Governing the black box in adjudication: Policy frameworks for algorithmic transparency in international arbitration

Generative artificial intelligence (GenAI), especially large language models (LLMs), is entering international arbitration as a research, analytics, document-review and drafting aid. The core instruments governing arbitral legitimacy remain technology-neutral, yet GenAI creates risks that ordinary procedural baselines do not operationalise: fabricated authorities and record assertions, automation bias, cognitive offloading, confidentiality exposure and uneven access to advanced tools. Using doctrinal and comparative legal analysis, this article treats the New York Convention and the UNCITRAL Model Law as the lex lata enforceability baseline, and advances the Algorithmic Disclosure Protocol as a de lege ferenda procedural framework. The Protocol adapts the EU AI Act's risk-tiered logic to arbitration's constraints of party autonomy, confidentiality, deliberative secrecy, due process, reasoned decision-making and finality. At the meso level, institutions adopt standard AI-use schedules, security defaults and competence expectations. At the micro level, tribunals apply a substantive-influence test that escalates disclosure and verification duties as AI use moves from administrative assistance to reasoning-sensitive functions. At the macro level, parties receive narrowly scoped, cure-oriented auditability rights, including reliance statements, preservation of verification artefacts and cost-sensitive remedial powers. Arbitration can use GenAI without accepting black-box adjudication if material AI influence is disclosed, verified and contestable before enforcement is jeopardised.