Predictive Coding, in Simple Terms
Predictive coding is a technology-assisted review (TAR) method that learns from human reviewers’ document decisions and uses those examples to rank or classify other electronic documents. Rather than reading every document in arbitrary order, a review team can focus first on documents the model predicts are more likely to matter. It is a workflow combining legal judgment, software, and quality control; not a substitute for the legal team. [1] [12]
Imagine reviewing a collection of 100,000 emails. Attorneys label examples as responsive or nonresponsive. A model learns patterns in those examples and assigns predictions to unreviewed emails. As review continues, the team checks the model’s results, corrects errors, and assesses whether additional review is necessary.
How Does Predictive Coding Work?
Define the review question
The team defines what “responsive,” “relevant,” or another target category means for the matter. [1]
Review and code examples
Human reviewers examine documents and apply consistent coding decisions. [1]
Train and rank
Software identifies patterns and assigns rankings or predicted categories to unreviewed material. [1]
Continue reviewing
Depending on the system, reviewers receive prioritized documents and their decisions further inform the model. [1]
Validate and document
Sampling, control sets, recall and precision estimates, reviewer quality checks, and documented decisions help assess whether the workflow is defensible. [1]
What Is Continuous Active Learning (CAL)?
Continuous active learning is a TAR workflow in which the system uses ongoing reviewer decisions to update document priorities. A high prediction score indicates a stronger model prediction for the defined category; it does not itself establish legal responsiveness, admissibility, or privilege. Different products use different scoring scales and update schedules. [1] [4] [7]
Predictive Coding vs. Generative AI
Traditional predictive coding learns from coded examples and predicts classifications or ranks. Generative-AI review tools may instead interpret written review instructions and produce document-level assessments and explanations. Both can assist document review, but their setup, validation, outputs, and risks differ. The names of vendor products should not be treated as interchangeable. [1] [8]
Predictive Coding in Everlaw
Everlaw offers a default model using its Hot/Warm/Cold document ratings and lets teams create custom models based on review codes and other criteria. Its model assigns documents scores from 0 to 100 and updates predictions as review progresses. Everlaw says a model starts producing predictions after at least 200 qualified reviewed documents, including at least 50 positive and 50 negative examples. Qualified documents must meet additional conditions, such as sufficient text and nonconflicting coding. [2] [3]
Review teams can use the scores to prioritize assignments, compare model predictions with human decisions, and assess performance. Everlaw also documents control-set and validation functionality, clustering overlays, and reusable multi-matter models. The product’s predictive coding is distinct from its other generative-AI features. [10] [2]
Predictive Coding in Nebula
Nebula’s Predictive Coding uses continuous active learning based on reviewer coding decisions, giving documents a score from 0 to 100 to prioritize likely responsive documents. Its documented workflow includes creating a predictive coding project, choosing a seed set, reviewing documents, running training sessions, monitoring scores, and quality-checking the remaining pool. [4] [5]
Nebula supports control sets and classifier-performance metrics. Its documentation specifies individual-document coding for training rather than simply relying on family or email-thread coding. KLDiscovery says its eDiscovery Technologist team handles predictive coding projects and consults with clients. This is a more explicitly specialist-supported workflow than merely turning on an automatic tag suggestion. [4] [5]
Predictive Coding in DISCO
DISCO AI provides tag predictions learned from past document-tagging decisions. Its documentation describes a score for each document-and-tag pair from −100 to +100: negative values indicate the model predicts the tag is unlikely to apply, while positive values indicate it is more likely. Predictions are updated as additional documents are tagged. [6]
For a responsiveness workflow, reviewers can use these predictions to prioritize documents likely to receive a responsive tag, while also applying the tool to other review tags. DISCO’s tag-prediction feature should be distinguished from any separate generative-AI review offering: the predictive score described here comes from learned tagging patterns, not a standalone legal determination. [6]
Predictive Coding in Relativity
RelativityOne’s Review Center offers prioritized review using a machine-learning classifier that learns from reviewer coding decisions. It ranks documents, generally from 0 to 100, and uses those rankings to serve documents to reviewers. Relativity’s documentation explains the support-vector-machine approach used in this workflow and provides monitoring and review-queue tools. [7]
Relativity also offers aiR Review (formerly aiR for Review), which is different: it uses generative AI and natural-language prompt criteria to analyze extracted document text, provide relevance or issue predictions, and offer rationales and citations where applicable. Its result scores use a different scale from traditional active-learning document ranks. A team evaluating “predictive coding in Relativity” should specify whether it means Review Center/active learning or aiR Review. [8] [9]
How Do These Four Platforms Differ?
