use of AI to prepare a case “closely resembles the kind of confidential, strategy-laden iterative work product that Rule 26(b)(3) was designed to protect.” She distinguished Heppner as a criminal case in which a represented defendant acted apart from his lawyer.

After deciding the work-product issue, the court addressed a separate question: when material covered by the protective order could be entered into an AI platform. The amended order barred entry unless the AI provider is contractually prohibited from:

“(1) storing or using inputs to train or improve its model; and (2) disclosing inputs to any third party except where such disclosure is essential to facilitating delivery of the service... In addition, the AI provider must contractually afford the party or authorized recipient the ability to remove or delete all CONFIDENTIAL information upon request.”

The court connected those requirements to the resource gap between large firms and other litigants. “[A]s large firms pour thousands of dollars into enterprise-grade AI and make their use of AI more secure, efficient, effective, and powerful, how will a pro se litigant or a litigant who cannot afford big-ticket legal services and better AI keep up?” Secure provider terms should not require a large-firm technology budget.

What remains unresolved

Warner and Morgan are work-product rulings by magistrate judges in pro se employment cases. Neither addressed privilege, and neither examined whether a represented lawyer’s own use of AI raises different questions. Heppner is a privilege ruling in a criminal case about a represented defendant who acted without his lawyers’ involvement. These three courts addressed different doctrines in different procedural contexts.

How Bells Up AI meets these requirements

Under Heppner’s reasoning, and consistent with the Morgan protective order, the provisions that matter are whether the AI provider may train on the user’s prompts and responses and whether it may disclose them to others. All of Bells Up’s provider agreements prohibit training on your prompts and responses and restrict their disclosure. For seven of our ten model families, the model developer never receives your prompts and responses, and the infrastructure provider (Amazon Bedrock or Fireworks) does not retain them.

Eight of our ten model families operate under zero data retention by the inference provider. The two exceptions are OpenAI (up to 30 days for abuse monitoring) and Google (up to 55 days). A firm can disable either family for all its users. Our privacy reference page identifies the terms applicable to each provider.

The following tables show how we believe Bells Up AI meets the concerns the Heppner court identified and the requirements of the Morgan protective order. (No court opinion has analyzed how these standards might apply to Bells Up AI or a platform like it.)

What the Heppner court found problematic

Anthropic consumer term the court cited Bells Up AI Our infrastructure providers Our model providers
Anthropic collects data on users’ inputs and outputs. Bells Up stores conversation data only when the user chooses to save. Unsaved prompts and responses are not retained; attached documents are held encrypted and deleted within 24 hours of inactivity. AWS stores saved conversation data under Bells Up’s encryption. Fireworks does not store conversation data. Eight of ten families operate under zero data retention: they do not store your prompts or responses. OpenAI retains for up to 30 days; Google for up to 55 days. Firm administrators can remove these models from the Bells Up interface for all users.
Anthropic uses collected data to train Claude. Bells Up AI does not train on user input. AWS does not train on customer data (AWS Data Processing Addendum). Fireworks is prohibited from training on customer data (Fireworks Data Processing Addendum). All ten model families are contractually prohibited from training on your prompts and responses.
Anthropic may disclose data to third parties, including governmental regulatory authorities. Bells Up will not voluntarily disclose conversation content to law enforcement or any government entity (Privacy Policy). AWS’s Data Processing Addendum restricts disclosure to providing the service or complying with law. Fireworks’s terms restrict disclosure to providing the service or complying with law. Eight of ten families retain nothing and have nothing to disclose. OpenAI restricts access to authorized employees and confidentiality-bound contractors (Services Agreement). Google restricts access to authorized employees (Data Processing Addendum).
Anthropic may disclose personal data in connection with claims, disputes, or litigation without a subpoena. Bells Up will produce content only in response to valid legal process, will challenge overbroad demands, and will notify the affected user (Privacy Policy). AWS’s terms restrict disclosure to what law requires. Fireworks’s terms restrict disclosure to what law requires. Eight of ten families retain nothing. OpenAI and Google restrict access during their retention periods. Firm administrators can remove these models from the Bells Up interface for all users at the firm.
The user had no reasonable expectation that inputs would not be shared. Users can reasonably expect confidentiality because Bells Up’s Terms of Service commit to treating conversations and documents as confidential and restrict disclosure to what is necessary to provide the service. AWS and Fireworks are each bound by data processing agreements that restrict use and disclosure of customer data. All ten families are under contracts that prohibit training and restrict disclosure. Eight of ten retain nothing.

What the Morgan protective order requires

Requirement Bells Up AI Our infrastructure providers Our model providers
The AI provider must be contractually prohibited from storing or using inputs to train or improve its model. Bells Up AI does not train on user prompts and responses (Terms of Service). AWS: inputs and outputs are not used to train models (AWS Bedrock Data Protection). Fireworks: prohibited from training on covered data (Fireworks Data Processing Addendum). All ten families are contractually prohibited. Anthropic: Commercial Terms §B. OpenAI: Services Agreement §4.2. Google: Gemini API Paid Services Terms. xAI: Enterprise Terms §3.3. Bedrock-routed developers never receive the input.
The AI provider must be contractually prohibited from disclosing inputs to any third party except where essential to delivering the service. Bells Up will not voluntarily disclose conversation content. Terms of Service restrict disclosure to what is necessary to provide, support, and secure the service or comply with law. AWS: Data Processing Addendum restricts disclosure to providing the service or complying with law. Fireworks: terms restrict disclosure to providing the service or complying with law; content is not retained beyond request duration. Eight of ten families retain nothing. OpenAI restricts access to authorized employees and confidentiality-bound contractors (Services Agreement). Google restricts access to authorized employees (Data Processing Addendum). Firm administrators can remove these models from the Bells Up interface for all users.
The AI provider must contractually afford the ability to remove or delete all confidential information upon request. Users can delete saved conversations immediately; deletion destroys the encryption key. Encrypted backup copies expire within approximately 30 days. Unsaved prompts and responses are not stored; attached documents are held encrypted and deleted within 24 hours of inactivity. AWS stores only encrypted data under Bells Up’s key management; deletion removes the decryption key. Fireworks retains nothing to delete. Eight of ten families retain nothing to delete. OpenAI retains abuse-monitoring logs for up to 30 days; Google for up to 55 days. Those copies expire on schedule. Firm administrators can remove these models from the Bells Up interface to prevent their use entirely.

For provider-by-provider contract terms and retention periods, see our privacy reference page.