India's judiciary is trying to answer a question most legal systems have dodged: where exactly does AI stop and judgment begin? On June 3, 2026, the Artificial Intelligence Committee of the Supreme Court of India published the draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026, and put it out for public comment. After stakeholders asked for more time, the Committee pushed the deadline from June 20 to July 15, 2026 — a 42-day consultation window, run through submissions emailed to the Committee's Member Secretary at the Supreme Court (official notice; Allahabad High Court circulation notice). The framework is expected to be finalized and notified once comments are processed.
What the Draft Actually Bars
The operative restriction is narrow and specific, not a blanket ban on courtroom AI. The draft prohibits AI from performing adjudication or sentencing without mandatory human-in-the-loop review, and it separately bars risk scoring in any form — assessing flight risk, predicting recidivism, evaluating bail eligibility, or scoring the credibility of parties and witnesses (Supreme Court Observer). Surveillance of judges, advocates, or litigants is also off-limits absent independent legal authorization. The drafting language is telling: "no judicial outcome shall be reached through automated decision-making alone," and AI systems must remain "strictly subservient to human judgment."
What It Explicitly Allows
Everything else is fair game, subject to committee approval: legal research, citation verification, document summarization, transcription and translation (with mandatory human review), cause-list preparation, docket prioritization, case management, litigant-facing chatbots, accessibility tools, and fraud detection on filings (Mondaq analysis).
The Case for the Line
The Committee isn't legislating in a vacuum. India's Supreme Court itself flagged a case this year involving AI-generated fake and hallucinated case law being relied upon in litigation — a live illustration of what happens when unverified machine output enters a courtroom. And the specific functions the draft prohibits are the ones with the worst track record globally: risk-scoring tools like the U.S. COMPAS system have been repeatedly shown to encode racial and socioeconomic bias into bail and sentencing outcomes, and the drafters appear to have written the regulation explicitly to avoid replicating that model in a jurisdiction where caste, religion, and economic status already shape unequal outcomes (Supreme Court Observer). A bright line against automated bail and sentencing, in a system already strained by backlog and under pressure to digitize fast, is a defensible precaution, not overreach.
Where Proportionality Holds — and Where It Doesn't
The substantive ban is right-sized. It targets liberty-affecting decisions specifically rather than AI use generally, which is the correct scope for a jurisdiction still building the evaluative capacity to audit these systems. Where the draft strains is on the compliance architecture wrapped around everything AI is still allowed to do. The framework creates an Apex Body chaired at the Supreme Court, five Standing Committees (Judicial, Technical, Infrastructure/Finance, Case/Data Management, Cyber Security), a High Court AI Committee in every state, a Centre of Research and Excellence on AI to test systems before deployment, and an AI Secretariat in each court led by an officer of District Judge rank or above (Supreme Court Observer). That is a heavy institutional stack to clear before a district court can deploy a transcription tool that carries none of the decisional risk the regulation is actually worried about.
The Gaps Critics Are Flagging
SFLC.in's July 15 submission to the Committee made the sharpest version of this point: the draft has no risk-based classification framework, so a chatbot answering procedural questions and a tool assisting bail-adjacent research face comparable governance friction. SFLC also flagged inadequate safeguards for sensitive judicial data and no defined mechanism for reviewing AI systems after they're updated post-approval — a real gap, since a model that passes CoRE-AI evaluation today can drift after a vendor pushes a new version. Supreme Court Observer's analysis separately notes that the mandatory disclosure requirement for lawyers using AI in filings doesn't define which tools trigger it — leaving open whether checking a citation with an AI research tool requires the same disclosure as AI-drafted argument, and whether disclosure rules will slow down urgent matters like habeas corpus petitions where hours matter.
The Better Design
None of this argues for weakening the substantive ban — that part of the draft is sound and should survive the consultation intact. It argues for tiering the governance layer to match actual risk, the way SFLC's submission urges: light-touch approval for tools with no bearing on liberty or outcome, and the full Apex Body-to-Secretariat gauntlet reserved for anything that touches case-relevant judgment, however indirectly. India's subordinate courts — where the backlog is worst and digitization is most needed — are also where a five-committee approval chain will bite hardest. A regulation that correctly protects judicial discretion from automation should not, as a side effect, protect the paperwork backlog from it too.