Inside SC’s proposed regulations for AI use in courts: What’s allowed, what’s absolutely barred

Is this binding?

Not automatically, and not uniformly. The Draft Regulations will come into force for the SC on any date notified by the Chief Justice of India. For High Courts and courts and tribunals under their jurisdiction, they will come into force separately on the dates notified by that High Court’s Chief Justice. Different provisions can be brought into force on different dates, meaning that AI adoption can be further phased to suit the courts’ requirements and circumstances.

What role for AI has the Supreme Court carved out?

The Draft Regulations propose to require courts to “actively seek opportunities” to deploy AI systems and tools that “demonstrably” improve access to justice, reduce delays, or enhance administrative efficiency.

AI use is explicitly permitted for a range of administrative and assistive functions, including case management, transcription, translation, legal research, document summarisation, accessibility, and court administration. All of these will require approval in writing by the Apex Body of the SC, or AI Committee of the high court/tribunal, as the case may be, and supervision and verification of officers nominated by the court.

Can AI decide a case?

No. The Draft Regulations are categorical that no judicial outcome shall be reached through algorithmic decision-making alone, or solely on the basis of AI-generated information. Human judicial authority is determinative in all adjudicative decisions. Therefore, the use of AI in the decision-making process related to a case, if any, will only be advisory, and subject to independent human (judicial) evaluation.

What is off-limits?

Several uses are prohibited in “absolute and non-derogable” terms, meaning no authority can later permit them under the Regulations. These include: ‘risk scoring’ to assess flight risk; predicting recidivism; evaluating bail eligibility; determining the credibility of witnesses; using AI to predict, profile, or infer the future conduct or behaviour of parties, accused persons, witnesses, or legal representatives more generally; submitting AI-generated output as independent evidence without full disclosure of its AI-generated character; and using blackbox (unexplainable) AI systems in matters affecting personal liberty.

Will litigants know if AI was used for their case?

Yes, the Draft Regulations propose that if a court uses an AI tool to “materially assist” it in any aspect of case management, document analysis, or judicial administration, it will be required to inform the parties (litigants and their counsel) in a timely and accessible manner. The standard is material assistance; therefore, litigants will not be informed of all instances where AI has been used in their case, but only when AI has provided material assistance to the judge.

Who will regulate?

The institutional architecture will be led at the highest level by an “Apex Body” at the SC, composed of sitting SC and High Court judges, an official of the Ministry of Electronics and Information (MeitY), and experts in finance and cybersecurity. This body will set out the minimum mandatory standards approving of the use of AI systems and issuing implementation guidelines. It will work through five specialised committees; the SC and each high court will constitute their own AI Committees, backed by an AI Secretariat.

A separate research body, the Centre of Research and Excellence on Artificial Intelligence (CoRE-AI), will evaluate tools and track developments to support the Apex Body.

Will private companies be involved?

Yes, but only with written approval from the relevant court authority, and subject to a mandatory list of contract terms. Vendor agreements must cover ownership of and access rights to court data and AI outputs, a bar on using sensitive judicial data, and an explicit prohibition on retaining or fine-tuning models using court data without the AI Committee’s written approval.

The Draft Regulations also state that vendors cannot claim exclusive IP over tools built substantially on judicial data or public resources.

What safety measures are in place?

The Draft Regulations take a lifecycle approach for safety measures: before, during, and after the deployment of the AI system. A Technical and Ethical Impact Assessment will be required, covering the system’s architecture, training data, bias, hallucination risks, and cybersecurity posture. Some systems may first be required to go through “Controlled Environment Testing” in an isolated setup. Upon deployment, systems will face technical, legal, and ethical audits, as well as separate cybersecurity audits on the same cycle. These are to be conducted in-house, since source code and training data can never be shared with a third-party for audit purposes.

Each Court will also be required to maintain an AI Register documenting approved systems and audit outcomes, while each AI Secretariat will maintain an AI Incident Database logging malfunctions, errors and biases, with a mandatory 24-hour notification requirement if a tool fails or is suspended. Each high court will also be required to maintain an emergency fall-back protocol to keep court processes running manually if a system goes down.

What recourse does a litigant have if the AI used in their case causes harm?

The grievance route will apply specifically where harm results from a prohibited use of AI (explained above). An affected party can file an application at the earliest opportunity with the court where the relevant AI system was or is being used. That court must give the affected party a hearing and pass “appropriate orders as it may deem fit”.

High Courts may separately lay down simplified procedures and complaint formats to make this accessible to people with limited legal literacy. This stream of grievance redressal will sit alongside any other legal remedy that may be already available under ordinary law.

The authors are Research Fellow and Senior Resident Fellow in the Applied Law and Research Centre at the Vidhi Centre for Legal Policy.