India's Judiciary Drafts Its AI Rulebook

The Stakes

For the first time, India's highest court is writing the rules for the machines that may one day help decide cases. On 3 June 2026, the Supreme Court of India's AI Committee published the draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026 and opened it for comment, "[s]eeking views/suggestions of all stakeholders and the general public" ¹. This is not a white paper or an aspirational policy note; it is the scaffolding of a binding governance regime for AI inside the justice system.

The window is short. Submissions close on 20 June 2026 ¹, leaving stakeholders barely a fortnight to engage with a document that will shape how algorithms touch litigants, evidence and liberty.

This piece takes a deliberately balanced view. The draft has real strengths worth defending, alongside fixable gaps worth flagging, particularly around prohibited uses, risk classification, automated decision-making, human-in-the-loop safeguards and personal liberty. Atherna is actively reviewing the draft and will file formal views before the deadline; what follows previews that analysis.

What the Draft Gets Right

Before turning to its shortcomings, the draft deserves credit for several genuinely well-conceived design choices.

Human primacy as the load-bearing principle. The instrument refuses to let AI drift from tool to decision-maker. Regulation 4 fixes AI in a strictly assistive role:

This single clause does more to protect judicial independence than any number of technical safeguards.

Accountability that cannot be outsourced. Regulation 8 closes the obvious escape hatch: blaming the machine.

Responsibility stays with the human, where it belongs.

A serious institutional architecture. Rather than a one-page advisory, the draft builds standing machinery: an Apex Body with specialised Committees, an AI Secretariat, a Centre of Research and Excellence on AI (CoRE-AI), a public AI Register, periodic Audits, an AI Incident Database, and an Annual Transparency Report. ⁴ Governance is treated as a continuing institutional function, not a launch-day formality.

Transparency baked into procedure. Regulation 43 requires parties to be informed where an AI tool materially assists proceedings, mandates litigant and court declarations through Annexures I and II, and empowers courts to compel disclosure of the system used. ⁵

Anchoring to existing law. A dedicated chapter ties judicial AI to India's data-protection and cyber-security regime, rather than legislating in a vacuum. ⁴

The Core: Prohibited Uses, ADM, HITL & Liberty

If the architecture surveyed above is the draft's skeleton, Regulation 20 is its spine. It enumerates a set of uses that are not merely discouraged but forbidden, and it does so in language that forecloses the usual escape routes.

Non-derogable, by design

The prohibitions in Reg. 20 are absolute. The draft states plainly that these lines hold against everyone:

This is a deliberate structural choice: where most of the framework operates through graduated oversight and discretion, Reg. 20 removes discretion entirely. No Apex Body resolution, no Committee carve-out, no local practice direction can dilute these lines. That non-derogability is what converts aspirational principle into enforceable constraint.

Humans as the sole determinative authority

Two interlocking prohibitions protect adjudication itself. First, the draft bars:

Second, it forbids "adjudication or sentencing by AI systems without mandatory HITL," and confirms that "any adjudicatory AI output is strictly advisory." ⁶ The Human-in-the-Loop requirement is therefore not a courtesy layer bolted onto automation; it is the condition precedent to any adjudicatory output having effect at all.

The risk-scoring ban: a deliberate divergence

The draft's most consequential line is its outright prohibition on:

This is a conscious rejection of the US trajectory, the COMPAS-style actuarial tools that have drawn sustained criticism for opacity and disparate impact. India's judiciary is choosing, ex ante, not to import that controversy.

That choice is reinforced by the draft's treatment of opaque systems. Deployment of AI that is "opaque or incapable of explanation shall be subject to heightened scrutiny and shall be restricted in high-risk applications affecting personal liberty or any lawful right of a person." ⁷ Read together, Reg. 20 and Reg. 7(3) build a liberty-protective core: the closer a system gets to a person's freedom, the less room it has to operate unexplained or unsupervised.

Where the Draft Leaves Gaps

For all its structural ambition, the draft's safeguard regime carries soft spots that a closing-date submission should target.

Discretion that can hollow out human-in-the-loop

The verification architecture is weakened by its own provisos. Reg. 8(3) permits a responsible officer to "dispense with prior verification for reasons recorded in writing" ³. A written-reasons formality is a thin guard: under docket pressure it can normalise after-the-fact rationalisation and quietly erode the prior-verification discipline the draft elsewhere insists upon.

