There was a moment, not long ago, when technology-assisted review in eDiscovery was a controversial subject. Law firms argued about whether predictive coding was reliable enough to substitute for linear review. Courts issued opinions establishing validation requirements. Practitioners debated defensibility.
That debate is largely settled. AI-assisted review is now the default methodology in large-scale eDiscovery. The new debate — still early and still unsettled — is about what comes next: how AI reshapes the discovery process in arbitration and mediation, how it is changing the analysis of litigation risk and case strategy, and what obligations practitioners carry when they deploy tools whose internal logic they may not fully understand.
The Duty of Technological Competence Is Not Optional
Comment 8 to ABA Model Rule 1.1 requires lawyers to "keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology." In practice, this means that a litigator who advises a client on eDiscovery without understanding how AI-assisted review tools work — their capabilities, their limitations, and the validation methodology that makes their results defensible — is operating below the professional standard the Model Rules establish.
This is not an academic point. Courts have imposed sanctions for eDiscovery failures that stemmed from counsel's failure to understand the technology their vendors were deploying. Reliance on a vendor does not transfer professional responsibility; the attorney remains accountable for the adequacy of the process and the reliability of the results. If the technology fails in a way the attorney would have caught with basic competence, the vendor contract does not insulate the lawyer.
How AI-Assisted Review Actually Works — and Where It Fails
Technology-assisted review (TAR) — also called predictive coding — uses machine learning algorithms trained on a sample of documents reviewed and coded by experienced counsel. The system learns to identify responsive, privileged, and non-privileged documents by pattern-matching against the training sample and extrapolating across the full document population.
The methodology works well when the training sample is representative, the relevant documents follow consistent patterns, and the model's performance is validated against a control set before reliance. It breaks down in identifiable circumstances that practitioners must watch for:
Atypical documents. AI models are trained on patterns. Documents that are responsive but atypical — because they use unusual language, were written by non-native speakers, or represent a factual development the training review did not anticipate — can be systematically misclassified. A thorough validation protocol includes testing the model against categories of documents known to be responsive to verify they are not falling through.
Privilege creep. Predictive coding can be used for privilege review, but with elevated risk. Privilege determinations often turn on fine-grained factual distinctions — who was on the communication, in what capacity, for what purpose — that are difficult to teach to a machine learning model. Counsel should consider targeted human review of documents flagged near the privilege classification boundary rather than relying entirely on model output.
Evolving fact patterns. In long-running matters where the relevant facts develop over time, a model trained on early-stage document samples may not adequately capture later-developed relevance categories. Periodic retraining or supplemental human review of later document tranches is necessary.
eDiscovery in ADR: A Growing Frontier with Fewer Rules
Arbitration and mediation have historically operated with lighter discovery obligations than federal court litigation. For many commercial parties, that was a feature, not a bug: limited discovery was part of the efficiency value proposition for ADR.
That calculus is shifting. As the disputes being sent to arbitration grow in complexity — intellectual property, employment class actions, financial fraud, data breach claims — parties increasingly need the same quality of data transparency that litigation provides. Arbitration panels are being asked to address ESI disputes that rival the complexity of MDL discovery practice. Mediators are facilitating settlements of data breach class actions where the damages analysis depends directly on the scope of what was exposed and to whom.
The practical consequence is that practitioners in ADR settings must now understand eDiscovery as a core competency, not a litigation-specific specialty. When an arbitration clause covers a dispute that will turn on document evidence, the discovery provisions of that clause — or the absence of adequate discovery provisions — directly affect the quality of the proceedings and the reliability of the outcome.
Poorly drafted arbitration clauses that omit document discovery protocols, fail to address preservation obligations, or provide no mechanism for resolving ESI disputes can transform the theoretical efficiency advantage of ADR into an actual disadvantage, producing proceedings that cannot be resolved on the merits because the relevant evidence was never produced.
AI as a Strategic Tool: Case Outcome Prediction
Beyond document review, AI tools are increasingly being used for case outcome prediction — analyzing large volumes of judicial decisions to identify fact-pattern correlations with favorable rulings, judge-specific tendencies, and probability-weighted outcome estimates for settlement analysis.
The capability is real and growing. AI tools trained on comprehensive judicial decision databases can identify patterns invisible to even experienced practitioners relying on case research. In class action analysis, for example, AI can process far more certification decisions, across more jurisdictions and fact patterns, than human research allows — and can model the probability of certification outcomes as a function of specific factual and legal variables.
The risk, however, is equally real. Outcome prediction tools produce statistical estimates, not certainties. A tool that shows a 70 percent historical win rate on a motion type in a particular jurisdiction is telling you something about base rates — not about your specific case, your specific judge, or the specific factual developments that will drive the analysis in your matter. Overreliance on statistical predictions can distort the judgment that experienced practitioners develop over years of practice, and clients are paying for that judgment, not for a number.
Used correctly — as one input into a broader strategic analysis, not as a substitute for it — AI outcome prediction tools provide genuine value. Used as a replacement for professional judgment, they introduce a different category of risk.
What Modern Practice Requires
The synthesis of these developments is straightforward: eDiscovery and AI competence are no longer specialties that litigators can delegate entirely to discovery counsel or vendors. They are core professional obligations. The practitioners who understand how AI-assisted review works, what makes its results defensible, where its limitations lie, and how its strategic applications in outcome analysis can inform but not replace professional judgment — those practitioners are better equipped across every matter they handle, in every forum.
Law schools are beginning to integrate this training. Bar associations are developing guidance. Courts are developing standards. The practitioner who waits for the profession to fully institutionalize these skills before developing them personally will be behind by the time the profession catches up.
How Law & Forensics Helps
Law & Forensics advises litigators, arbitrators, and in-house counsel on AI-assisted review and its defensibility — designing and validating TAR protocols, auditing vendor workflows, serving as forensic neutral or special master on ESI disputes in court and in arbitration, and providing expert analysis when a review process is challenged. We help counsel meet the competence standard the rules now demand and keep AI an instrument of professional judgment rather than a substitute for it.
Key Takeaway: Practitioners handling matters with significant eDiscovery should be able to answer three questions about any AI-assisted review process in their case: What validation methodology was used to verify the model's performance? What categories of potentially responsive documents were specifically tested to confirm they were not being systematically excluded? Who is personally accountable for the defensibility of the review protocol? If the answers to any of these questions come entirely from the vendor rather than from counsel's own understanding, the attorney's professional exposure deserves attention.