Predictive coding in eDiscovery is a machine-learning method for sorting large document collections during litigation, training software on human coding decisions so it can score every file in a dataset for relevance. Also called technology-assisted review, or TAR, it replaces much of the manual document review that would otherwise consume thousands of attorney hours on collections that routinely run into the millions of files. Federal courts have approved it since 2012, and in large-scale reviews it is now the expected approach rather than an experimental one.
How the Technology Actually Works
Predictive coding converts documents into mathematical representations and measures how similar they are to each other. The software maps every file as a point in a multi-dimensional space using word frequencies, word proximity, and structural patterns in the text. Metadata such as sender names, timestamps, and file types add further dimensions. When a new document enters the system, its mathematical fingerprint gets compared against the patterns the algorithm has already learned from human-coded examples, and the system assigns a relevance probability score.
That approach catches things a keyword search cannot. A search for “merger” will find documents containing the word but miss emails discussing “the deal,” “Project Falcon,” or whatever code name the parties used. Predictive coding picks up contextual relationships between words and concepts, flagging documents that are substantively about the merger even when the term never appears.
Research by Grossman and Cormack found that TAR systems achieved roughly 77% recall and 85% precision on average, while manual human review achieved only about 59% recall and 32% precision. The result is counterintuitive but well-documented: TAR does not just save money, it tends to find more relevant documents than teams of lawyers reading every page.
TAR 1.0 vs. TAR 2.0
Two dominant approaches exist, and they work differently enough that choosing the wrong one adds real cost. Understand the distinction before you negotiate a protocol with opposing counsel.
TAR 1.0: Simple Active Learning
TAR 1.0 separates training from review. A subject matter expert, typically lead counsel or a senior attorney with deep case knowledge, reviews and codes a seed set of documents, usually 500 to 2,000 files. The software learns from those decisions, and the expert then reviews additional rounds of algorithmically selected documents until the system hits a target recall level. Only after training is complete does the broader review team begin reviewing scored results above a cutoff threshold. Documents below the cutoff are treated as non-responsive, subject to validation sampling.
The drawback is rigidity. TAR 1.0 gives you essentially one shot at training, and it requires significant time from expensive subject matter experts before any large-scale review begins. It also struggles with rolling data loads, which are common when new custodians are added or productions arrive in stages.
TAR 2.0: Continuous Active Learning
TAR 2.0 eliminates the distinction between training and review. Every coding decision by every reviewer feeds back into the algorithm in a continuous loop, and the system constantly reprioritizes documents so the most likely relevant files surface next. There is no separate seed set phase, no fixed control set, and no predetermined stopping point for training. The system keeps learning until it runs out of relevant documents to find.
This approach consistently outperforms TAR 1.0 on both efficiency and effectiveness. In one comparison, a review team would have needed to manually examine 50,000 additional documents under a TAR 1.0 workflow to match the recall that continuous active learning delivered automatically. TAR 2.0 also handles rolling data loads naturally, since new documents simply enter the existing feedback loop.
Culling the Data Before Review
Running predictive coding on raw, unculled data wastes money. Before any TAR process begins, the data needs to be cleaned, and the quality of that preparation directly affects both cost and accuracy.
Global deduplication uses cryptographic hashing to generate a digital fingerprint of each file’s content. When the system encounters an identical hash, it suppresses the duplicate while recording which custodians held copies. That prevents your review team from coding the same email fourteen times across five mailboxes. Document review typically accounts for 60 to 70% of total eDiscovery spending, so eliminating redundant files before review begins is one of the most impactful cost-saving steps available.
Near-duplicate identification goes further, grouping documents that are substantially similar but not byte-for-byte identical, such as successive drafts of a contract or email chains where each reply adds a few lines to the same body. Clustering those near-duplicates lets reviewers code them in batches. DeNIST filtering removes known system files (operating system components, standard application files) that carry no evidentiary value. Together, these culling steps can reduce the volume entering TAR by a significant margin.
