A Guardrail Grounded in Evidence
The prohibition on algorithmic bail and recidivism scoring in India's new draft is not a precautionary instinct — it is a response to documented harm. ProPublica's analysis of the COMPAS algorithm, widely used in US state courts, found Black defendants were 45 percent more likely to be assigned high-risk scores than white defendants even after controlling for criminal history, age, and gender. Black defendants who did not reoffend were flagged as high-risk at nearly twice the rate of comparable white defendants.
India's draft takes the strongest available regulatory position: not a compliance pathway with audits and documentation requirements, but an outright ban. The EU AI Act (Regulation 2024/1689) classifies AI systems that assist judicial authorities in interpreting facts and applying law as categorically high-risk under Annex III — requiring pre-deployment conformity assessments, human oversight, and data governance frameworks before deployment. India's Supreme Court AI Committee goes further, ruling such tools off the table for adjudicatory purposes entirely.
Both positions are defensible, but the Indian context sharpens the stakes. India's courts carry more than 54 million pending cases — 17.2 million of which have dragged on for more than five years — and the average judge manages upwards of 2,200 active matters against a national judiciary budget of just 0.08 percent of public expenditure. The pressure to automate is real. The draft correctly concludes that the bias risks outweigh it, at least until better tools and stronger safeguards exist.
What the Draft Actually Prohibits and Permits
Released on June 3, 2026 for public comment closing June 20, the Regulations for Use of Artificial Intelligence in Courts, 2026 establish three structural pillars.
The first is an absolute prohibition on AI adjudication. No AI system may determine bail eligibility, score a defendant's recidivism or flight risk, assess witness or party credibility, or issue any judicial order bearing legal consequences. Human judges hold sole authority over every outcome.
The second pillar is mandatory disclosure. Lawyers and litigants must declare in a prescribed format any AI assistance in preparing pleadings, submissions, or evidence. This is the practical answer to a documented problem: AI-generated citations that produce plausible-looking but entirely non-existent case references have already embarrassed counsel in Indian, US, and UK courts. Placing accountability squarely on the filing lawyer is the proportionate response.
The third pillar is a layered governance architecture. A permanent Apex Body at the Supreme Court — comprising two Supreme Court judges, two Chief Justices of High Courts, additional High Court judges, representatives from the Ministry of Electronics and IT, cybersecurity experts, and technology law advocates — will approve and certify AI tools for all courts and tribunals. Below it sit High Court AI Committees, a Centre of Research and Excellence (CoRE-AI), and AI Secretariats at each court, all required to publish annual compliance reports.
Permitted uses are scoped tightly to non-adjudicatory functions: case management and scheduling, automated transcription with human verification, legal research and precedent retrieval, multilingual translation, accessibility tools for disabled parties, and litigant-facing chatbots under judicial oversight.
Three Structural Gaps That Must Close
A close reading of the draft reveals three weaknesses that could render the framework symbolic rather than operational.
No penalty provisions. The draft prohibits AI bail scoring but specifies no disciplinary consequence, fine, or conviction-review mechanism for a court that uses such a tool anyway. A prohibition without an enforcement sanction is a statement of aspiration. The committee should specify, at minimum, the professional disciplinary process triggered by a violation and establish a mechanism to review bail orders or convictions where AI was used in breach of the regulations.
Governance circularity. The Apex Body that will regulate AI use in courts is itself appointed by the judiciary it oversees. There is no external audit, independent technology ombudsman, or parliamentary review of its decisions. For a framework premised on accountability and transparency, this internal loop creates a structural conflict. An independent oversight mechanism — even a parliamentary standing committee with reporting rights — would materially improve credibility.
No litigant challenge pathway. Under the current draft, a party who believes AI improperly influenced their case must raise the challenge before the same court that allegedly violated the regulations. This is procedurally incoherent, particularly for bail decisions where defendants are detained, time-sensitive, and least resourced to mount technical challenges. The final regulations need a designated independent channel for such complaints.
The Over-Prohibition Risk
The draft's caution runs in one direction. By confining AI strictly to administrative tasks, it excludes supervised AI-assisted legal research — where outputs would be visible to all parties, clearly flagged as AI-generated, and subject to adversarial challenge. The EU's framework regulates rather than bans research-assistance AI for the judiciary. Given that 17 million cases have waited over five years for resolution, excluding AI entirely from analytical support imposes a real cost — one that falls on litigants, not on regulators. A more calibrated final regulation would permit supervised research assistance while maintaining the absolute prohibition on anything that touches outcome.
What Comes Next
The public comment window has closed. The Apex Body has not yet been constituted; CoRE-AI does not yet exist. The committee's next move is to process submissions, finalize the text, and publish the criteria by which AI tools will be certified. How that certification process is designed will determine whether the framework becomes a genuine quality-control gate or a rubber-stamp exercise.
India's draft is a serious piece of judicial AI governance, more specific and more protective of fundamental rights than most equivalent documents globally. The prohibitions are the right ones. The delivery mechanism needs to match the ambition.