Home

Why AI Document Review Matters for eDiscovery

Insight

October 16, 2026

AI document review uses software, trained or guided by lawyers, to find responsive, privileged and key documents faster than reading every file by hand. It matters because it cuts review time and cost while still allowing a defensible result, provided the process is validated by sampling, supervised by people and documented.

Document review is usually the largest cost and the longest phase of a discovery effort. Data volumes keep growing, and deadlines do not move. Legal teams need a way to review more material without lowering the standard of care.

This is where AI document review earns its place. It does not replace attorney judgment. It changes where attorney time goes. Instead of reading every document in sequence, reviewers train, check and correct a system that ranks or classifies documents at scale.

The legal footing is not new. In a March 2015 opinion in Rio Tinto PLC v. Vale S.A., a federal magistrate judge in the Southern District of New York wrote that "it is now black letter law that where the producing party wants to utilize TAR for document review, courts will permit it." The open question today is how to run the process so it holds up if challenged.

Key takeaways

AI document review can shorten eDiscovery document review by letting software prioritize or classify documents, with lawyers training and checking the results.
Courts have accepted technology-assisted review (TAR) for years. They have also held it to the same reasonableness standard as keywords or manual review, not a higher one.
Defensibility rests on validation: statistical sampling, recall and precision measurement, and a written record of decisions.
Federal Rule 26(g) requires a "reasonable inquiry" behind every discovery response, whichever method you use.
Human review stays in the loop. AI output is a recommendation that qualified people confirm.

What is AI document review in eDiscovery?

AI document review is the use of machine learning or generative AI to help decide which documents are relevant, privileged or important to an issue. Humans set the criteria and verify the output.

The EDRM, the group that publishes the standard eDiscovery reference model, defines technology-assisted review as "a review process in which humans work with software to train it to identify relevant documents." That phrase captures the core idea. The software learns from human decisions, and humans check what the software does.

In practice, AI eDiscovery tools take a few forms:

Predictive coding or TAR. Reviewers code a sample of documents. The system learns from those calls and scores the rest.
Prioritization. The system pushes likely responsive documents to the front of the queue so reviewers see them first.
Categorization. Documents are sorted by issue, privilege likelihood or other tags.
Quality control. The system flags documents where a reviewer's call looks inconsistent with similar documents.
Generative AI review. Large language models read each document against written instructions and return a call with a short explanation.

The EDRM TAR Guidelines cover several of these uses, including prioritization, categorization, privilege review and quality control.

How does AI doc review speed up eDiscovery document review?

AI doc review speeds up the work by removing the need for a person to read every document before a decision about it is made. The largest gains come from the first pass.

Traditional linear review assigns documents to reviewers in batches. Most of the time goes to reading material that turns out to be irrelevant. AI changes the sequence in three ways.

Fewer documents need first-pass human reading. The system handles bulk classification, and people concentrate on borderline and high-value documents.
Important documents surface earlier. Case teams learn the facts sooner, which helps with case assessment, witness preparation and settlement talks.
Consistency improves. A system applies the same criteria to document one and document one million. Human reviewers drift, tire and interpret instructions differently.

Speed is only half of the Rule 26(b)(1) picture. That rule limits discovery to nonprivileged matter "relevant to any party's claim or defense and proportional to the needs of the case." Factors include the amount in controversy, the importance of the discovery and "whether the burden or expense of the proposed discovery outweighs its likely benefit." A review method that reduces burden without sacrificing quality supports a proportionality argument.

What do courts say about AI eDiscovery and TAR?

Courts have accepted TAR as a legitimate method and have declined to hold it to a higher bar than other methods. They have not, however, treated it as a free pass.

The Sedona Conference TAR Case Law Primer, Second Edition (public comment version, February 2023) traces the development. It notes that Da Silva Moore (2012) established computer-assisted review as an available tool, and that Rio Tinto (2015) described acceptance as black letter law.

Three points from the case law matter for practitioners:

No higher standard. Judge Peck wrote in Rio Tinto that "it is inappropriate to hold TAR to a higher standard than keywords or manual review."
Responding parties choose the method. The Primer reports that courts have generally declined to force a responding party to use TAR when it prefers keyword searching, citing Hyles v. New York City.
Transparency is encouraged, not always required. Rio Tinto "highly recommend[ed]" that counsel be willing to discuss transparency in the process. The Primer notes that later decisions such as Entrata held the Federal Rules do not require transparent disclosures as a requirement of using TAR, with disclosure becoming an issue when production deficiencies emerge or an ESI protocol is violated.

Do not treat this as legal advice. Case law varies by court and by judge, and local rules and ESI protocols can add requirements. Confirm the standards that apply in your matter.

How do you keep AI document review defensible?

You keep AI document review defensible by validating results, using qualified human reviewers, and documenting each decision. The method matters less than your ability to show it was reasonable.

