On June 3, 2026, the Artificial Intelligence Committee of the Supreme Court of India published a preliminary draft of the "Regulations for Use of Artificial Intelligence in Courts, 2026," inviting public comment. The original window closed June 20; a follow-up notice extended it to July 15, 2026 — meaning the draft is still, as this is written, open to revision. That matters, because the single most consequential clause in the 35-page document is also its most legally fragile.
What the draft actually does
The regulations apply to the Supreme Court, every High Court, and all subordinate courts, tribunals and statutory commissions performing adjudicatory functions. They rest on five principles — human primacy, transparency, accountability, data protection and judicial independence — and draw a hard line between administration and adjudication. AI may assist with legal research, citation verification, drafting, translation, transcription, scheduling and case management. It may never decide a case, recommend bail, determine a sentence, evaluate witness credibility, predict a litigant's future conduct or influence judicial deliberation.
Layered on top is a liability regime. Regulation 8 makes every AI-assisted judicial or administrative decision "advisory," attributed entirely to the officer who acted on it; that officer cannot invoke a black-box system or an AI hallucination as a defense. Regulation 43(6) extends the same no-fault standard to lawyers and litigants: submit an AI-fabricated citation, and "reasonable care" taken to verify it beforehand is not, under the current draft language, a valid defense.
The case for zero excuses
The Court's instinct here is not paranoid. Fabricated case law generated by large language models has already reached real courtrooms — most famously in the 2023 Mata v. Avianca sanctions in the Southern District of New York, where a $5,000 fine followed a brief citing six cases that did not exist. India's own courts, already straining under roughly 50 million pending cases nationwide, cannot afford a wave of appeals built on invented precedent. A rule that lets "the model hallucinated" function as an excuse would gut the incentive to check anything at all, and in a system where litigants — often unrepresented, always unequal in resources — have no way to audit a judge's or opposing counsel's AI workflow, some hard floor on accountability is a legitimate ask.
Where the draft overreaches
The trouble is that Regulations 8 and 43(6) collapse two very different failures into one standard. A lawyer who pastes an unverified AI output straight into a filing is negligent and should answer for it. A lawyer or judicial officer who uses a system the Apex Body itself approved under Regulation 22's audit and impact-assessment process, checks the output against available sources, and is still misled by an undetectable fabrication has exercised exactly the diligence the regulations elsewhere demand — yet the draft, as written, denies them any defense at all. That is strict liability in the classic sense: untethered from fault, and imposed on the person least able to control the underlying failure, since the vendor whose model was certified for court use faces no comparable exposure under this draft. Legal commentary on the text has already flagged this as vulnerable to an Article 14 arbitrariness challenge, and separately notes the regulations name who is liable, in Regulations 20 and 46, without specifying penalties or a remedy mechanism — a gap one analysis calls "the liability void." The Bar Council of India's position — that AI has no legal personality under the Advocates Act, 1961, so liability must fall on the human professional — is correct as far as it goes, but it doesn't resolve which humans should bear how much of it, or under what standard.
A global outlier
Compare this to the EU AI Act, which classifies AI used to help judicial authorities research and interpret law as high-risk, but places most of the compliance burden — risk management, technical documentation, human-oversight design — on the provider, with obligations for high-risk systems phasing in from August 2, 2026. US courts, in the years after Mata, have overwhelmingly responded with certification and disclosure rules enforced through existing fault-based sanctions regimes, not blanket strict liability on whoever happened to be holding the pen. India's draft is unusual in placing no-fault liability directly on the individual end user rather than distributing it toward the parties with more actual control over whether a tool hallucinates: the Apex Body that certifies it, or the vendor that built it.
Fix the liability, not the ambition
None of this is an argument against regulating AI in courts. A judiciary carrying 50 million pending cases has every reason to want faster research and translation tools, and that backlog will not shrink by refusing every efficiency gain out of caution. But a liability rule that punishes diligence exactly as harshly as recklessness will do the opposite of what the Court intends: it will push careful judges and lawyers away from Apex Body-vetted tools altogether, while doing nothing to stop the reckless filers who never verified anything to begin with. With the comment window open until July 15, the fix is straightforward — pair the no-hallucination-excuse principle with a genuine safe harbor for demonstrated, documented reasonable care on court-approved systems. The Court has a few weeks left to build that in before "advisory" liability hardens into strict liability by default.