The surprising case for AI judges

Summary of The surprising case for AI judges

by The Verge

1h 13m•February 12, 2026
Listen to Episode

Overview of Decoder — The surprising case for AI judges

This episode of Decoder (host Neal A. Patel) features Bridget McCormack — former Chief Justice of the Michigan Supreme Court and now President & CEO of the American Arbitration Association (AAA) — discussing the AAA’s AI Arbitrator: a narrow, agent-based system that can decide certain arbitration disputes (currently documents‑only construction cases). The conversation covers why automation might improve predictability and access to justice, how the system is built and governed, where it makes sense (and where it doesn’t), the risks (hallucinations, bias, accountability), and the safeguards AAA is using as it begins real-world deployments.

Key points and main takeaways

  • Why automate dispute resolution?
    • Many legal outcomes today are probabilistic and inconsistent because courts are human-run, underresourced, and often inaccessible—especially for people and small businesses who can’t afford lawyers.
    • AI can increase predictability for many routine disputes, let parties “feel heard” by validating claims and summarizing evidence, and lower cost/time barriers to resolution.
  • What AAA built
    • An AI-native case management platform plus an “AI Arbitrator” composed of multiple agents (~20+) that parse pleadings/evidence, organize claims, reason, draft awards — with humans in the loop.
    • Launched narrowly: documents-only construction disputes (no live witness credibility assessment). As of the interview there was one live case; AAA ran retrospective tests and an academic review.
  • Governance and safeguards
    • Narrow scope, human arbitrator oversight, transparent audits, due‑process checks on arbitration clauses, an academic white paper review, and careful training/grounding of agents on industry-specific documents.
  • Risks and trade-offs
    • Hallucinations and model errors remain a real danger; institutional control, auditability, and human oversight are essential.
    • Arbitration’s private, fee‑driven nature raises fairness and accountability concerns—especially in consumer/contracts-of-adhesion contexts.
    • Some disputes should remain public (e.g., criminal matters, government actions) to preserve transparency and public accountability.
  • Where this can expand
    • Other documents-only dispute niches: supplier disputes, insurer-provider (payer-provider) disputes, smaller construction matters, internal dispute resolution, early case evaluation tools.
  • Timeline and uncertainty
    • McCormack expects steady expansion industry-by-industry; she predicts human judges will handle fewer routine decisions over years/decades, but exact timing is uncertain (years rather than months).

How the AI Arbitrator works

  • Platform: AI-native case management system accessible via web (and same infrastructure powering other AAA services later).
  • Agents:
    • Front-end agents parse party submissions, extract claims/elements, and iterate with parties to confirm understanding (“Did I get this right?”).
    • Reasoning agents analyze the grounded facts and legal framework.
    • Draft-award agents prepare an award for human review.
  • Human-in-the-loop:
    • Parties validate the parsed case summary.
    • A cohort of experienced construction arbitrators serve as the human arbitrator who reviews, edits, and issues the final award.
  • Scope & limits:
    • Documents-only disputes (no live witness credibility assessments).
    • Built from AAA’s library of past construction cases and a constructed handbook; trained with help from construction arbitrators and lawyers.
  • Development:
    • Internal AI engineers + partnering with QuantumBlack (McKinsey’s AI team) for the MVP; AAA staff trained broadly on LLMs.

Arguments in favor of AI judges/arbitrators

  • Access to justice: many citizens and small businesses are effectively shut out of courts; faster cheaper dispute resolution can extend coverage.
  • Predictability and efficiency: deterministic guidance for routine disputes reduces needless litigation and operational friction.
  • Feeling heard: iterative agent summaries can make parties confident their issues were understood — an underappreciated component of procedural fairness.
  • Auditability: a governed, transparent system can show how a decision was reached (audit trails), potentially more traceable than many human trial-court decisions that aren’t fully written or explained.

