AI Esquire
Menu
Plans from $497/monthBuy Intake AI
Professional responsibility · AI confidentiality · Protective orders

The Confidentiality Breach Happened Before the AI Answered.

The document carried a header saying it was covered by a tribunal's protective order. A Justice Department lawyer uploaded it to a public artificial intelligence platform anyway. A later review found another protected document and Department files containing employees' personally identifiable information, including medical information. The Justice Department's Office of Professional Responsibility concluded that the lawyer knowingly and intentionally violated confidentiality duties and knowingly violated professional-conduct rules and the tribunal's order. OPR published the summary on November 17 without naming the lawyer, the platform, the matter, or the tribunal. Those omissions limit what anyone can responsibly infer about the underlying case. They do not obscure the central lesson. The most consequential event was not a hallucinated citation or a bad answer. It was the transfer. By the time the prompt was submitted, the professional-responsibility problem already existed.

OPR found three different failures in one upload pattern

OPR's summary is short, but its legal characterization is precise. A Department component reported the attorney's unapproved use of a public AI platform for government work after detecting the upload of a document whose header showed it was covered by a protective order. Subsequent review identified a second document under the same order and other documents containing Department employees' personal information, including medical information.

OPR separated the conduct into distinct findings. It concluded that the attorney recklessly violated professional-conduct rules by using the public platform contrary to Department policy. It concluded that the attorney knowingly and intentionally violated professional-conduct rules by uploading confidential Department documents. It further concluded that the attorney knowingly violated professional-conduct rules and the tribunal's protective order by uploading two documents known to be covered. OPR referred the findings to the Professional Misconduct Review Unit.

Those are OPR's findings, not a court's adjudication or a state disciplinary judgment. The office says its public summaries omit names and identifying information and may exclude material protected by court order, privilege, grand-jury secrecy, privacy law, or policy. The summary also does not reveal what the platform did with the files, whether any human accessed them, whether the material was retained, or whether it influenced later model training. None of those facts should be invented.

The limited record is still enough to reject a common framing. The risk did not depend on a false output. It did not depend on public reposting. OPR treated the upload to a third party as the disclosure. The system's response, accurate or otherwise, was beside that point.

Legal AI has an input problem and an output problem

The profession has spent years discussing hallucinations because they are visible. A fictional quotation appears in a filed brief. A nonexistent case is easy for a judge to expose. A second OPR summary published the same day illustrates that familiar failure: an Assistant U.S. Attorney filed a pleading with fabricated quotations and misstated holdings, then, according to OPR, made false and misleading representations about the use of AI when the court demanded an explanation.

That matter concerns output and candor. The protective-order matter concerns input and confidentiality. Putting them side by side reveals why a single instruction to verify AI work is inadequate. Verification asks whether the proposed answer is accurate, supported, and fit for use. It cannot retrieve information already disclosed to a system that was not authorized to receive it.

A careful lawyer therefore evaluates two transactions. First, what is being transmitted, to whom, under what terms, for what purpose, with what retention, access, and reuse? Second, what did the system produce, how was it verified, and how will it be used? The same tool can be acceptable for one transaction and unacceptable for the other. A public system might competently summarize a published opinion while remaining an improper destination for a sealed exhibit, client chronology, personnel record, or privileged interview memorandum.

Calling the software a drafting assistant does not change the information flow. The interface may feel like a blank page, but submitting a prompt sends data beyond the lawyer's immediate custody. For confidentiality analysis, that destination matters more than the conversational design.

A protective order governs transmission, not just publication

Protective orders are sometimes discussed as if they merely prevent a party from posting discovery on the internet. Their function is broader. Federal Rule of Civil Procedure 26(c), for example, authorizes courts to prescribe terms for disclosure or discovery, limit who may be present, require confidential handling, and protect trade secrets or other sensitive information. The actual order in the actual matter controls, and state, administrative, and criminal proceedings have their own authorities and practices.

That matter-specific text matters because an AI vendor may not fit the order's definition of an authorized person, contractor, expert, or service provider. A confidentiality designation may require acknowledgments, access controls, notice, return or destruction, or use only for the litigation. A lawyer cannot answer those questions by asking whether the platform is popular or whether its output will remain inside the firm.

Professional confidentiality is broader still. ABA Model Rule 1.6, where adopted or adapted, protects information relating to a representation and requires reasonable efforts against unauthorized disclosure or access. ABA Formal Opinion 512 applies that framework to generative AI and directs lawyers to understand a tool's policies and practices, including information security, retention, use of submitted information, and potential disclosure to others. Governing state rules and opinions may differ, and a government lawyer may also face agency policy and statutes governing official records.