Training the Algorithm
In a TAR 1.0 workflow, the training phase is where the algorithm gets its marching orders. The subject matter expert reviews the seed set and tags each document as responsive or non-responsive to the litigation issues. Many platforms also allow coding for specific issue tags, flagging documents related to particular claims, contract provisions, or time periods.
Consistency during training is critical. The algorithm replicates the logic it is shown, so contradictory tagging decisions produce an unreliable scoring model. If one document about a contract amendment is tagged responsive and a nearly identical document is tagged non-responsive, the system receives conflicting signals that degrade its performance across the entire dataset. That is why training is typically limited to one or two senior reviewers rather than distributed across a large team.
Training runs in iterative rounds. After each round, the software surfaces borderline documents, the files it finds hardest to classify, for the expert to resolve. Those edge cases sharpen the model faster than reviewing randomly selected documents would. Training continues until the system shows stable performance, at which point the team transitions to full-scale review or, in TAR 1.0, applies a relevance score cutoff to separate the review population from the non-review set.
Privileged documents require separate handling throughout. Any file flagged as attorney-client privileged or work product gets segregated from the production set and logged on a privilege log describing the document’s nature without revealing its protected content.
Validating Results: Recall, Precision, and Elusion
Statistical validation is what makes predictive coding defensible in court. Without it, you are trusting the software. Three metrics matter most, and understanding what each actually measures will save you from accepting a validation report that looks good on paper but hides real problems.
- Recall is the percentage of all truly relevant documents in the collection that the system identified. A recall rate of 80% means the algorithm found four out of every five relevant documents. Low recall means you are missing evidence. Protocols commonly target recall in the 75 to 80% range, though the appropriate target depends on the case.
- Precision is the percentage of documents the system flagged as relevant that actually are relevant. Low precision means your review team wastes time on false positives.
- F1 score is a composite of precision and recall that ranges from 0 to 1, with higher scores indicating a better balance between the two. The F1 score helps you spot situations where high recall masks terrible precision, or vice versa.
Elusion testing targets the documents the software classified as non-responsive. A random sample from that discard pile gets reviewed by a human to check for relevant files the algorithm missed. The elusion rate, the percentage of responsive documents incorrectly classified as non-responsive, gives you a direct measure of what you are leaving behind. Parties typically negotiate an acceptable elusion rate as part of the TAR protocol. A rate under a few percentage points is often the target, though no single universal threshold exists. What matters is that the rate reflects a reasonable and defensible search given the stakes of the case.
These metrics interact. Pushing recall higher usually drives precision lower, because the system starts flagging marginal documents to avoid missing anything. The right balance depends on the proportionality analysis for your case. A bet-the-company patent dispute warrants a higher recall target than a routine contract claim.
What It Costs
The economics become compelling once you understand what predictive coding replaces. A manual review of one million documents at attorney billing rates measured in hundreds of dollars per hour would be ruinous for most litigants. Even staffing with contract reviewers at lower hourly rates still means paying humans to read documents one at a time, most of which turn out to be irrelevant.
Predictive coding reduces the number of documents that require human eyes. In a TAR 2.0 workflow, the algorithm continuously pushes the most likely relevant documents to the front of the queue, so your review team spends time on files that matter rather than plowing through junk. Savings scale with the size of the collection: the larger the dataset, the greater the share of documents the software handles without human intervention.
Line items to budget for include eDiscovery platform licensing or per-gigabyte processing and hosting fees, which vary widely across vendors. Review attorney time remains the largest expense, but TAR dramatically reduces the volume those attorneys need to touch. Training time from subject matter experts adds cost in TAR 1.0 workflows specifically, since senior attorneys must spend significant hours coding seed and training sets before the broader team can begin. TAR 2.0 reduces that burden because any reviewer’s coding decisions train the model continuously.
Judicial Acceptance and Key Case Law
Federal courts have endorsed predictive coding for over a decade. The case law has matured to the point where parties using TAR face less judicial skepticism than parties relying on keyword searches alone.
