The Supreme Court draft regulations on AI use can be accessed here.
The Supreme Court of India has published draft regulations on the use of artificial intelligence (AI) in courts, dated June 3, and has invited public comments by June 20. The regulations apply to every court in India, from the Supreme Court down to tribunals and statutory commissions.
The core principle: AI can assist courts, but it cannot replace judicial decision-making. Every AI output serves only an advisory role, and final authority over law, facts, and justice rests exclusively with the judge.
What AI can do in courts:
- Case management: Scheduling, cause-list preparation, docket prioritisation, and defect identification in new filings.
- Transcription: Of court proceedings, where a human must review and certify the output.
- Translation: Translating judgments, orders, and pleadings, subject to human verification.
- Legal research: Retrieving precedents, verifying citations, and summarising documents.
- Administrative tasks: Assisting with case filing, record management, and the auto-generation of notices and summonses.
- Litigant assistance: Powering chatbots that help litigants understand court procedures, under human oversight.
- Accessibility: Supporting text-to-speech, speech-to-text, Braille translation, and visual-assistance tools.
- Anonymisation: Anonymising judgments and records for publication in the public domain.
- Fraud detection: Detecting fraud in administrative processes, subject to human review.
What AI absolutely cannot do: These prohibitions are absolute and non-derogable. No AI system can:
- Reach a judicial outcome through algorithmic decision-making (ADM) alone, meaning a process in which an algorithm, rather than a human, makes the determination.
- Perform risk scoring, including predicting reoffending probability, bail eligibility, or flight risk.
- Operate as a black box, meaning a system whose decision-making logic cannot be explained, in any process affecting personal liberty or legal rights.
- Conduct surveillance of judges, lawyers, or litigants in connection with court premises.
- Profile or predict the future behaviour of any party, witness, or legal representative.
- Enter the record as independent evidence unless the submitting party fully discloses its AI-generated nature.
- Use personal data for training unless approved by the appropriate authority.
- Compromise the confidentiality of judicial deliberations.
The disclosure requirement: Lawyers who use AI to prepare any pleading, document, or evidence must declare it at the time of submission. Courts that use AI in case management must inform the parties. Anyone using synthetic data, AI-generated audio, visual, or text content that mimics real data, must also disclose its use.
The institutional structure:
- Apex Body at the Supreme Court: A permanent, full-time body that governs AI across the judiciary. It will comprise two Supreme Court judges, two Chief Justices of High Courts, two High Court judges, a Joint Secretary from the Ministry of Electronics and Information Technology (MeitY), and experts in cybersecurity, finance, and technology law. It will set minimum mandatory standards for all courts.
- AI Committees at every High Court: These committees will comprise judges responsible for approving AI systems, monitoring compliance, and overseeing the AI Secretariat. They must meet at least once every three months.
- AI Secretariat at each High Court: Headed by an officer of district judge rank, the Secretariat will maintain the AI Register and incident database, conduct audits, and handle approvals.
- Centre of Research and Excellence on AI (CoRE-AI): This body will conduct research, evaluate AI tools, track international jurisprudence, and publish white papers.
- AI Content Verification Authority: This authority will oversee the verification of Generative AI (GenAI)-generated content.
Before deployment: Every AI system must clear a Technical and Ethical Impact Assessment covering its architecture, training data quality, risks of bias and hallucination, cybersecurity vulnerabilities, explainability, and incident reporting mechanisms. The appropriate authority must prescribe a standard assessment format within six months of the regulations coming into force.
Ongoing oversight:
- Every court must maintain an AI Register documenting approved systems, their scope, vendors, and audit outcomes.
- Every AI system must undergo technical, legal, and ethical audits at least once a year.
- Each AI Secretariat must maintain an AI Incident Database tracking malfunctions and biases, and share learnings across jurisdictions.
- Every High Court must publish an Annual Transparency Report.
On private vendors: No private entity can participate without prior written approval. Vendor agreements must:
- Prohibit vendors from using sensitive judicial data beyond the scope of the engagement.
- Prohibit vendors from retraining or fine-tuning models on court data without written approval from the AI Committee.
- Require on-premises or sovereign-cloud deployment for sensitive judicial data.
- Vest ownership of any tool built on judicial data in the court and prohibit vendors from claiming intellectual property rights over tools developed primarily using judicial or public resources.
- Prevent vendors from sharing source code and training data with third parties, except for authorised audits.
On data protection: Sensitive judicial data cannot leave court systems without written authorisation. Courts must anonymise personal data before using it for training, to the extent feasible without compromising utility, and should prefer systems that require less data processing.
Grievance redressal: Any party harmed by a prohibited use of AI may apply to the court where the system was used. That court must hear the matter and pass appropriate orders.
The retraining question: AI training data versus the right to be forgotten
On May 29, 2026, the Delhi High Court delivered a 144-page judgment directing Google and Indian Kanoon to de-index name-based search results for petitioners in cases ending in acquittal, settlement, or discharge. The judgments remain accessible through case numbers and citations, but the platforms must remove the individuals’ names from search results.
The draft regulations require courts to anonymise personal data before training AI systems going forward. However, they do not address systems that have already been trained on judicial data that is later ordered to be de-indexed.
Key questions remain:
- What happens to a system that already trained on a judgment before a court orders it de-indexed or a name anonymised?
- Does the draft provide a mechanism for a party who has obtained a de-indexing order to remove their data from an existing model’s training set?
- If a court’s legal research tool surfaces a name that a court has ordered removed, what steps must the court take? De-indexing orders against Google and Indian Kanoon do not affect the training data of models already in operation within courts.
The Delhi High Court judgment already covers more than 30 petitions filed since 2016, while the Supreme Court is separately examining the broader scope of the right to be forgotten.
Other open questions:
- Anonymisation versus precedent retrieval: The draft allows AI systems to retrieve and cite past cases, a function that often depends on identifiable names. At the same time, it directs courts to anonymise training data wherever feasible. Can legal research tools effectively surface named precedents if identifying information is removed from training datasets?
- In-house audit capacity: The draft requires courts to conduct AI audits in-house and restricts the sharing of source code and data outside court premises for audit purposes. Do courts possess the technical expertise necessary to evaluate complex private AI systems, particularly opaque or black-box models?
- The verification dispensation: The draft generally requires officers to verify AI outputs before use. However, it allows the responsible officer to waive verification “for reasons to be recorded in writing” and exempts certified administrative tools altogether. Could these exceptions become routine rather than exceptional?
- Sole accountability versus vendor liability: The draft places accountability for AI-assisted decisions solely on the responsible officer while also requiring vendor agreements to allocate liability between courts and vendors. If harm results from a defective vendor-supplied model, who ultimately bears responsibility, the officer, the court, or the vendor?
Also read:
