Document Review With AI and the Privilege Log Question

Document Review With AI and the Privilege Log Question

AI-assisted document review has been defensible in federal court since at least 2012, when Judge Andrew Peck blessed predictive coding in Da Silva Moore v. Publicis Groupe. A decade later, courts routinely accept technology-assisted review (TAR) as reasonable, and in some cases have found it superior to manual review. But there is a specific problem that keeps coming up in practice, and it does not get nearly enough attention: what happens when AI touches privileged documents, and how do you defend the resulting privilege log?

The Baseline: What Courts Actually Require

Federal Rule of Civil Procedure 26(b)(5)(A) is straightforward. If you withhold documents on privilege grounds, you have to describe them in a way that lets the other side assess the claim "without revealing information itself privileged." That means your privilege log needs to identify the document, the privilege asserted, and enough factual basis to support the claim. Most local rules add specificity requirements on top of that.

The practical problem is volume. In large commercial litigation, review populations of 500,000 to several million documents are common. A 2023 RAND Institute study on civil discovery costs found that document review accounts for roughly 70% of total e-discovery spending, which can push into eight figures on major matters. When you are dealing with that kind of scale, manual privilege review of every document is economically brutal and, frankly, error-prone. Reviewers get tired. Consistency degrades. Privileged documents slip through.

Where AI Fits in the Privilege Workflow

Most teams using AI for privilege review are doing one of two things, sometimes both. First, they use TAR or large language models to classify documents as potentially privileged, creating a priority queue for human reviewers. Second, they use AI to generate draft privilege log entries, pulling out the key metadata and descriptive fields that Rule 26(b)(5) requires.

Both of these are defensible, but the details matter considerably.

On classification, the key case is still Winfield v. City of New York (S.D.N.Y. 2017), where the court accepted TAR for privilege review but emphasized that human attorneys had to make the final privilege designations. The AI narrows the field; the lawyer makes the call. More recently, Judge Rochon's standing order in the D.C. District (updated January 2024) explicitly permits AI-assisted review but requires disclosure of the methodology and validation metrics.

On privilege log generation, the risk is different. If your AI drafts log entries that are vague or formulaic, opposing counsel will move to compel, and courts are not sympathetic. In Chevron Corp. v. Donziger, the Southern District of New York found boilerplate privilege log descriptions inadequate and ordered in camera review of thousands of documents. That was a manual review case, but the principle applies with even more force when the descriptions are machine-generated. A judge who sees identical phrasing across 4,000 log entries will reasonably question whether anyone actually reviewed the underlying documents.

Defending the Methodology

If opposing counsel challenges your AI-assisted privilege review, you need to be ready with three things.

1. A documented protocol

This should exist before review begins, not after a motion to compel lands on your desk. The protocol should specify which tools you used, how the training set was constructed, what validation steps were taken, and where human review intersects with machine classification. The Sedona Conference's Commentary on Achieving Quality in the E-Discovery Process (2023 update) recommends documenting recall and precision metrics for any TAR workflow. If you can show 90%+ recall with reasonable precision on your privilege classifications, you are in strong shape.

2. Quality control sampling

Courts want to see that you checked the AI's work. Random sampling of both the privileged and non-privileged populations is standard practice. In In re Broiler Chicken Antitrust Litigation (N.D. Ill. 2018), the court found a TAR protocol reasonable in part because the producing party had conducted statistically valid quality control sampling at a 95% confidence level. For privilege specifically, you should be sampling documents the AI flagged as non-privileged to confirm that privileged material is not leaking into production. A missed privilege call is a potential waiver, and while Federal Rule of Evidence 502(b) provides some protection for inadvertent disclosures, relying on it as a safety net is not a strategy.

3. Transparency with the court

This is where many teams stumble. There is a temptation to be vague about AI involvement, either because the technology feels novel or because you do not want to invite scrutiny. That instinct is wrong. Courts have consistently rewarded transparency. In Livingston v. City of Chicago (N.D. Ill. 2022), the court praised the producing party for proactively disclosing its TAR methodology and denied a motion to compel additional review. Conversely, in cases where parties have been evasive about their review methods, courts have ordered do-overs at the producing party's expense.

The FRE 502(d) Order: Get One

If you are using AI anywhere in your privilege workflow, you should be seeking a Federal Rule of Evidence 502(d) order at the outset of the case. A 502(d) order provides that disclosure of privileged material in the litigation does not constitute waiver in any other federal or state proceeding. This is not a substitute for careful review, but it is essential risk mitigation. The order is non-controversial; most courts grant them on stipulation. Yet a surprising number of litigators still do not request them. In a 2022 survey by Exterro, only 41% of respondents reported routinely seeking 502(d) orders in cases involving large-scale document review. If you are running AI-assisted review without one, you are taking on unnecessary exposure.

The Clawback Protocol

Related but distinct from the 502(d) order is your clawback agreement under FRCP 26(b)(5)(B). When AI is involved in the review, your clawback protocol should specifically address the scenario where machine classification errors result in production of privileged material. The protocol should include timelines for notification, procedures for sequestration, and agreement on the treatment of any copies. Draft this with specificity. A generic clawback clause will not help you when opposing counsel has already forwarded a privileged email to their expert.

Practical Considerations for Privilege Logs at Scale

A few things worth noting for teams building these workflows:

• Attorney-client vs. work product classifications require different training. The linguistic markers for attorney-client privilege (communications with counsel, legal advice sought or rendered) differ meaningfully from work product indicators (litigation anticipation, mental impressions). Training a single model to handle both without distinction will degrade accuracy.
• Email threading creates privilege chain problems. A single thread may contain both privileged and non-privileged segments. AI tools that classify at the document level rather than the communication level will over-designate or under-designate. Look for tools that can parse threads.
• Categorical privilege logs are your friend. Many courts now accept categorical logs under United States v. Construction Products Research and its progeny. If your AI can reliably cluster documents into privilege categories, you can often negotiate a categorical log format that is both more efficient and more defensible than a document-by-document approach.

How FirmAdapt Addresses This

FirmAdapt's architecture is built around auditability, which turns out to be exactly what courts want to see when AI-assisted review is challenged. Every classification decision is logged with the underlying confidence score, the model version, and the human review disposition. When you need to produce a protocol document or respond to a motion to compel, the audit trail is already there, structured and exportable.

On the privilege log side specifically, FirmAdapt supports granular classification workflows that separate attorney-client and work product analysis, handle email thread segmentation, and generate log entries that include document-specific descriptions rather than boilerplate. The platform also flags low-confidence privilege classifications for priority human review, which directly supports the quality control sampling that courts expect. For teams handling large-scale review in federal litigation, this kind of built-in compliance infrastructure reduces both the risk of privilege waiver and the cost of defending your methodology after the fact.