THE SUPREME COURT OF INDIA’S Draft Regulations for Use of Artificial Intelligence in Courts, 2026 (‘Draft Regulations’) are based on five principles: human primacy, transparency, accountability, data protection, and judicial independence. These principles visibly translate through the Draft Regulations in the general principles which deal with governing AI systems, permissible and prohibited use of AI, institutional architecture, oversight architecture, procurement, data protection, training programmes, and grievance redressal.
The Draft Regulations use a careful vocabulary with their framers clearly alive to what can go wrong when AI enters adjudication. But what they do less well, and what this piece argues, is reckon with what it will actually take for an Indian judiciary of uneven technical capacity, spanning the Supreme Court down to tribunals and statutory commissions, to deliver on what they propose.
This piece does not highlight what the Draft Regulations say in its entirety, which is accounted for in my companion explainer. It follows what assumes that reading and asks a difficult question: whether the Indian Supreme Court’s Draft Regulations on the use of AI in courts have the institutional weight to uphold the promises it makes?
The infrastructure question precedes the regulatory one
Every provision under the Draft Regulations creates certain obligatory functions, assuming the required capacity to discharge them. It is largely silent on where that capacity will originate from. For instance, Regulation 38 of the Draft Regulations requires periodic in-house technical, legal, and ethical AI audits. This sounds like a thoughtful security position, but it also presumes that AI secretariats will possess the expertise to audit AI systems unaided.
District courts in several states are still building reliable case management software. At this backdrop, the proposal of initiating in-house AI auditing is not simply a small requirement. It does not say about where the capacity of in-house AI auditing will come from; all that it says is that AI auditing is a must.
Who governs the Apex Body?
The Draft Regulations envision an oversight structure comprising an Apex Body, AI Committees, AI Secretariats, and a Centre of Research and Excellence on Artificial Intelligence. This tier-ed structure, in compact form, has the power to approve, audit, direct remedial measures, and suspend a non-compliant system.
By the Draft Regulations, the Apex Body at the Supreme Court shall set the minimum mandatory standards for AI use across all courts in India. Under Regulation 57, the Supreme Court AI committee has the task of periodically reviewing these Draft Regulations, folding self-review into the body whose decisions are most consequential.
AI governance errors are frequently invisible until someone with technical literacy looks for them. The Draft Regulations have left open, rather than answered, who monitors the gap in the oversight structure when there’s an error.
There is nothing in the Draft Regulations that requires an external audit of the Apex Body’s own judgment calls. There is no independent technical Ombudsperson outside the proposed institutional hierarchy. In some sense, courts have been self-reviewing and with jurisdiction under Articles 32 and 226 of the Constitution, remain available in principle, and so the absence of an external audit is not necessarily fatal.
Literacy and training are two different things
The Draft Regulations require regular and structured training for judges, advocates, and court staff, on the technical, legal and ethical dimensions of AI, at a minimum, addressing the functioning, capabilities and limitations of AI systems: the identification and mitigation of AI bias, hallucinations and technical errors; the legal and ethical framework governing AI in judicial context, including the rights of litigants and the obligations of judicial officers under these regulations; data protection principles, cybersecurity awareness and the handling of sensitive judicial data; and the correct procedure for reporting AI incidents, raising concerns and utilizing grievance literacy mechanisms.
This seems to be a credible curriculum, but not without a stronger guarantee of competence. UNESCO’s 2024 Global Judges’ Initiative: survey on the use of AI systems by judicial operators found that only nine per cent of operators who had used AI tools had received any related training, even as forty-four per cent were already using such tools. This is a gap between exposure and literacy that a static curriculum cannot close.
The Draft Regulations have left open, rather than answered, who monitors the gap in the oversight structure when there’s an error.
Draft Regulations require the AI Secretariat to design training programmes and review them at least once in every two years. But what it does not specify is the duration, recurrence, assessment, or what follows when an officer demonstrably fails to internalise and apply the material, say, through a flawed verification practice, or a simple unfamiliarity with what a hallucinated citation looks like before it reaches a judgment.