The OPR summary's reference to employee medical information shows why labels should not be collapsed. A single upload may implicate a tribunal's order, professional rules, employer policy, privacy requirements, contractual terms, and duties to people who are not the lawyer's client. One source of authority does not cancel the others. Client consent, even when relevant, would not authorize violation of a court order or erase duties owed to third parties.

Tool approval is not document approval

Organizations often respond to AI risk by creating an approved-tools list. That is useful, but it can produce false confidence. Approval usually means a product passed some combination of security, procurement, privacy, and operational review for defined uses. It does not mean every user may upload every category of information from every matter.

A firm needs decisions at more than one level. The organization decides which systems may be used. The practice group decides which tasks are appropriate. The matter team identifies restrictions imposed by the client, court, regulator, contract, or data category. The lawyer then decides whether this particular document, in this form, may be transmitted for this purpose. Redaction or de-identification can change the analysis, but only if it actually removes the protected information and preserves the usefulness of the task.

Technical controls should reflect those distinctions. Blocking unapproved public platforms is different from limiting uploads to an approved enterprise environment. Data-loss prevention can flag medical information, confidentiality legends, and unusual transfers, but it will not understand every protective order. Matter workspaces, role-based access, retention settings, audit logs, and vendor restrictions help. None replaces a lawyer who reads the order and knows what the document contains.

The OPR matter was detected, which also raises the incident-response question. Once a questionable upload is found, the organization should preserve logs, identify the files and account, stop further transfer, determine retention and access, contact the provider when appropriate, and evaluate notice, court, client, employment, privacy, and disciplinary obligations. The sequence should be directed by the actual facts and governing law. Quiet deletion from a browser history is not remediation.

The clean output can still conceal the serious violation

The harder governance problem is that an input breach may leave excellent work product. The summary may be accurate. The chronology may save time. The draft may contain no invented authority. A reviewer focused only on quality can approve the result without seeing how it was obtained.

That makes confidentiality controls less visible and more important. Courts catch fake citations because the error reaches the docket. Clients and third parties may never learn that protected material reached an unauthorized system. The absence of a public embarrassment is not evidence that the workflow was sound.

Lawyers should resist turning the OPR summaries into a morality play about two careless government attorneys. The more useful question is whether ordinary systems make the same conduct easy. Can a lawyer open a public tool on a work device? Can protected text be pasted without a warning? Does the firm know which vendors retain prompts? Is responsibility divided so thinly among security, procurement, knowledge management, and ethics counsel that nobody owns the document-level decision? Individual judgment matters, but architecture influences which mistakes become one click away.

The profession's first AI lesson was to check the answer. That lesson remains necessary. It is no longer sufficient. A lawyer must also know what the system was asked to hold, who was permitted to receive it, and whether the act of asking violated a duty that no accurate answer could repair.

Sources and further reading

Primary and industry sources used to support this page. External guidance should be reviewed in context and for your jurisdiction.

  1. DOJ OPR, 2026 Investigative Summary 6Official November 17, 2026 summary of OPR findings concerning unapproved public AI use, confidential Department documents, employee personal and medical information, and two documents covered by a tribunal's protective order.
  2. DOJ OPR, 2026 Investigative Summary 5Official November 17, 2026 summary of separate findings involving fabricated quotations, misstated holdings, and false or misleading explanations to a court about AI use.
  3. ABA Formal Opinion 512National model guidance addressing competence, confidentiality, communication, candor, supervision, and fees when lawyers use generative AI. Governing jurisdictional rules may differ.
  4. ABA Model Rule 1.6Model confidentiality rule and comments addressing information relating to representation and reasonable efforts to prevent unauthorized disclosure or access.
  5. Federal Rule of Civil Procedure 26(c)Federal civil rule authorizing protective orders that prescribe disclosure terms and protect confidential, commercial, and other sensitive information. The order in each matter controls.
  6. DOJ, Overview of Privacy Act Disclosures to Third PartiesDepartment overview explaining the Privacy Act's general rule against agency disclosure of covered records without consent, subject to statutory exceptions.
Put the framework to work

See Intake AI handle your firm's real workflow.

Bring the intake questions, routing rules, or coverage gap you want to improve. We will demonstrate the system against them.

Call Intake AI now (941) 941-6967Book a 30-minute working session