Da Silva Moore v. Publicis Groupe (2012)
This was the first federal decision to formally approve technology-assisted review. Magistrate Judge Andrew Peck held that “computer-assisted review is an acceptable way to search for relevant ESI in appropriate cases” and that the legal community should “seriously consider” TAR “for use in large-data-volume cases where it may save the producing party significant amounts of legal fees.”1Justia. Da Silva Moore v. Publicis Groupe et al The district court adopted Judge Peck’s ruling, finding no basis to conclude it was clearly erroneous.2Justia. Da Silva Moore v. Publicis Groupe et al
Rio Tinto PLC v. Vale S.A. (2015)
Three years later, Judge Peck declared that TAR had reached the status of “black letter law,” meaning courts will permit a producing party to use it as a matter of established legal principle rather than as an experimental novelty.3Justia. Rio Tinto PLC v. Vale, S.A. et al, No. 1:2014cv03042 – Document 207 (S.D.N.Y. 2015) The court also emphasized that the standard for TAR is not perfection but reasonableness and proportionality under the circumstances.4Justia. Rio Tinto PLC v. Vale, S.A. et al
Hyles v. City of New York (2016)
Hyles set an important limit: a requesting party cannot force the producing party to use predictive coding. Judge Peck ruled that while he was “a judicial advocate for the use of TAR in appropriate cases,” cooperation principles do not give the opposing side or the court power to dictate the producing party’s review methodology.5Justia. Hyles v. City of New York et al The producing party is “best situated to evaluate the procedures, methodologies, and technologies appropriate for preserving and producing their own electronically stored information.” If you are the requesting party and want TAR, persuasion and proportionality arguments are your tools, not a court order.
Proportionality Under Rule 26
The legal foundation for predictive coding runs through Federal Rule of Civil Procedure 26(b)(1), which limits discovery to information that is relevant to any party’s claim or defense and proportional to the needs of the case. The rule requires courts to weigh the importance of the issues, the amount in controversy, the parties’ relative access to information, and whether the burden of the proposed discovery outweighs its likely benefit.6Legal Information Institute. Federal Rules of Civil Procedure Rule 26 – Duty to Disclose; General Provisions Governing Discovery
Predictive coding maps neatly onto that framework because it provides a statistically measurable way to demonstrate that a search was both comprehensive and cost-proportionate. When opposing counsel challenges a production as inadequate, you can point to specific recall and precision rates, validation sample results, and documented training decisions. Keyword searches cannot produce the same kind of record. Rule 26(g) also requires the attorney signing a discovery response to certify that it is the product of a reasonable inquiry and is not unreasonably burdensome given the needs of the case, and a well-documented TAR process makes that certification far more defensible than keyword lists selected by instinct.
Negotiating a TAR Protocol
Courts expect transparency when parties implement predictive coding, and working out the details upfront avoids disputes that can derail timelines. A TAR protocol typically addresses the software platform, whether the workflow follows TAR 1.0 or TAR 2.0, what validation metrics and thresholds the parties will accept, and how disputes about the results will be resolved.
One recurring flashpoint is whether to share seed sets with opposing counsel. Some courts have encouraged or required disclosure so that both sides can evaluate whether training documents fairly represent the issues. Others leave seed-set disclosure to the parties’ agreement. If you are the producing party, sharing the seed set may actually work in your favor. It builds a record of cooperation that makes the production harder to challenge later.
Failure to cooperate on protocol terms can lead to sanctions or court-ordered re-reviews at significant expense. Judges have limited patience for parties who stonewall methodology discussions, particularly when the law on TAR cooperation has been clear for over a decade. Getting the protocol into a court order or stipulation before review begins protects both sides and gives the process judicial backing if challenged.