Real AI literacy can only be built through scenario-based exercise and repetition, and not simply by an onboarding set of modules.
Following Regulation 50, maintaining a repository of best practices on AI incidents is a sound instinct and a practice towards institutional memory. But it is only as useful as the obligation to consult it, and the Draft Regulations do not make consultation mandatory for deployment decisions.
Failures: isolated instances and patterns
In one case, an instance of a single hallucinated citation is a correctable error. That hopefully can be caught by the verification duty, as under Regulation 8(3). In a difficult category of events, another case being: suppose a hallucination pattern runs quietly through a legal research tool which is deployed across the district courts of a state, that will result in a systematic failure masquerading as unconnected individual failures.
Primarily, the Draft Regulations’ incident architecture is built around the individual case basis. Regulation 39 requires AI incidents to be logged in a database at each AI secretariat, then should be communicated with across the AI Secretariats of the other High Courts and to the Apex Body, so that corrective measures may be adopted across jurisdictions.
There’s no doubt that this is useful, but the Draft Regulations do not attempt to mandate a static analytical function where someone can be responsible for the job to scan the database for clustering, to ask whether how fifteen unrelated complaints, say, a transcription tool powered by AI across five district courts in a quarter, are in fact one incident wearing fifteen case numbers. This process, in turn, will make the AI incident database an archive rather than an early warning system without that layer.
The accountability mechanisms under the Draft Regulations are almost entirely procedural. None of these steps independently tests whether an AI system performs safely with the court’s language diversity and data.
Process compliance and safety
This is perhaps the most consequential structural choice of the Draft Regulations that it has taken. Notably, such an approach is not unique to India. A similar default posture can be observed in nearly every court AI framework written so far.
The accountability mechanisms under the Draft Regulations are almost entirely procedural: approval before deployment, usage documentation, followed by an audit. None of these steps independently tests whether an AI system performs safely with the court’s language diversity and data.
A court which satisfies every requirement under the Draft Regulations, for instance, fills an impact assessment, logs the system, conducts an annual audit, trains the staff, and can still deploy an AI system that causes real harm. Therefore, the focus should not be on “did the institution follow its own rules”, which is a procedural compliance question; instead, it should be “is the AI system safe to deploy and safe to use?”
England and Wales’s AI Guidance for Judicial Office Holders gestures at this distinction by treating personal responsibility (not the process) as the operative safeguard. On similar lines, Regulation 8 of the Draft Regulations vests accountability in an officer rather than on the AI tool in use. However, under Regulation 8 itself, the proviso allows officers to dispense with verification for recorded reasons. Further, allowance for administrative tools certified as reliable on a class basis without prior verification. Now, both of these create exactly the kind of shortcut that circumvents scrutiny and compliance — where such scrutiny matters the most.
The day after something goes wrong
As Regulation 52 states, an aggrieved party has the right to apply to the court where the AI system was used, and the court shall “pass appropriate orders as it may deem fit”. Further, in respect to any harm caused by the use of AI in court processes, a person may seek other legal remedies as available under any law in force (Regulation 53).
Some quick operational questions. What happens to an aggrieved party in a district court with no legal representation? Within 24 hours of a serious incident, who is actually notified – is it only the AI secretariat? Or does the aggrieved party learn anything before the matter resurfaces later? Who will investigate these matters, and further, with what independence from the body that approved the AI system in the very first place?
The proportionality of the grievance redressal architecture does not appear to mitigate the risks the Draft Regulations themselves identify. It is harder to assess, as timelines for investigation, an interim relief, and a guaranteed channel through which an unrepresented person would learn of a logged AI incident are all left unspecified.
What the rest of the world has built around its rules
Compared against the prominent comparative practices, India’s Draft Regulations are unusually comprehensive on paper, and unusually quiet within the ecosystem that would make such paper rules operative.
