In June, the Supreme Court proposed the first rules for how artificial intelligence may be used inside India's courts. The rules arrived because something had already gone wrong.
A Rule That Followed a Problem
Most regulation arrives before a technology spreads. This one arrived after. By the time the Supreme Court published its draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026 on 3 June, courtrooms across the country had already spent more than a year dealing with a problem the rules were written to address: legal filings full of cases that do not exist.
The pattern became familiar to judges. A petition would arrive with citations that looked entirely legitimate, formatted like a standard Supreme Court Cases or All India Reporter reference. A judge would try to locate the reasoning, only to find the judgment was never decided. The cause was rarely deliberate forgery. It was generative AI, used for research and copied into a brief without anyone verifying that the cases were real.
In January 2026, the Bombay High Court imposed a cost of fifty thousand rupees on a party for placing fabricated case law into written submissions, noting that the filing carried the obvious signs of raw AI output. In September 2025, a petition before the Delhi High Court was withdrawn after opposing counsel showed that several cited precedents were fictitious, including references to paragraphs of a judgment that contained far fewer paragraphs than claimed. These were not isolated embarrassments. They were becoming routine.
From Mistake to Misconduct
The turning point came on 27 February 2026. Taking suo motu notice of a trial court order from Andhra Pradesh that had relied on four fabricated judgments, a Bench of Justice P.S. Narasimha and Justice Alok Aradhe drew a line that the legal profession had not heard before. A decision built on non-existent and fake judgments, the Court said, is not an error in the decision-making process. It is misconduct, and legal consequences follow.
"A decision based on such non-existent and fake alleged judgments is not an error in the decision-making process. It would be a misconduct and legal consequence shall follow." — Supreme Court of India, 27 February 2026
That framing mattered. An error invites correction. Misconduct invites accountability. The Court issued notices to the Attorney General, the Solicitor General and the Bar Council of India, and appointed Senior Advocate Shyam Divan as amicus curiae to help examine the wider implications. The message to lawyers was direct. The defence that the AI made it up, and the person filing did not know, would no longer be accepted.
High Courts moved in parallel. On 4 April 2026, the Gujarat High Court issued a policy that confined AI with unusual precision, barring it from judicial reasoning, evidence handling and drafting, requiring independent verification of any AI-generated reference, and fixing personal responsibility on judges and staff. The direction of travel was clear well before the Supreme Court's draft arrived.
What the Draft Rules Actually Say
The draft Regulations, prepared under the Supreme Court's AI Committee and opened for public comment until 20 June 2026, try to do something more lasting than penalise individual lapses. They set boundaries for the entire adjudicatory system, the Supreme Court, the High Courts, district courts, tribunals and statutory commissions.
At the centre sits one principle the drafters call human primacy. AI systems may function only in an assistive capacity. They can never replace judicial decision-making. The draft states that no judicial outcome, or any part of it, may rest solely on algorithmic decision-making or on AI-generated information. Where AI output is used at all, it is to be treated as advisory, and reasonable care must be taken to verify its accuracy.
┌────────────────────────────────────────────────────────────┐
│ WHAT THE DRAFT RULES PROHIBIT OUTRIGHT │
│ │
│ • AI deciding any case, or any part of a case │
│ • AI-driven bail or risk assessment of an accused │
│ • Behavioural profiling of individuals │
│ • Algorithmic prediction of case outcomes as a basis │
│ for decisions │
└────────────────────────────────────────────────────────────┘
Alongside the prohibitions, the draft sets out where AI may help: research, drafting assistance, translation and transcription, all subject to supervision and written approval by the appropriate authority. The framework is built on a stated philosophy its authors describe as innovation over restraint, encouraging responsible adoption rather than a blanket ban. The point is not to keep technology out of the courtroom. It is to keep judgment human.
For practitioners, the most immediate change is disclosure. The draft contemplates a record of how AI was used in a filing, including which tools were used, the underlying sources, and the human review involved. In plain terms, a lawyer who uses a tool to prepare a document will be expected to say so and to stand behind every citation in it. Responsibility for the content stays with the person, not the software.
The Machinery Behind the Rules
The draft does not stop at principles. It proposes an institutional structure to make them work: a permanent Apex Body at the Supreme Court level, made up of judges, technologists, cybersecurity and finance experts, practitioners and government representatives, to set national standards, approve AI systems, and publish annual governance reports. Below it sit specialised committees, an AI register, an incident database, a content verification authority and a grievance process.
This is worth pausing on. The courts are not only telling lawyers to be careful. They are building permanent administrative plumbing to supervise a technology that did not meaningfully exist in courtrooms three years ago. That tells you how seriously the problem is being taken, and how quickly it grew.
The judiciary is not approaching this from a standing start. The Supreme Court already runs SUVAS, its translation software, which has rendered tens of thousands of judgments into regional languages, and TERES, which transcribes Constitution Bench hearings in real time. The institution is comfortable using AI as a tool. What it is now doing is drawing a hard boundary between the tool and the decision.
Why This Reaches Beyond Lawyers
It would be easy to read this as an internal professional matter, a question of how advocates prepare briefs. It is more than that. The same hallucination problem that produced fake citations in court is the problem ordinary people now face every time they ask a chatbot a legal question and act on the answer.
A tenant researching eviction rights, a small business owner checking a contract clause, an employee reading up on notice periods, all increasingly turn to AI tools that sound confident and are sometimes wrong. The court's response sets a useful standard for everyone, not only the bar. Treat the output as a draft, not an answer. Verify against the actual source. Remember that the responsibility for acting on information stays with the person who acts, not the system that produced it.
There is also a deeper question the rules only begin to touch. As AI-generated text, images, audio and video become harder to distinguish from the real thing, the reliability of evidence itself comes under pressure. A fabricated citation is one problem. A fabricated recording or a synthetic document is another, and courts will face both. The draft rules are a first attempt to hold a line that technology will keep testing.
Where Things Stand
The consultation window closed on 20 June. What follows is the slower work of turning a draft into a working framework, refining how disclosure should operate, how verification is enforced, and how the Apex Body functions in practice. None of that will be settled quickly, and the technology will not wait for it.
But the central judgment has already been made, and it is a conservative one in the best sense. The Indian judiciary has decided that AI belongs in the courtroom as an assistant and nowhere near the seat of decision. A machine can help a judge read faster. It cannot be the judge. A machine can help a lawyer draft. It cannot take responsibility for the draft.
That distinction sounds obvious when stated plainly. The reason it had to be written into a rule is that, for a worrying stretch of months, a number of people in the system had stopped observing it.