Risks and concerns

  • Hallucinations and factual errors: LLMs can invent facts or misapply law; stacking agents can amplify errors.
  • Bias and unfair training data: AI can encode and reproduce systemic biases unless datasets are carefully audited and de-biased.
  • Accountability and opacity: cloud-based agentic systems may feel less accountable than human judges who have reputations and are publicly appointed/elected.
  • Power imbalance in B2C: consumer arbitration often arises from non-negotiable standardized contracts; private dispute forums can favor repeat/wealthy players unless regulated.
  • Scope mismatch: criminal prosecutions and government actions require public scrutiny; privatized AI decision-making here is inappropriate.

Safeguards AAA is using / proposed

  • Narrow roll-out: start with documents-only construction cases where outcomes are well-grounded in contemporaneous records.
  • Human arbitrators throughout: human arbitrators review and sign awards; parties can request human involvement.
  • Due-process protocols: AAA requires businesses to submit clauses for review and meet standards for consumer fairness before the provider will administer those disputes.
  • Transparent auditing and white papers: AAA is publishing audits and allowed academic review (e.g., John Choi’s engagement) to benchmark performance vs. human baselines.
  • Grounded training: agents are trained on domain-specific case libraries and a handbook tailored to the dispute type to reduce off‑domain hallucinations.
  • Iterative party validation: front-end agents repeatedly confirm facts/claims with parties so parties feel heard and mischaracterizations are caught early.

Potential use cases and roadmap

  • Immediate: documents-only construction disputes (current product).
  • Near-term candidates: supplier disputes in energy and manufacturing, payer-provider claims (hospital vs insurer), other B2B documents-only matters.
  • Additional features: early case evaluation tools for parties to assess merits before spending money; internal corporate dispute resolution.
  • Not planned (for now): live witness credibility determinations, criminal cases, government-backed public adjudication.
  • Timeline: stepwise expansion industry-by-industry over years; McCormack expects substantial change in a decade-scale horizon but acknowledges uncertainty.

Policy and ethical considerations

  • Public vs private forum: McCormack argues criminal & government cases should remain in public courts; many regulatory decisions should ensure transparency and accountability.
  • Consumer protections: regulatory fixes (Congress/FAA) or marketplace pressure could limit B2C arbitration or require stricter due‑process and disclosure.
  • Standard-setting and audits: industry standards for explainability, audit trails, independent review, and vendor neutrality will be essential.
  • Redistribution of legal work: as automation displaces routine legal tasks, legal education and professional training need redesign to prepare lawyers for new roles (oversight, complex reasoning, institutional design).

Notable quotes / insights

  • “If it were more deterministic, we would have fewer disputes.” — McCormack on predictability reducing friction.
  • “One thing AI systems can do is just make you feel heard.” — McCormack on a core procedural fairness advantage of agentic interfaces.
  • “You can de-bias a data set a lot easier than you can de-bias a human.” — argument in favor of carefully governed AI systems.
  • “There are disputes where this will never be appropriate — criminal cases or cases against the government.” — boundary-setting on public interest disputes.

Actionable recommendations (who should do what)

  • For regulators:
    • Require transparency, standardized audits, and independent third-party testing for any AI dispute-resolution systems used in consumer contexts.
    • Consider targeted rules for B2C arbitration clauses (disclosure, opt-in, due-process standards).
  • For providers and vendors:
    • Limit early rollouts to narrowly defined, documentable dispute types and publish detailed audit results.
    • Maintain human oversight, party validation steps, and an immutable audit trail of agent decisions and data sources.
  • For companies that use arbitration clauses:
    • Evaluate whether clauses meet due‑process benchmarks and whether the provider publishes clear governance/audit materials.
  • For consumers:
    • Seek providers that disclose due-process protections and auditability; push for transparency if a provider is used by your vendor.
  • For legal/tech educators:
    • Reframe legal training to emphasize oversight of algorithmic systems, interdisciplinary skills (data governance + law), and institutional design.

Short conclusion

The AAA’s AI Arbitrator is an early, cautious experiment: narrow in scope, human-supervised, and built on domain-grounded data. Its promise is real — faster, cheaper, and potentially fairer dispute resolution for many routine cases — but so are the risks (hallucinations, accountability, power imbalances). The episode frames this as a live policy and technical experiment: success requires rigorous governance, transparency, careful rollout, and clarity about which disputes should remain public.