Rule 26(g)(1) is the anchor. An attorney signing a discovery response certifies that, "to the best of the person's knowledge, information, and belief formed after a reasonable inquiry," the response is consistent with the rules and is not unreasonable or unduly burdensome. The Sedona Primer notes that courts increasingly look to this "reasonable inquiry" standard when they consider validation, and that validation can apply to keyword search as well as TAR.

A defensible workflow usually includes:

Written objectives. Define relevance, issues and privilege criteria before anything is coded.
Expert human training and calls. Use senior, well-briefed reviewers for training decisions. Poor training produces poor results.
Sampling during the review. Check the system's calls against human judgment on random samples and correct course.
Validation at the end. Measure recall and precision on a statistical sample of both the documents set aside and the documents selected.
Privilege and sensitive-data checks. Apply human review before production.
A written record. Keep the protocol, sample sizes, results, reviewer instructions and the reasons for key decisions.

Rio Tinto points to the same idea from the requesting party's side. The court said requesting parties can confirm that training and review were done appropriately "by other means, such as statistical estimation of recall at the conclusion of the review as well as by whether there are gaps in the production."

What do recall and precision tell you?

Recall tells you how many of the responsive documents you found. Precision tells you how many of the documents you selected were actually responsive.

The Sedona Primer defines recall as the proportion of responsive documents actually identified, and precision as the proportion of identified documents that are actually responsive. High recall protects you against missing documents. High precision protects you against over-producing and spending reviewer time on noise.

Measure	Question it answers	Main risk if low
Recall	Did we find the responsive documents?	Missed documents, production gaps, challenges to the process
Precision	Of what we selected, how much was responsive?	Wasted review cost, over-production

The Primer reports that courts treat these metrics as important but apply proportionality analysis rather than a rigid threshold. There is no single number that makes a review defensible in every case. The target should fit the matter, and the reasoning should be recorded.

Where does human review still fit?

Human review remains essential at the points where judgment, context and accountability matter. AI output supports attorneys. It does not take over their duties.

People still need to:

Decide what counts as relevant and what the issues are.
Train the system and resolve disagreements between the system and reviewers.
Review privileged, confidential and sensitive documents before production.
Interpret ambiguous documents and context that a system may miss.
Sign off on the process under Rule 26(g).

Preservation is a separate duty, and no review tool cures a failure there. Rule 37(e) addresses electronically stored information that "should have been preserved in the anticipation or conduct of litigation" and is lost because a party "failed to take reasonable steps to preserve it." AI review speeds analysis of what you have. It cannot recreate what was lost, so legal holds and preservation steps still come first.

How Gemean helps

Gemean offers electronic discovery services that include data hosting and review platforms, eDiscovery consulting, managed review and technology-assisted review. It also offers PythAI for Document Review, which Gemean describes as a custom generative AI solution for litigation and investigation document review. Gemean's site lists uses such as first-level review of large document sets, early case insights, quality control for document review and analysis of opposing document productions. Gemean states that PythAI operates with expert consultants and keeps data in a private, dedicated environment.

If you are weighing AI document review for a matter, Gemean can help you scope the workflow, set up sampling and validation, and document the process. For related background, see Gemean's post The AI-Powered Edge.

Talk to Gemean

FAQs

Is AI document review accepted by courts?

Yes, in general. In Rio Tinto v. Vale (S.D.N.Y. 2015), the court wrote that courts will permit a producing party to use TAR for document review. Courts also say TAR should not be held to a higher standard than keywords or manual review. Rules and judges vary, so check your court's requirements and any ESI protocol.

What is the difference between TAR and AI document review?

TAR is a review process in which humans train software to identify relevant documents, per the EDRM. AI document review is a broader term. It includes TAR and also newer methods such as generative AI that reads documents against written instructions. Both need validation and human oversight.

Do I have to tell the other side I am using AI eDiscovery?

Not always. The Sedona Primer reports that courts encourage transparency but generally do not mandate disclosure of TAR processes. Disclosure can become required if production deficiencies emerge or an ESI protocol is violated. Many parties negotiate the approach early. Ask counsel what your court and protocol require.

How do you validate AI doc review results?

Validate by sampling. Draw random samples from the documents classified as responsive and not responsive, have qualified reviewers code them, and estimate recall and precision. Record sample sizes, results and any corrective action. This documentation supports your Rule 26(g) "reasonable inquiry."

What are recall and precision in eDiscovery document review?

Recall is the share of all responsive documents that your process found. Precision is the share of the documents you selected that are actually responsive. Courts treat both as useful measures and apply proportionality rather than a fixed cutoff, so the right target depends on the case.

Does AI replace human reviewers?

No. AI classifies, ranks and flags documents at scale, but attorneys and trained reviewers set criteria, train the system, check samples, handle privilege and sign off on the process. The EDRM definition itself describes humans working with software, not software working alone.