What is Predictive Coding in eDiscovery?

Predictive coding is a supervised machine learning workflow used in eDiscovery document review. During this EDRM stage, legal teams decide if documents are relevant, privileged, or responsive. 

Attorneys manually tag a sample of documents (often called a “seed set”). The predictive coding AI model learns phrase and metadata patterns that separate relevant documents from non-relevant ones. It then scores every unreviewed document.

Predictive Coding vs TAR in eDiscovery

The terms “Predictive Coding” and “technology-assisted review” (TAR) are often used interchangeably. However, they are not exactly the same.

TAR is a broad term for automated document review using eDiscovery software. Predictive coding is the machine learning method used inside TAR. In the past, it meant training a model once on a fixed seed set.

Today, most TAR predictive coding uses continuous active learning. The model updates in real time as reviewers code documents.

The model does not write or summarize content. It also does not make final decisions. Attorneys stay responsible for coding calls and for the defensibility of the process.

How Does Predictive Coding Work?

Predictive coding follows a repeatable cycle of human coding and machine prediction:

1. Defining the review set: The team collects, processes, deduplicates, and culls the data before training starts.
2. Tagging a seed set: Subject matter experts tag a sample of documents as relevant or not relevant.
3. Training the model: The predictive coding machine learning model learns which text and metadata link to each decision.
4. Dataset scoring and ranking: The model gives every remaining document a relevance score. Likely relevant material then rises to the top.
5. Reviewing and refining: Human reviewers inspect the high-scoring documents. Their decisions feed back into the model to continuously update and improve its accuracy.
6. Validating the results: The legal team samples the remaining unreviewed documents to measure recall (the percentage of total relevant documents found) and precision (the accuracy of the AI's predictions). This proves the process is defensible before production.

Types of Predictive Coding

Two approaches are widely recognized:

• TAR 1.0 (simple active or passive learning): The model is trained on a fixed sample, often checked against a control set, and then applied to the rest of the documents.
  
  Best for: Linear datasets where the legal issues are clear from day one and will not change.
  
  
• TAR 2.0 (continuous active learning, or CAL): The model updates as reviewers code documents. Training and review happen together. No separate training stage is needed.
  
  Best for: Fast cases, rolling data collections, and investigations where relevance shifts over time.

Why Predictive Coding Matters

Predictive coding changes the cost and speed of document review. In linear review, attorneys read every document in order. That is slow and costly for large datasets. Predictive coding brings key evidence to the surface early. Legal teams can then shape their case strategy sooner.

It also supports proportionality under Federal Rule of Civil Procedure 26(b)(1) and defensibility. Courts have accepted predictive coding since Da Silva Moore v. Publicis Groupe (2012), provided the process is documented and validated with recall and precision measures. 

Discover how Venio’s built-in predictive coding accelerates document review and lowers costs. Book a demo today.

‍

Frequently Asked Questions

How is predictive coding used in eDiscovery employment cases?

Predictive coding helps employment teams sort large volumes of emails, chat messages, and HR records. These records come from managers, HR staff, and employees involved in a claim. It is very effective in complex commercial litigation, regulatory investigations, antitrust matters, and employment cases, where teams must review large email archives and chat histories.

Does predictive coding replace human document reviewers?

No, predictive coding amplifies human effort rather than replacing it. Subject matter experts are essential to review the initial seed set, train the algorithm, and validate the results. The software simply prioritizes the most relevant documents so attorneys can focus their time on complex legal analysis.

How does predictive coding make the eDiscovery process more efficient?

Predictive coding uses machine learning to automatically sort large datasets. This greatly reduces the time and cost of manual document review.