On February 10, 2026, Judge Jed Rakoff of the Southern District of New York issued an opinion in United States v. Heppner that should send a chill through every law firm in the country. The ruling is straightforward, and that is what makes it so disruptive: documents generated by an AI system and subsequently shared with counsel are not protected by attorney-client privilege. The AI is not the client. The AI is not the lawyer. And the communication between a machine and a human, Judge Rakoff wrote, does not satisfy the foundational requirements of the privilege.
The immediate reaction from legal Twitter was predictable. Some called it obvious. Others called it catastrophic. Both camps are right, in their own way. The ruling is doctrinally sound. And it will fundamentally change how lawyers use AI in litigation.
What Happened in Heppner
The facts are instructive. The defendant's legal team used an AI system to analyze a large volume of internal corporate documents. The AI generated summaries, flagged potentially responsive documents, and produced analytical memoranda identifying legal risks. These AI-generated outputs were shared with outside counsel, who used them to develop litigation strategy. When the government sought production of the AI-generated documents in discovery, the defense asserted attorney-client privilege and work product protection.
Judge Rakoff rejected both arguments. On privilege, his reasoning was precise. Attorney-client privilege protects confidential communications between a client and their attorney, made for the purpose of obtaining legal advice. The AI system is neither the client nor the attorney. Its outputs are not "communications" in the sense the privilege contemplates. They are machine-generated analyses. The fact that these analyses were later shared with counsel does not retroactively clothe them in privilege, any more than sharing a publicly available research report with your lawyer would make that report privileged.
On work product, the analysis was slightly more nuanced but reached the same result. Work product protection applies to documents prepared in anticipation of litigation by or for a party. The AI-generated documents were prepared by a machine, not by counsel or at counsel's direction in the traditional sense. Judge Rakoff acknowledged that the legal team configured the AI's prompts and parameters, but concluded that the AI's outputs reflected the machine's analysis, not counsel's mental impressions or legal theories.
The privilege exists to protect the human relationship between attorney and client. When a machine interposes itself into that relationship, the privilege does not automatically extend to the machine's outputs.
The Implications for Litigation Holds and Discovery
The practical impact is enormous. Consider the standard workflow at a large firm handling complex litigation. The litigation team receives a massive document production. They use AI tools to review, categorize, and summarize the documents. The AI generates analytical outputs: risk assessments, privilege logs, issue summaries, strategic memoranda. These outputs are shared among the legal team and used to develop case strategy.
After Heppner, every one of those AI-generated outputs is potentially discoverable. The opposing party can argue, with the weight of Judge Rakoff's opinion behind them, that these documents are not privileged communications and not attorney work product. They are machine outputs that happen to have been shared with lawyers.
This has immediate implications for litigation holds. Organizations must now consider whether AI-generated documents fall within the scope of their preservation obligations. If an AI system generates a risk assessment that identifies potential legal exposure, that document may need to be preserved and produced in discovery. The instinct to treat AI outputs as internal, privileged analysis is, after Heppner, a dangerous instinct.
It also raises questions about document review workflows. Many firms now use AI as a first-pass reviewer, with the AI's categorizations and summaries serving as the foundation for human review. If the AI's work product is discoverable, opposing counsel can see not just the documents themselves but the AI's analysis of those documents, including its assessment of which documents are most damaging and why.
What Lawyers Need to Change Right Now
Segregate AI outputs from privileged communications. Do not comingle AI-generated analyses with attorney-client communications. If you use AI to review documents, keep the AI's outputs in a separate system from your privileged case files. When attorneys develop strategy based on AI analysis, the attorney's own memoranda reflecting their legal judgment should be documented separately from the AI's outputs.
Treat AI as you would treat a non-lawyer consultant. The analogy is useful. If you hire a forensic accountant to analyze financial records, the accountant's report is not privileged merely because it was prepared for litigation. It may be protected as work product if prepared at counsel's direction in anticipation of litigation, but even then, the factual portions are discoverable. Apply the same logic to AI outputs.
Implement prompt hygiene. The prompts you give to AI systems are themselves potentially discoverable documents. If your prompts reveal litigation strategy, legal theories, or privileged information, you have created a discoverable record of your strategic thinking. Design prompts that are functional without being strategically revealing.
Review your AI vendor agreements. Many AI tools process data on remote servers. If your client's confidential information is being processed by a third-party AI system, you may have a privilege waiver problem independent of Heppner. Ensure your vendor agreements include confidentiality provisions adequate to preserve whatever privilege arguments remain available.
Update your litigation hold procedures. Your litigation hold notice should now explicitly address AI-generated documents. Custodians need to understand that AI outputs, including summaries, analyses, and categorizations, must be preserved alongside the underlying documents.
The FRE 707 Connection
The Heppner ruling intersects with the proposed Federal Rule of Evidence 707 in an important way. FRE 707 addresses the admissibility of machine-generated evidence, requiring proponents to demonstrate system reliability, appropriate methodology, and known error rates. After Heppner, AI-generated documents that were previously shielded by privilege claims may now enter evidence, where they will face scrutiny under the FRE 707 framework.
This creates a two-stage problem for litigants who relied on AI. First, the AI's outputs are discoverable because they are not privileged. Second, when the opposing party seeks to introduce those outputs as evidence, the party that generated them may need to defend the AI system's reliability under FRE 707, even though they did not intend for those outputs to be used as evidence at all.
Imagine the scenario: your AI document review system flagged a particular email as "highly relevant to potential securities fraud liability." That assessment, which you generated for internal strategic purposes, is now produced in discovery and offered by the opposing party as evidence that your client recognized its fraud exposure. You now find yourself arguing that your own AI system's analysis is unreliable, which undermines your broader use of the system, or conceding that the analysis is reliable, which supports the opposing party's case.
This is not a hypothetical trap. It is the predictable consequence of Heppner combined with the FRE 707 framework. Lawyers must anticipate it now.
The Broader Trajectory
Judge Rakoff's opinion is a district court ruling, not binding precedent outside the Southern District of New York. Other courts may reach different conclusions, particularly on the work product question, where the analysis is closer. But the opinion is from one of the most respected federal judges in the country, and its reasoning is difficult to distinguish. The core insight, that an AI system is not a party to the attorney-client relationship and its outputs do not automatically inherit the privilege that attaches to that relationship, is likely to be broadly persuasive.
The legal profession's adoption of AI has been rapid and enthusiastic. Heppner is the first major judicial check on that adoption. It does not prohibit the use of AI in legal practice. It does not even discourage it. What it does is clarify that the rules of professional responsibility and evidence apply to AI-assisted legal work just as they apply to every other form of legal work. AI does not create a new category of privilege. It does not automatically generate work product. And lawyers who assume otherwise are exposing their clients to significant risk.
The firms that adapt quickly will have a competitive advantage. They will develop AI workflows that are both effective and litigation-proof, that leverage AI's analytical capabilities without creating discoverable records of strategic thinking. The firms that ignore Heppner will learn its lessons the hard way, in the discovery phase of their next major case.
The Criterion AI provides expert witness services and litigation support for matters involving artificial intelligence, machine learning, and algorithmic decision-making. For a confidential consultation on an active or anticipated matter, contact us at info@thecriterionai.com or call (617) 798-9715.