Beyond the Courtroom: Can India's Draft AI Regulations Shape the Future of Arbitration?

Judiciaries worldwide have increasingly sought to regulate Artificial Intelligence (“AI”) through scattered practice directions and ad hoc guidance. On 3 June 2026, the Supreme Court of India (“Supreme Court”) broke from that pattern. Its Draft Regulations for the Use of AI in Courts (“Regulations”), running into 10 chapters and 57 regulations, are the most granular attempt yet by any judiciary to separate permissible-AI-use from prohibited-AI-use in adjudication. Understandably, arbitration practitioners have started asking what, if anything, the Regulations offer their own field.

The honest answer is that arbitration should not adopt the Regulations wholesale. A court is a unitary, sovereign, publicly accountable institution – the judge, the registry and the system are one actor, so the Regulations can speak of “the Court” using AI without distinguishing user from adjudicator. Arbitration cannot follow suit. Its diverse actors use AI for different reasons, and overlooking these distinctions is the largest error arbitration could make in borrowing court concepts.

We start with what the Regulations have gotten right and what already unites the fields of litigation and arbitration: human decision-making at the centre. Under the Regulations, AI must remain “strictly subservient to human judgment”, with accountability resting “exclusively” on the human-decision-maker. The Supreme Court, in a recent judgment issued on 2 July 2026, condemned the National Company Law Tribunal and National Company Law Appellate Tribunal’s reliance on hallucinated AI-generated material “as if it were a precedent”. It affirmed that AI may be used only “in aid of adjudication”, with humans retaining total control over adjudication.

The Regulations offer helpful guardrails for the “use, deployment, or integration” of AI in any adjudicatory function of the courts. However, they provide next-to-no regulation on its use by counsel or parties, beyond an isolated disclosure rule.1 The irresponsible use of AI by users is capable of creating a butterfly effect on the outcome, as seen across the world – in December 2025, the Abu Dhabi Global Market Court penalised a law-firm for citing AI-generated authorities, and ordered payment of AED 282,508 for the waste of opposing counsel’s time in researching hallucinatory authorities, and to disincentivise misleading information. The risk is higher in arbitration, as the confidential nature of the proceedings may bury such transgressions completely.

That said, the Regulations force a question that arbitration’s own soft law, from the SVAMC Guidelines to the CIArb Guideline and the AAA-ICDR Guidance, has so far answered only in general terms: where exactly does assistance end and adjudication begin? With this question in mind, the post considers three important aspects of the Regulations: what AI may do in arbitration, what it must never do, and who actually gets to use it (and how). The aim is to set up a framework for deciding how the Regulations may be adapted, if at all. This piece does not address the confidentiality and data-governance risks raised by AI use, which arise for counsel, tribunal, and institutional use alike and merit separate treatment given arbitration's confidentiality guarantee.

Permissible uses

Unlike a court, arbitration necessitates a more tailored approach from the three distinct users of AI whose duties differ sharply.

Counsel already rely on AI for legal research, contract and disclosure review, and machine translation of multilingual records. The 2025 White & Case Survey found 77% of respondents are comfortable with AI-assisted calculation of damages, costs and interest, because that task is arithmetic. Document production is arbitration’s most mechanical stage and its most improvable: the Armesto Schedule (discussed here) already structures requests around defined, machine-checkable criteria, and the ASA User Council’s 2025 whitepaper anticipates AI screening requests before they reach a Redfern schedule (discussed here).

Tribunal use may be permissible over a narrower band, and CIArb’s template procedural order already contemplates it. Drafting a procedural timetable, a first-cut of procedural order no. 1, or the chronology of undisputed facts are administrative outputs a tribunal can generate and then adopt as its own. Summarising submissions is harder. The White & Case survey recorded 66% support for AI-use in summarising submissions, albeit with dissent insisting that summarising the facts is itself part of the cognitive process of decision-making. This objection locates the actual boundary: AI may compress an agreed record, but when compression requires choosing among competing accounts, it enters the territory of evaluation, and the SVAMC Guideline 7 duty to disclose and invite comment should then apply.

Institutional use is the least visible and the most defensible, provided it is not used for pending decisions. Case administration, docket management, conflict screening and machine translation are functions leading arbitral institutions perform manually. AI can absorb these costs without absorbing any adjudicative function. The AAA-ICDR’s pilot of an AI system trained on more than 1,500 construction awards, where an AI system drafts awards for arbitrators to review, tests exactly this limit. That kind of drafting is defensible only if the arbitrators meaningfully review the draft, rather than simply rubber stamp on what the AI produced. No court has yet had to decide whether a reviewing arbitrator's sign-off was genuine or perfunctory.

Regulation 19 of the Regulations is where tribunals and institutions can draw directly from the court model. It sets out the purposes for which AI may be used subject to written approval, including case management, transcription, translation and the like. It also permits AI for research and summarisation. The regulation is precise, forward-looking and liberal. Because it governs procedural functions, it does not carry the risk of hallucination corrupting the ultimate outcome.

