Electronic discovery has undergone a profound transformation over the past decade, driven by the application of machine learning and artificial intelligence to the challenge of reviewing vast document collections in litigation. What was once a labor-intensive process requiring armies of contract attorneys to review documents one by one has evolved into a technology-assisted workflow where algorithms identify relevant documents, cluster related concepts, and prioritize review queues. Understanding how these tools work, their legal acceptance, and their limitations is essential for any legal professional involved in modern litigation.
The Scale Problem in Modern eDiscovery
The volume of electronically stored information (ESI) in modern organizations is staggering. A single custodian in a large corporation may generate tens of thousands of emails, documents, and messages per year. In complex commercial litigation, the universe of potentially relevant documents can easily reach millions or tens of millions. Traditional linear review, where each document is read and coded by a human reviewer, is neither economically feasible nor timely at this scale.
The cost implications are substantial. Linear review of a million documents at an average rate of 50 documents per hour and a billing rate of $50 per hour would cost approximately $1 million in review fees alone, not counting quality control, project management, or second-level review. Machine learning technologies can reduce these costs by 60 to 80 percent while maintaining or improving accuracy.
Technology-Assisted Review: Core Concepts
Technology-Assisted Review (TAR) is the umbrella term for eDiscovery processes that use machine learning algorithms to classify documents as relevant or non-relevant. The two primary TAR methodologies are TAR 1.0 (also called predictive coding) and TAR 2.0 (continuous active learning).
TAR 1.0: Predictive Coding
In TAR 1.0, a subject matter expert reviews a random sample of documents, called the seed set, coding each as relevant or non-relevant. The algorithm trains on this seed set and then applies the learned patterns to the full document population, generating a relevance score for each document. The scored population is then validated through statistical sampling to confirm that the algorithm has achieved acceptable recall and precision rates.
TAR 1.0 works well when the concept of relevance is relatively stable and can be captured in a single training round. However, it struggles with document populations where relevance is multifaceted or where new issues emerge during the course of review.
TAR 2.0: Continuous Active Learning
TAR 2.0 addresses the limitations of TAR 1.0 by implementing continuous learning. Instead of training on a single seed set, the algorithm continuously updates its model as reviewers code documents throughout the review process. The system prioritizes the most likely relevant documents for human review, then uses the reviewer's coding decisions to refine its predictions in real time.
This approach is generally more efficient and flexible than TAR 1.0. It adapts to evolving definitions of relevance, handles multiple issues more effectively, and can achieve high recall rates with fewer human review decisions. Research by Grossman and Cormack has demonstrated that TAR 2.0 consistently matches or outperforms human review teams in identifying relevant documents.
Advanced AI Applications in eDiscovery
Concept Clustering
Beyond binary relevance classification, AI tools can cluster documents by concept, identifying groups of related documents that address similar topics or themes. This is particularly valuable in the early stages of litigation when legal teams need to understand the contours of a document population quickly. Concept clustering can reveal unexpected themes, identify key custodians, and inform case strategy before formal review begins.
Email Threading
AI-powered email threading groups emails and their replies into conversational threads, allowing reviewers to assess entire conversations in context rather than reviewing individual messages in isolation. This improves both efficiency and comprehension, as the relevance of an individual email often depends on the broader conversation in which it appears.
Sentiment and Communication Pattern Analysis
More advanced AI applications can analyze communication patterns and sentiment within document collections. These tools can identify unusual spikes in communication between specific individuals, detect changes in tone or language that may indicate awareness of wrongdoing, and map social networks within organizations. While these analyses are not dispositive, they can guide review strategies and help investigators identify key documents and custodians more efficiently.
Large Language Models in Document Review
The emergence of large language models has opened new possibilities for eDiscovery. LLMs can summarize documents, answer specific questions about document content, translate foreign-language documents, and identify privileged material with greater contextual understanding than traditional keyword-based approaches. However, their use raises questions about accuracy, hallucination risk, and the defensibility of LLM-assisted review processes.
Legal Acceptance and Defensibility
US courts have broadly accepted technology-assisted review as a legitimate and defensible approach to document review. The landmark case Da Silva Moore v. Publicis Groupe (2012) was the first federal court decision to approve the use of predictive coding, with Magistrate Judge Andrew Peck endorsing it as an acceptable method for identifying relevant documents. Subsequent decisions have reinforced this acceptance.
In Rio Tinto PLC v. Vale S.A. (2015), Judge Peck went further, stating that TAR is now widely accepted and that the real question is not whether TAR is acceptable but whether it is being used properly. Courts have generally focused on the reasonableness and transparency of the TAR process rather than the specific technology employed.
- Transparency: Courts expect parties to disclose their use of TAR and provide sufficient information about the methodology employed
- Validation: Statistical validation of results, typically through sampling and quality control measures, is expected
- Cooperation: Courts favor cooperative approaches where parties agree on TAR protocols, though judicial intervention can resolve disputes
Ethical Considerations
The use of AI in eDiscovery raises ethical considerations that legal professionals must navigate carefully. Lawyers remain responsible for supervising the review process and ensuring its adequacy, regardless of the technology employed. Blind reliance on algorithmic outputs without human oversight could constitute a failure of professional responsibility.
Privilege review is an area requiring particular caution. While AI tools can assist in identifying potentially privileged documents, the determination of whether attorney-client privilege or work product protection applies to a specific document requires legal judgment that cannot be fully delegated to an algorithm. Inadvertent production of privileged documents due to over-reliance on AI tools could have serious consequences.
Implementation Best Practices
Organizations implementing AI-powered eDiscovery should establish clear protocols that define the role of technology in the review process, the qualifications and training of human reviewers who interact with the system, quality control procedures, and validation methodologies. Documentation of these protocols is essential for defensibility.
Training data quality is critical to outcomes. The subject matter experts who train AI models must understand the legal issues in the case, apply consistent coding decisions, and provide the algorithm with a representative sample of both relevant and non-relevant documents. Poor training data produces poor results, regardless of the sophistication of the underlying algorithm.
Finally, legal teams should view AI as a complement to human expertise, not a replacement. The most effective eDiscovery workflows combine the efficiency and consistency of machine learning with the judgment, contextual understanding, and ethical obligations of experienced legal professionals.
