The Supreme Court AI Draft Re-Audits Tools Already Running

India’s Supreme Court wants every AI tool already running in a courtroom reviewed within a year of new rules taking effect. The 57-regulation draft published on 3 June is still not in force. Its first targets are the transcription bots, translation engines and research aids already inside registries, plus the lawyers who use chatbots to draft.

Chief Justice of India Surya Kant’s court put the text out through the Artificial Intelligence Committee chaired by Justice P.S. Narasimha and asked for comments by 20 June 2026. The draft applies to the Supreme Court, every High Court, district courts, tribunals and statutory bodies that decide cases. Commencement is staggered. The Chief Justice of India notifies the Supreme Court date; each High Court Chief Justice notifies the rest.

The Software Is Already Inside the Courtroom

The draft does not arrive in an empty building. A February 2026 government note described AI speech tools turning Constitution Bench arguments into near real-time transcripts, with human checks still required before the record is treated as final. The same note said the Supreme Court Vidhik Anuvaad Software, known as SUVAS, handles translation of judgments into 18 Indian languages for the e-SCR portal.

The Supreme Court Portal for Assistance in Court Efficiency, or SUPACE, is still experimental. It is built to surface facts and precedents for judges, not to propose a result. The National Informatics Centre and IIT Madras have been testing defect-flagging tools on e-filings, with access limited to a small group of Advocates-on-Record.

Private tools are further along than the court’s own research portal. The Oxford Institute of Technology and Justice, in a briefing updated as at March 2026, said Adalat AI, a legal-tech nonprofit, had been integrated in over 4,000 courts across nine states, with a stated aim of 7,500 active courtrooms by 2027. The same briefing said Nyaay AI, a commercial platform co-founded by PanScience Innovations, is used in the Supreme Court and 16 of India’s 25 High Courts.

AI TOOLS ALREADY IN INDIAN COURTS

• SUVAS: Machine translation of judgments and orders into 18 Indian languages, with human review still required.
• SUPACE: A research aid for judges that is still in an experimental stage and is not in regular use.
• Adalat AI: Real-time transcription of depositions, cross-examination and dictated orders in over 4,000 courts as at March 2026.
• Nyaay AI: A commercial suite used in the Supreme Court and 16 High Courts for tasks such as defect detection and live transcription.
• Digital Courts 2.1: Voice-to-text dictation (ASR-SHRUTI) and translation support (PANINI) inside the court’s own apps.

Regulation 41 is the clause that turns that inventory into a compliance problem. AI systems already in use when the rules commence “shall be reviewed by the AI Secretariat for compliance within a period of one year”, and the Appropriate Authority then decides what happens to any system that fails.

Every Listed Use Still Needs a Written Yes

Regulation 4 is the sentence every explainer quotes. AI in court processes “shall at all times remain strictly subservient to human judgment and judicial authority.” Judges keep the last word on law, fact and justice. That line is real. It is also the easy part.

The harder line sits in Regulation 19. Even the uses the draft likes, transcription, translation, research, chatbots, cause-list work, are “subject to prior approval in writing by the Appropriate Authority” and to named officers who must check the output. Anything not on the illustrative list needs a written yes, with reasons recorded for a grant or a refusal.

Regulation 16 tells every court to look for tools that cut delay, and says that unless proved otherwise the presumption is in favour of responsible adoption. Regulation 17 adds that, other things being equal, active adoption should be preferred over restraint. Those clauses sit in the same chapter as a written-approval gate that applies to the listed uses themselves.

WHAT THE DRAFT ALLOWS AND WHAT IT BARS

The ban on algorithms deciding cases is written as non-derogable. No authority under the draft, including the power to approve new uses, can relax it. Opaque systems are barred wherever they may materially affect lawful rights or personal liberty. That is the constitutional fence. The operational fence is the signature required before a district court keeps the transcription tool it already has.

Who Pays When a Citation Does Not Exist?

Lawyers may use AI to prepare pleadings. They must say so. Regulation 43(3) requires a declaration in a prescribed form that the material has an “AI-assisted character.” The court can then demand the system used, how far it helped, and what checks were done. If a filing is false or misleading because of AI output, the person who filed it “shall bear full responsibility” and cannot blame the model.