A Federal Rule of Evidence 502(d) order, obtained early in the case, provides a safety net by ensuring that any inadvertent disclosure during litigation does not waive privilege in any other proceeding.7Legal Information Institute. Federal Rules of Evidence Rule 502 – Attorney-Client Privilege and Work Product; Limitations on Waiver Getting a 502(d) order in place before production begins is one of the smartest protective steps in any TAR workflow, and many legal teams overlook it.
Ethical Duties for Supervising Attorneys
Attorneys using predictive coding carry ethical responsibilities beyond running the software. ABA Model Rule 1.1 requires competent representation, and the accompanying commentary states that competence includes staying current on “the benefits and risks associated with relevant technology.” The ABA clarified in its 2012 amendment to Comment 8 that this is not a new obligation, just a reminder that technology literacy has always been part of a lawyer’s duty of competence.8American Bar Association. ABA Issues First Ethics Guidance on a Lawyer’s Use of AI Tools
In practice, the attorney overseeing a TAR process needs to understand how the algorithm works well enough to explain the methodology to a court. You do not need to be a data scientist. You do need to know what recall and precision measure, why seed set quality matters, and what your validation results actually prove. Delegating the entire process to a vendor without meaningful oversight is a competence problem waiting to become a sanctions problem.
Supervising attorneys also have a duty to ensure that contract reviewers and paralegals working within the TAR workflow understand how their coding decisions affect training. In a continuous active learning system, every reviewer’s decisions feed back into the model. One poorly trained reviewer coding inconsistently can degrade the entire system’s performance, and the supervising attorney bears ethical responsibility for the outcome.
Generative AI and New Privilege Risks
Large language models in eDiscovery workflows have created new legal risks that courts are only beginning to address. The core concern is straightforward: uploading privileged or confidential documents into a public AI tool may waive the protections those documents carry.
Federal courts have begun splitting on how to treat these interactions. In early 2026, a court in the Southern District of New York ruled that communications with a public generative AI platform were not protected by attorney-client privilege or the work product doctrine, finding that users lacked a reasonable expectation of confidentiality because the platform’s terms reserved the right to use inputs for training and share data with third parties. Courts in other districts have taken a more permissive approach in civil matters, treating AI platforms as software tools rather than adversaries and declining to find a privilege waiver.
The safest path is to use enterprise or closed AI systems that do not train on user inputs and that process data within a controlled environment. Some courts have begun updating protective orders to restrict or prohibit the use of public AI tools for any discovery materials, not just confidential documents. If your case involves a protective order, check whether it addresses AI tools before uploading anything, even if the documents are not designated as confidential.
ABA Formal Opinion 512, issued in 2024, reinforced that lawyers using generative AI tools must exercise their own skill and judgment rather than rely on AI output alone, and must verify AI-generated work for accuracy before submitting anything to a court.8American Bar Association. ABA Issues First Ethics Guidance on a Lawyer’s Use of AI Tools High-profile incidents of attorneys submitting AI-fabricated case citations have made judges acutely sensitive to this issue, and the duty of candor to the tribunal applies with full force to any AI-assisted work product.
When Predictive Coding Is Not the Right Tool
TAR is not the answer for every case. The technology performs best on large, text-heavy document collections where the cost of manual review would be disproportionate to the stakes. For smaller datasets, generally under about 10,000 to 20,000 documents, the overhead of setting up a workflow, training the algorithm, and running validation may exceed what it would cost to simply have reviewers read every file.
Collections dominated by non-text content also present challenges. Scanned images without optical character recognition, audio recordings, video files, and spreadsheets with primarily numerical data give the algorithm less textual content to analyze. Predictive coding works by comparing language patterns, so files with little or no extractable text do not fit naturally into the model. A hybrid approach often makes sense, using TAR for the email and document populations while handling non-text files through targeted manual review or specialized tools.
Cases where relevance turns on context external to the document, such as whether a financial transaction occurred on a specific date, can also challenge TAR’s effectiveness. The algorithm assesses textual similarity rather than factual accuracy. In those situations, structured data analytics or database queries often work better than document classification algorithms.