Prohibited uses

The Regulations prohibit Algorithmic Decision-Making in adjudication and forbid AI from assessing witness credibility or predicting outcomes, echoing the EU AI Act’s classification of AI tools used in legal analysis and adjudication as high-risk (discussed here) – a classification the French Cour de Cassation’s working group also endorsed in April 2025 in concluding that adjudication must remain human-led. Arbitration reaches the same conclusion by its own route. In ARIHQ v. Santé Québec (discussed here), a Québec court set aside an award because the arbitrator’s reasoning rested on hallucinated authorities, holding that reliance on fabricated sources is, in substance, delegation of the decision-making, a defect treated as a serious procedural irregularity under provisions mirroring Article V of the New York Convention and Article 34(2)(a)(iv) of the Model Law.

The deeper problem is legitimacy rather than accuracy (enforcement concerns are discussed here). An award whose reasoning is probabilistic rather than deliberative may satisfy the appearance of a reasoned decision without satisfying its substance. This may trigger Article V(1)(b) wherever the parties never had the chance to contest the basis on which they were actually judged, and Article V(2)(b) wherever the enforcing court treats an opaque, undisclosed process as offending public policy – a risk  identified as early as 2020.

The prohibition should attach weight to what a tribunal does with an output, not the technology producing it. Predictive tools that model settlement value or quantify enforcement risk serve a legitimate function in deciding whether to arbitrate at all, and proposals to use predictive analytics in arbitrator-challenge analysis raise due process questions no more disqualifying than those raised by human intuition alone. A functional rule – one that bars a tribunal from treating an AI-derived credibility or probability score as a substitute for its own reasoning – travels across seats and rules better than a blanket ban on AI. It is also a rule an enforcing court applying Article V can actually police.

Institutions have started testing this functional line. Proposals for AI Ethical Review Committees – pre-award bodies that audit explainability, disclosure and the risk of improper delegation before an award issues – borrow the logic of Article 37 of the ICC Rules, which already permits institutional scrutiny of draft awards for enforceability. A parallel, multilateral track is moving in the same direction: UNCITRAL’s Working Group II devoted its February 2026 colloquium to AI in dispute resolution, and the OECD’s AI Principles, revised in 2024, supply value-based principles such as transparency, accountability, human oversight and inclusive growth that both India’s Regulations and arbitration’s soft law are converging on independently. This convergence matters because agreements reached from different institutional starting points is more durable than agreements imposed from one.

Inclusivity and access

India frames inclusivity as ‘access to justice’ for litigants unfamiliar with court procedure. Arbitration’s version of the problem is different: the barrier is price rather than procedure. The White & Case survey recorded that 18% of respondents flagged unequal access to AI tools as a risk in itself, with SMEs and counsel from developing jurisdictions unable to match the premium products available to better-resourced opponents. One respondent called this the divide between the ‘haves and the have-nots’. Layered onto arbitration’s existing cost problem (discussed here), the risk is that AI, which should narrow the price gap between well-resourced and under-resourced parties by cutting the cost of document review, translation and research, instead widens it – because only one side can afford the tool that does the cutting.

The upside is real. Machine translation lowers the cost of multilingual proceedings for institutions handling disputes across Asia, the Gulf and Latin America. The IBA’s Generative-AI Guidelines, though written for mediation, already model AI-assisted translation and language support that arbitration has yet to formalise. The downside is that training data skewed toward English-language and common-law materials risks disadvantaging submissions grounded in other legal traditions. AI literacy is also unevenly distributed among counsel, so the same tool can be used well by one side and poorly by the other without either breaching any disclosed duty. Equality of arms, arbitration’s most basic procedural guarantee, now has a technological dimension. Tribunals conducting case management conferences should treat relative access to AI as a legitimate agenda item rather than an unspoken asymmetry.

The way forward

For arbitrators: apply the ‘party, tribunal and institution’ distinction, hold the output-based line against ‘evidentiary and reasoning’ delegation, raise access asymmetry at the first case management conference and consult the parties before using AI yourself.

For counsel: treat verification of AI output as a non-delegable professional duty and disclose AI use where it is material to the tribunal’s understanding.

For institutions: publish voluntary ‘opt-in’ protocols, impact-assessment templates, draft AI use clauses, and an incident-learning repository, but resist central registries and mandatory public reporting like the Regulations prescribe.

For parties: contract for AI use in the arbitration agreement – provide for all scenarios (to the extent possible) where AI use cannot be later objected to.

This answers the fair complaint made on this Blog, that AI arbitration guidelines arrived before practice existed to justify them.

The Regulations are worth reading closely: they force the question of where assistance ends – a question arbitration must now answer for itself, rather than by importing an answer built for a public court.

• 1Regulation 20(h) provides that “no AI-generated output shall be submitted to a Court as an independent source of evidence without full and transparent disclosure of its AI-generated character”.