The Supreme Court Advocates-on-Record Association, the body for lawyers who can file in the apex court, told the committee that blanket disclosure is unworkable. Advocates are already bound by the Advocates Act, 1961, and the Supreme Court Rules, 2013, SCAORA wrote, and “the ultimate filing remains the sole responsibility of that advocate, who personally vouches for the accuracy of its contents.”

Requiring lawyers to constantly file declarations for using everyday software is unworkable as it can prejudice the party or the presiding Judicial Officer.

Supreme Court Advocates-on-Record Association, comments on the draft regulations

SCAORA’s alternative is a line in the existing special-leave-petition certificate confirming that every citation and statute has been personally checked. That would treat the hallucination problem as a verification duty, not a labelling duty. The association also warned that Regulation 20(1)(c), which bars AI from adjudication “without mandatory Human-in-the-Loop,” can be read as allowing a machine to sentence if a person is parked nearby. “Mere presence of a human cannot be allowed to legitimise AI-based adjudication,” it wrote.

That fight is the one that will show up in registries first. Fake citations are already a live nuisance. A disclosure stamp on every AI-touched paragraph will either become a ritual, or it will be used to discount a brief before anyone reads the law in it. SCAORA’s fear is the second outcome. The draft’s authors are writing for the first, after benches have had to chase judgments that do not exist.

Private Vendors Face a Data-Ownership Gate

Chapter VI is where the companies that already sit in courtrooms meet a new contract. No private vendor may “undertake, participate in, or provide any service in connection with an AI System deployed in Court processes” without prior written approval. Bids are to be scored on technical skill, legal compliance, ethics, data security and money.

The mandatory clauses are the point. Court data and AI outputs must have clear ownership and access terms. Sensitive judicial data cannot be used beyond the job. The AI Secretariat gets audit and inspection rights, including over underlying data. Models cannot be retrained or fine-tuned on court data without written approval from the AI Committee. Systems that handle sensitive judicial data must run on-premise or on a sovereign cloud. Vendors must indemnify the court. If a tool is built mainly with court data or public resources, the court keeps ownership or a perpetual royalty-free licence, and no private party can claim exclusive intellectual property.

There is a fast lane. The AI Secretariat may grant approval within 30 days for a tool used only for administrative work, that does not touch personal data of parties, that does not affect adjudication, and that is functionally similar to a tool already approved. That path is built for clones of listing and notice software. It is not built for a transcription engine that hears witness names in an atrocity trial.

Industry body Nasscom asked the court to define “high-risk applications” and to make clear that technical audits will not require handing over source code. Software Freedom Law Centre, India, separately flagged the absence of a risk-based classification and of a set review when a model is updated. Those are vendor problems dressed as legal ones. A transcription nonprofit that is already in thousands of rooms now has to prove it can survive an audit clause written for a full-stack supplier.

XKDR’s Case Against a Tech-Shaped Institution

The draft builds a permanent Apex Body at the Supreme Court, five standing committees (judicial, technical, infrastructure and finance, case and data, cyber security), a Centre of Research and Excellence on Artificial Intelligence, an AI Committee and an AI Secretariat in every High Court, an AI register, an incident database, annual transparency reports, and an AI Content Verification Authority. High Courts would have to publish the systems they use, audit results and incidents.

Pavithra Manivannan and Supriya Sankaran, writing for XKDR Forum in comments dated July 2026, listed four structural concerns with the proposed structure: overlapping mandates, weak continuity, judges spending time on work that needs specialists, and high costs for anyone trying to ship a tool. Their spine sentence is blunt. “Institutional structures must not be built around individual technologies.” Computerisation, video hearings and AI, they wrote, are successive tools in the same change to court administration, and the judiciary needs one standing body that can absorb the e-Committee’s work plus whatever comes next.

The second gap they named is standards. The draft says who approves a system. It does not say what error rate is tolerable in a scheduling tool versus a research summary, what a system must be able to explain, or where training data may come from. Without those numbers, each of the 25 High Courts invents its own test. Capacity is not even across those courts. Duplicate evaluations of the same product are the predictable result, which is also the complaint SCAORA made about a “pyramid type of governance system” that makes it hard to fix accountability.

SCAORA asked that its president, or a nominee, sit as a permanent member of the Apex Body. The draft already lets the Chief Justice nominate one or more advocates with technology or privacy experience. Whether the Bar gets a standing seat is still open, because the text is still a draft.