The iterative, incident-driven culture can be found in England and Wales. The first published AI guidance was released in December 2023, replaced in April 2025, and refreshed again on October 31, 2025, an approach built around revision rather than a single frozen statement of principle. Cases like Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin) are a visible warning case. In Ayinde, the Divisional Court of the High Court of Justice was confronted with false authorities and inaccurate legal material linked to the use of generative AI. The response did not stop with the bench: the Bar Council in England and Wales updated its generative AI guidance on November 26, 2025. The Civil Justice Council set up a working group in January 2025 to consider whether rules are needed for the use of AI by legal representatives in preparing court documents. It treats personal responsibility as the load-bearing safeguard. It warns that material entered into a public AI tool should be treated as published to the world, a confidentiality risk that the Indian Draft Regulations address only obliquely.
In this context, the United States federal judiciary illustrates a different stage of institutional AI governance. There is no single binding nationwide AI rule. There is an evolving mix of Administrative Office interim guidance, Federal Judicial Center materials, and local standing or general orders issued by individual courts. In early 2025, the Administrative Office established an advisory AI Task Force and said its interim guidance was intended to permit experimentation while preserving the integrity and independence of the federal courts. Subsequently, the 2025 Strategic Plan called for an AI governance framework. Senate Judiciary Committee pressure intensified after Chairman Chuck Grassley disclosed that staff in two federal judges’ chambers had used generative AI to draft inaccurate orders that the judges approved. The US has long governed AI in courts through incremental guidance, local practice, and oversight. India’s Supreme Court, on the other hand, has chosen to intervene with a comprehensive governance architecture of its own.
One of the most operationally developed models can be found in Singapore. Its Registrar’s Circular No. 1 of 2024, which applies from October 1, 2024 across the Supreme Court, State Courts and Family Justice Courts. The Circular did not prohibit the use of generative AI. But at the same time, it places full responsibility on the court user and requires independent verification of the output. Pre-emptive disclosure is generally not required unless the court asks for it. Justice Aidan XU stated in a July 2025 speech that the Singapore judiciary is exploring and testing AI products, trying judgment-drafting use cases carefully, and subjecting them to safeguards, supervision, and even red-teaming — that public-facing compliance guide sits alongside an internal culture of experimentation — is exactly the kind of ecosystem the Indian Draft Regulations gesture toward through CoRE-AI, but yet to kick off alive and function.
One of the most operationally developed models can be found in Singapore. It did not prohibit the use of generative AI. But at the same time, it placed full responsibility on the court user and required independent verification of the output.
The European Commission for the Efficiency of Justice (CEPEJ) is the closest institutional comparator. It began with the 2018 European Ethical Charter, then the Operational Assessment Tool and AI Advisory Board, and more recently, Guidelines on the Use of Generative AI for Courts. CEPEJ has moved beyond prescribing principles to constructing mechanisms through which the principles envisioned in the ethical charter can be assessed, operationalised, and periodically revisited. This offers a useful blueprint for the direction that Indian Draft Regulations will require in the future.
Written proactively, the Indian Draft Regulations have the advantage of foresight, but are yet to be tested in practice and when errors necessitate institutional measures.
What the institution still has to prove
The arguments raised here are not intended to take a position against the Draft Regulations or the Supreme Court of India’s decision to develop an AI governance framework for the use of AI in Indian courts — that is an appreciative, timely, and necessary step.
The anchoring general principles, on which the Draft Regulations are based, are correct. The questions raised in this piece, however, remain to be proven.
The objective of creating any regulations is to have conditions under which an institution can function and succeed. Regulations, per se, do not supply the competence, the funding, or the culture of taking compliance and scrutiny seriously rather than performing it. That work falls on the table of the Apex Body at the Supreme Court, the AI secretariats under the High Courts, and every officer who will have to decide, case by case, whether to trust what a device screen displays.
The stakeholder and public comment process, which closed on June 20, 2026, was the first test of whether the institution understands the difference between writing a good rule and building the capacity to live by it. The real test of these Draft Regulations, however, will be whether the institutions expected to implement them acquire the technical capacity, professional culture, and willingness to enforce them.
To read our explainer on the Supreme Court’s Draft AI Regulations, click here.