Introduction

The integration of artificial intelligence into legal proceedings has moved from theoretical speculation to practical experimentation. In 2023, several arbitral institutions began piloting AI-assisted document review, and by 2024, at least one jurisdiction had formally permitted the use of large language models in preliminary procedural rulings. This article examines whether the next logical step — AI as a decision-maker in commercial arbitration — is legally permissible, practically desirable, and epistemically sound.

The Due Process Threshold

Arbitration derives its legitimacy from the parties' consent and the arbitrator's capacity for reasoned, impartial adjudication. Any substitution of the human arbitrator by an algorithmic system must clear a high threshold: the resulting award must be capable of enforcement under the New York Convention and domestic arbitration statutes, all of which presuppose a human or institutional decision-maker accountable under law.

The core challenge is not computational capacity but rather epistemic transparency. An AI tribunal must not only reach a defensible outcome; it must articulate a reasoning chain that satisfies the requirement of a reasoned award. Current large language models can produce plausible legal prose, yet their internal processes remain opaque — what practitioners have begun calling the black-box problem.

Consent and Party Autonomy

Party autonomy is arbitration's foundational principle. If both parties expressly consent to AI adjudication — with full disclosure of the model architecture, training data, and conflict-screening methodology — the autonomy argument weighs in favour of permissibility. The harder case arises in adhesion contracts where arbitration clauses are non-negotiable: can consent to arbitration reasonably be construed as consent to algorithmic adjudication?

This author's view is that it cannot. The right to a human decision-maker is not merely procedural — it reflects the parties' legitimate expectation that their dispute will be understood by a reasoning agent capable of grasping context, nuance, and the equitable dimensions that purely rule-based systems systematically underweight.

A Proposed Framework

Rather than wholesale adoption or rejection, a tiered framework deserves consideration:

  • Tier 1 — AI as assistant: Document review, translation, scheduling, preliminary procedural rulings on uncontested matters. Widely permissible today.
  • Tier 2 — AI as co-arbitrator: A hybrid panel in which one seat is occupied by an AI system, with human arbitrators retaining veto power over the final award. Permissible with express party consent.
  • Tier 3 — AI as sole arbitrator: Reserved for low-value, high-volume disputes (consumer arbitration, small commercial claims) where speed and cost reduction justify reduced procedural safeguards. Requires specific statutory authorization.

Enforcement Risk

Even where parties consent and institutional rules permit, the enforcing court retains discretion to refuse recognition on public policy grounds. National courts in France, England, and the United States have consistently held that procedural fairness requires a decision-maker capable of being held to account. Until AI systems can be subjected to meaningful accountability mechanisms — professional discipline, liability, disqualification — enforcement risk at Tier 2 and Tier 3 remains substantial.

Conclusion

Artificial intelligence will reshape arbitration profoundly, but the reshaping will be evolutionary rather than revolutionary. The appropriate role for AI in the near term is augmentative: enhancing the capacity of human arbitrators without displacing their judgment. The question is not whether algorithms can replicate the outputs of legal reasoning, but whether they can be trusted — by parties, institutions, and courts — to bear the normative weight that an arbitral award must carry. On current evidence, that trust has not yet been earned.