A 5.64-Crore Pile the Draft Cannot Hear

Phase III of the eCourts project was approved on 13 September 2023 with an outlay of Rs. 7,210 crore. The draft is the judicial rulebook meant to sit on top of that spend. It cannot appoint judges, and it cannot hear a trial.

Arjun Ram Meghwal, Minister of State for Law and Justice, told the Rajya Sabha, using National Judicial Data Grid figures as on 16 July 2026, that the Supreme Court had 96,024 pending matters, the High Courts 64,72,536, and district and subordinate courts 4,98,45,538. Added, that stack is 56,414,098 matters, which the minister put at 5.64 crore. Of the Supreme Court pile, 10,094 matters were more than ten years old. High Courts held 80,660 matters more than thirty years old.

PENDING CASES ON THE NATIONAL JUDICIAL DATA GRID, 16 JULY 2026

The uses the draft blesses, listing, translation, transcripts, defect flags, are the ones that eat staff time. They do not dispose of a suit. Docket prioritisation is allowed as an administrative aid, which is as close as the text comes to touching the queue. Risk scoring for bail and reoffending is banned. That is a choice to keep delay rather than import the COMPAS-style tools that other systems tried and then had to defend. The backlog will still be a judge-and-vacancy problem after the last chatbot is certified.

England Wrote Guidance, India Wrote a Code

India is not the first court system to tell judges that a chatbot is not a colleague. The Courts and Tribunals Judiciary of England and Wales issued updated judicial AI guidance in October 2025, replacing an April 2025 note. Lord Justice Colin Birss, the Lead Judge for Artificial Intelligence, put the duty on the human who signs the work.

The use of AI by the judiciary must be consistent with its overarching obligation to protect the integrity of the administration of justice and uphold the rule of law.

Lord Justice Colin Birss, Lead Judge for Artificial Intelligence, Courts and Tribunals Judiciary, 31 October 2025

The British document is guidance for judicial office holders, their clerks and staff. It warns against pasting private facts into public tools and treats AI output as something a judge must check. It is not a 57-clause code with secretariats in every High Court. United States practice is patchier still: individual federal judges have issued standing orders on disclosure after fabricated citations, without a single national court-AI statute.

India’s draft is heavier because it tries to govern vendors, litigants, registries and 25 High Courts in one instrument, and because it would run on top of tools that are already live. That is the second-order bet. A code can stop a bad bail algorithm that was never deployed. It can also slow a transcription engine that is already taking dictation in a trial court while the Apex Body, five committees and a Centre of Research work out the minimum mandatory standards.

THE DRAFT’S PATH SINCE JUNE

1. 3 June 2026: The AI Committee publishes the draft and invites comments to office.regcc@sci.nic.in.
2. 20 June 2026: The notice’s deadline for stakeholder comments.
3. July 2026: SCAORA’s executive committee presents its comments to Justice Narasimha and Justice Alok Aradhe; Narasimha says the draft will be deliberated in the coming weeks.
4. 18 July 2026: Chief Justice Surya Kant tells reporters the AI regulations are on the Supreme Court website and that use will expand “in a regulatory manner.”
5. 13 September 2026: Supreme Court judge Justice N.V. Anjaria, speaking at a digital ADR event in Gandhinagar, says AI “can be part of the justice delivery system, but cannot be the decision-maker.”

The one-year review clock does not start on publication. It starts when the Chief Justice of India, and then each High Court Chief Justice, notifies the rules. Until those dates are gazetted, Adalat can keep typing, SUVAS can keep translating, and a lawyer who files a chatbot citation still answers to the professional codes already on the books, plus whatever a particular bench decides to do with the brief.

Disclaimer: This article is news reporting and analysis of a draft issued by the Supreme Court of India’s Artificial Intelligence Committee. It is informational only and is not legal advice, a practice guide, or a substitute for reading the draft text and any later notified version. Readers who must decide how to file, procure, or deploy AI in a court process should consult a qualified advocate or the relevant court’s registry before acting. Figures, quotes and the draft’s status reflect the official notice, the National Judicial Data Grid snapshot used in Parliament, and other sources cited in the piece as those documents stood on the dates given, and all of them can change if the committee revises the text or a Chief Justice notifies a commencement date.