The Legal AI 'Privilege Log' Trap: Why AI-Assisted e-Discovery Is Generating Privilege Assertions That Won't Survive a Motion to Compel

The privilege log has always been litigation's most tedious artifact. Thousands of entries, each requiring a specific factual predicate, each potentially the subject of a motion to compel. So when AI-assisted review tools promised to automate the process—flagging privileged documents, populating log fields, generating descriptions—litigation teams embraced the efficiency. That was a mistake many are now paying for in courtrooms.

Courts are pushing back. Hard. And the problem isn't that AI is reviewing documents for privilege. The problem is that AI pattern-matching and the actual legal standard for privilege are doing fundamentally different things, and nobody has adequately bridged that gap.

What AI Tools Are Actually Doing

Most AI-assisted privilege review tools operate on a combination of custodian identification, domain filtering, and keyword or semantic pattern recognition. The system flags documents that contain lawyer names, legal domain addresses, or phrases associated with legal advice. Some more sophisticated tools use large language models to generate log descriptions.

What these tools are not doing is applying In re Grand Jury Subpoena analysis. They are not asking whether the primary purpose of a communication was to obtain or provide legal advice, as required under the dominant purpose test adopted in In re Kellogg Brown & Root, Inc., 756 F.3d 754 (D.C. Cir. 2014). They are not distinguishing between a lawyer acting as business advisor and a lawyer acting as legal counsel—the exact distinction that United States v. Deloitte LLP, 610 F.3d 129 (D.C. Cir. 2010), requires you to draw. They are matching patterns and generating assertions. Those are not the same thing.

The result is privilege logs that describe every email with a lawyer's email address as "confidential communication reflecting legal advice," regardless of whether the lawyer is weighing in on a regulatory question or scheduling a client dinner. Courts have seen this before from humans. They're increasingly seeing it from machines, at scale.

What's Happening in Motion Practice

Opposing counsel files a motion to compel. They pick ten entries off the log. Half of them are CC'd communications where the lawyer received a business update. Two are emails where in-house counsel is explicitly described in the log as providing "legal advice" but the document itself is a PowerPoint on market strategy. One is a draft contract with no attorney annotations.

You now have a problem. Not just on those ten documents, but on your entire log's credibility.

In Chevron Corp. v. Donziger, Judge Kaplan's sustained scrutiny of privilege assertions illustrated how courts respond to overbroad, formulaic logs: document-by-document in camera review, adverse inferences about the quality of review, and fee-shifting. More recently, courts applying the District of Delaware's Default Standard for Discovery have increasingly demanded that privilege logs include specific factual descriptions sufficient to allow the court to assess the claimed privilege without looking at the document itself. A log entry that reads "Email re: legal advice regarding contract" fails that standard. It fails it whether a paralegal wrote it or a language model did.

The automation problem is worse because AI-generated descriptions tend toward uniformity. When a human reviewer writes five hundred log entries, there's natural variation in how they describe documents. When an AI generates five thousand, the descriptions converge on the same half-dozen templates. That uniformity is itself a signal to experienced opposing counsel that the log was machine-generated and inadequately reviewed.

Why Large Document Sets Make This Worse

The economic logic of AI-assisted review depends on scale. The larger the document set, the greater the efficiency gain. But large document sets are also where the worst privilege assertion problems cluster.

In a bet-the-company case with two million documents, privilege review often occurs under extreme time pressure, with AI handling first-pass identification and humans reviewing only documents the system flags with high confidence. Documents in the middle confidence range—the ones where a lawyer might be providing business advice that touches on legal risk—get underreviewed. The AI's training data pushes toward false positives because the cost of waiver, in theory, exceeds the cost of over-assertion. In practice, courts disagree with that risk calculus.

The work product doctrine creates an additional layer of complexity that AI handles badly. The doctrine protects documents prepared in anticipation of litigation, but the "anticipation of litigation" inquiry is fact-specific and contested. Adlman, the leading Second Circuit case on the issue, requires analysis of whether litigation was a primary or secondary motivation for the document's creation. AI tools routinely stamp documents as work product because they were created by lawyers during a period when litigation was ongoing. That's not the standard.

What a Defensible Workflow Actually Requires

If you're using AI-assisted tools for privilege review, the workflow has to treat AI output as a first-pass triage mechanism, not a finished product. That means several things concretely.

Attorney review of a statistically valid sample of AI-flagged documents before finalizing the log is not optional. You need to test false positive rates against actual legal standards, not just against the vendor's confidence scores.

Log descriptions must be written—or substantially revised—by humans who can apply the primary purpose test. "Email from in-house counsel providing legal advice regarding pending regulatory inquiry" is a defensible entry. "Communication re: legal matters" is not.

Custodian-context matters and must be documented. A log entry for an in-house lawyer who splits time between legal and business functions needs more factual support than one for outside litigation counsel.

Finally, document your review process itself. Courts are starting to ask. In the event of a motion to compel, you want to be able to demonstrate a defensible methodology, not just point to a vendor's marketing materials about accuracy rates.

The Bill Is Coming Due

The efficiency gains from AI-assisted privilege review are real. The legal risk from deploying those tools without adequate attorney oversight is also real, and it's materializing in motion practice right now. A privilege log is not a document you can afford to lose credibility on. Every entry you cannot defend is a potential waiver, and in complex litigation, a string of waivers can compromise your entire privilege posture.

Use the tools. But own the log. Courts are not going to accept "the AI flagged it" as a response to a targeted motion to compel, and neither should you.