In February 2026, Judge Jed Rakoff of the Southern District of New York ruled on what he called a question of first impression nationwide.
Bradley Heppner, facing federal securities and wire fraud charges, had used a publicly available AI platform to prepare 31 documents outlining his defense strategy, his legal exposure, and his anticipated arguments. He later shared those documents with his attorney. His counsel asserted privilege. The government moved to compel production. Judge Rakoff ruled from the bench: no privilege, no work product protection.
The reason wasn’t complicated. The platform’s terms of service permitted data collection, model training, and disclosure to third parties, including government authorities. By entering his communications into that platform, Heppner had voluntarily disclosed them to a third party. There was no reasonable expectation of confidentiality. Privilege requires confidentiality. The protection was gone.
The Harvard Law Review called it a ruling that “veers toward categorically excluding a client’s use of generative AI from attorney-client privilege.”
That’s the case. Now here’s why it matters to you.
Most biopharma legal departments are using AI tools right now. Some of those tools are enterprise platforms with contractual confidentiality protections. Many of them are not. And the line between the two is almost never clear to the attorney or paralegal who opens a browser tab and starts typing.
ChatGPT, Claude’s consumer interface, Gemini, Copilot without an enterprise license — these are Consumer AI Tools. Their terms of service, in most cases, permit the vendor to collect inputs, use them to improve models, and disclose them under certain circumstances. When a GC drafts a legal memo using one of these tools, enters client facts into a prompt, or uses the platform to analyze a contract — and the platform’s terms permit what Heppner’s platform permitted — the privilege analysis is the same.
The case didn’t create a new rule for AI. It applied traditional privilege doctrine to a new fact pattern. And the doctrine is clear: voluntary disclosure to a third party, without authorization, waives privilege.
The question for every biopharma GC right now is a simple one: do you know which AI tools your legal team is using, and do you know whether those tools have terms of service that could waive privilege over the work they’re helping produce?
Most don’t. Not because they’re careless, but because nobody told them this was a question they needed to answer.
The K&L Gates analysis of this line of cases, published in February 2026, gives practical guidance that I’d treat as the current standard of care:
Use enterprise platforms with terms of service that expressly prohibit the vendor from retaining inputs, using them for model training, or disclosing them to third parties. Treat AI-generated work product as a supervised assistant’s draft — produced at counsel’s direction, reviewed by counsel, incorporated into counsel’s professional judgment. Document that supervision. Prompts and outputs are potential work product. Treat them that way.
That’s not a complicated framework. It’s three things: use the right tools, supervise the outputs, document the process.
There are two other developments worth flagging alongside Heppner, because they’ll surface in your next board AI conversation whether you raise them or not.
The first is FDA’s January 2025 draft guidance on AI in regulatory decision-making. If your company uses AI in clinical trial management, safety signal detection, or regulatory submission preparation, FDA expects you to have a documented credibility framework — a risk-based assessment of what the AI is doing, how it’s validated for that use, and what level of human oversight is applied. That documentation standard is already in the draft. It will be in the final guidance.
The second is HHS’s proposed HIPAA Security Rule updates, published January 6, 2025. The proposal explicitly requires covered entities and business associates to include AI tools in their security risk analysis programs. If your company is a covered entity or works with one, and you’re using AI tools that touch patient data, those tools need to be in your risk assessment. OCR collected over $9.9 million in enforcement actions in 2024. The agency is active.
These aren’t future concerns. They’re current ones.
The three things I’d do this week if I were a GC at a PE-backed biopharma company:
First, inventory every AI tool your legal team is using. Not just the ones IT approved. All of them. The browser tabs people have open. The tools they started using because a colleague recommended them. You cannot govern what you haven’t counted.
Second, pull the terms of service for each tool and answer one question: does this vendor prohibit itself from using our inputs for model training and from disclosing our inputs to third parties? If the answer is no, or you can’t find the answer, that tool is a Consumer AI Tool for privilege purposes until proven otherwise.
Third, build a policy before your board asks for one. Not because a policy eliminates the risk, but because a board that asks “what’s our AI governance framework?” and gets a written answer is a different conversation than one that doesn’t. The GC who walks in with a framework — even a preliminary one — controls that conversation. The GC who doesn’t is explaining.
I’ve been building an AI governance policy suite for biopharma legal departments over the past several weeks. Five documents covering AI use policy, data classification, HIPAA and regulatory compliance, vendor assessment, and incident response — each grounded in the current regulatory and case law landscape. If you want a copy or want to talk through how it applies to your company’s situation, contact me via email (john@la
roccaadvisory.com).