Undefined risk tiers

The text repeatedly turns on "high-risk applications" (Reg. 7(3) restricts opaque systems "in high-risk applications affecting personal liberty or any lawful right of a person" ⁷), yet nowhere defines the term. With heightened scrutiny and downstream obligations hinging on an undefined threshold, classification is left to ad hoc judgment.

Line-drawing and independence gaps

Three further tensions warrant attention:

• Profiling versus analytics. The risk-scoring prohibition lacks a bright line separating banned profiling from the permitted performance and backlog analytics, inviting characterisation disputes.
• In-house-only audits. Reg. 38(2) requires audits to be conducted "in-house," and bars sharing data, source code, datasets or architecture "with third parties or private entities for audits outside court premises" ⁸. Sound for data security, this forecloses arms-length scrutiny and external red-teaming.
• Reactive, court-bound redress. Grievance handling sits with the deploying court and internal AI Committee inquiry, with no independent appellate or ombuds channel for AI-caused harm.

Benchmarking: the EU AI Act & Global Norms

India's draft does not legislate in a vacuum. Reading it against the leading external benchmarks shows both where it leads the field and where it could profitably borrow.

The EU model: high-risk but permitted

The EU AI Act takes a fundamentally different posture. It classifies AI intended to assist a judicial authority "in researching and interpreting facts and the law and in applying the law to a concrete set of facts" as high-risk under Annex III, subject to a narrow exemption for purely procedural tasks ⁹. Crucially, the EU permits such systems while channelling them through conformity assessment, technical documentation, and registration obligations. It regulates risk; it does not forbid the function.

India goes further on the substance. As established above, the draft outright prohibits AI risk scoring and bars adjudicatory outcomes by automated decision-making alone: categorical bans the EU framework does not impose. On liberty protection, India is the more conservative instrument.

What the EU has that India lacks

The trade-off is structural discipline. The EU pairs its permissive stance with graduated risk tiers and mandatory third-party conformity assessment. India's draft, by contrast, relies on Ethical Impact Assessments without a formal risk-classification taxonomy or external certification, the very gap flagged above.

Convergence on principles

On normative fundamentals (human oversight, transparency, non-discrimination), the draft aligns closely with the CEPEJ European Ethical Charter and the UNESCO/Bangalore principles. The draft itself anticipates this: the Apex Body is mandated to "benchmark and integrate global standard AI systems into Indian Courts." ¹⁰ The mandate exists; the structured discipline to fulfil it is what remains to be built.

Concrete Recommendations

The gaps identified above admit of targeted, drafting-level fixes. We offer six, framed as constructive consultation input.

• Narrow the verification waiver. Reg. 8(3)'s Proviso 1 lets responsible officers "dispense with prior verification for reasons recorded in writing" ³. That latitude should be barred outright for any output touching personal liberty, and the recorded reasons should be contemporaneous and reviewable by the AI Committee rather than self-certifying.
• Codify a risk-tier taxonomy. Insert an explicit low/medium/high-risk schedule defining "high-risk" so that Reg. 7, 12 and 35 attach at predictable thresholds, adopting the EU's graduated logic without abandoning India's prohibition-led model.
• Draw the profiling/analytics bright line. Distinguish banned profiling and risk-scoring from permitted administrative analytics (Reg. 19(j)) through illustrative examples set out in the rules or an annexed schedule, so compliant case-management work is not chilled.
• Add independent assurance. Supplement in-house audits with periodic independent or peer audits under confidentiality safeguards, and mandate red-teaming before any high-risk deployment.
• Create an independent redress channel. Reg. 52 currently confines a harmed party to "an application with the court where the system is deployed." ¹¹ An independent grievance and appellate route, outside the deploying court, should sit alongside it.
• Default to explainability and sunset review. Require explainability-by-default for every high-risk tool, plus a sunset/periodic-review clause for each entry in the AI Register, so authorisations expire absent reassessment.

A Window That Closes on 20 June

The draft is a genuinely thoughtful, liberty-first foundation. But a foundation is not yet a finished building: consultation is precisely where good drafts become great rules, and the gaps identified above are fixable within the window the Committee has opened.

Atherna will file its formal views with the Supreme Court AI Committee ahead of the 20.06.2026 deadline, pressing the recommendations set out in this piece.

We urge fellow practitioners, litigants and technologists to do the same: read the draft, test it against your own practice, and submit. The window is short, and a rulebook this consequential deserves a crowded comment record.