Everlaw emphasizes configurable models grounded in existing review ratings or codes, with a documented training threshold. Nebula describes a CAL project workflow with training sessions, control sets, and KLDiscovery technologist involvement. DISCO describes ongoing tag-level predictions on a −100 to +100 scale. Relativity’s Review Center provides active-learning prioritization, while aiR Review separately uses generative-AI prompt-based analysis. These are descriptions of documented workflows, not a ranking of quality, cost, or legal defensibility. [2] [3] [4] [5] [6] [7] [8]
What Does a Prediction Score Actually Mean?
A score is a tool-specific prediction, not a universal percentage of legal certainty. Everlaw, Nebula, and Relativity use 0–100-style ranking in the cited workflows; DISCO describes −100 to +100 tag scores; aiR Review has its own categorical scoring system. A score of 80 in one platform should not be assumed equivalent to 80 in another. [2] [4] [6] [7] [9]
Can Predictive Coding Replace Human Document Review?
Predictive coding may reduce the amount of manual review needed, but legal teams still need clear criteria, training examples, quality control, privilege handling, and an appropriate validation plan. EDRM’s TAR Guidelines describe human involvement, testing, and validation as core parts of the process. A model can reproduce inconsistent human coding, overlook rare document types, or perform poorly if the review question changes. [1] [12]
Why Predictive Coding Matters for eDiscovery in Hawaiʻi
Hawaiʻi litigation and investigations can involve substantial electronic evidence just like matters elsewhere. Predictive coding is one way legal teams may organize large collections and focus review resources. Whether and how it may be used in a specific matter depends on applicable discovery rules, orders, agreements, and the team’s documented process; software functionality does not itself establish compliance with a court’s requirements. [1] [11]
A Simple Predictive Coding Example
In a contract dispute, the team collects 75,000 emails and documents. Reviewers code a starting group using a consistent definition of responsiveness. The platform ranks the remaining documents. Reviewers examine high-priority results, continue coding, and evaluate quality-control samples from lower-ranked documents. The team uses those findings (not the prediction score alone) to decide what additional review and production steps are appropriate.
Frequently Asked Questions
Is predictive coding the same as TAR?
The terms are often used interchangeably, although TAR is sometimes used more broadly for multiple technology-assisted workflows. [1]
Is continuous active learning the same as predictive coding?
CAL is a type of predictive-coding workflow in which ongoing reviewer decisions inform the model. [1]
Does predictive coding automatically decide privilege?
Not by itself. Privilege is a separate legal determination that needs appropriate review and safeguards. [1]
Does a low score mean a document can be deleted?
No. A low score is a prediction within a model, not permission to delete or disregard preservation and discovery obligations. [1]
Is Relativity aiR Review the same as Active Learning?
No. aiR Review is a generative-AI, prompt-based analysis product; Review Center uses a reviewer-trained classification and prioritization workflow. [7] [8]
Dictionary of Key Terms
Predictive coding: Machine-learning-assisted classification or prioritization of documents using human-coded examples.
TAR: Technology-assisted review; a human-and-software review process.
CAL: Continuous active learning; a process that updates predictions as review decisions accumulate.
ESI: Electronically stored information, such as emails, messages, and electronic files.
Responsive: Within the scope of a discovery request under the governing criteria.
Seed set: An initial group of documents selected for model training or early review.
Training set: Reviewed and coded examples used to inform a predictive model.
Control set: A sample used to assess model performance, often separated from training decisions.
Recall: The proportion of all actually relevant documents that a workflow identifies as relevant.
Precision: The proportion of documents identified as relevant that are actually relevant.
Elusion: Relevant documents found in a population predicted or designated nonresponsive.
Relevance rank: A model-generated score used to order documents by predicted relevance.
Coding: Applying review labels or tags to documents.
Validation: Testing and documenting whether a review process meets its intended objectives.
Quality control: Checks designed to identify errors and assess consistency and completeness.
SVM: Support vector machine, a machine-learning classification technique.
Generative AI: AI that generates analyses or text from instructions and input content; not identical to reviewer-trained predictive coding.
Privilege: A legal protection potentially shielding qualifying information from disclosure.
Sources and Further Reading
[1] EDRM — Technology Assisted Review
[2] Everlaw — Beginner’s Guide to Predictive Coding
[3] Everlaw — Predictive Coding FAQs
[4] Nebula — Predictive Coding
[5] Nebula — Predictive Coding Workflow
[6] DISCO — AI Tag Predictions
[7] Relativity — Review Center
[9] Relativity — Analyzing aiR Review Results
[10] Everlaw — Predictive Coding Product
[11] Hawaiʻi Judiciary — Rules of Court
[12] EDRM — TAR Guidelines PDF
Legal Disclaimer
This article is general educational information, not legal advice or a guarantee of any platform’s capabilities, availability, pricing, or compliance. Features and names can change. Review current vendor documentation and governing rules for a particular matter.