As of June 14, 2026, India’s subordinate courts carried a crushing backlog of over 4.94 crore pending cases. High Courts had another 64.2 lakh matters waiting for disposal. The Supreme Court itself crossed 93,676 pending cases. As per NJDG data, over 3,37,600 cases across the system have been pending for more than 20 years. The numbers make the case for artificial intelligence almost by themselves. For a system hemorrhaging time, any credible intervention is welcome.
Earlier this month, the Supreme Court’s AI Committee released the draft Regulations for Use of Artificial Intelligence in Courts, 2026, inviting public comments until June 20. The document has 57 regulations across ten chapters.
The Supreme Court’s SUVAS platform already translates judgments into 18 Indian languages. Adalat AI, a startup-developed transcription tool, is deployed across more than 4,000 courtrooms across several states, generating real-time transcripts of proceedings and reducing adjournments caused by disputes over what was said in court.
A tool called LegRAA (Legal Research Analysis Assistant) has been developed to assist judges in legal research. And Digital Courts 2.1 now provides judges a single paperless interface for case management, combined with voice-to-text (ASR-SHRUTI) and translation (PANINI) capabilities.
Yet the government’s own parliamentary statement acknowledges that “the current scope of AI-based solutions remains limited to controlled pilot deployments”, making the draft regulations the first attempt to formalise what has been building informally for years.
The core structure of the draft, legal experts say, is broadly defensible. AI may assist, but it cannot adjudicate.
Permissible uses include transcription, translation, legal research, case scheduling, anonymisation of judgments, chatbot guidance for litigants, accessibility tools, and fraud detection in filings. The prohibited list includes that no AI system may deliver a judgment, determine bail, conduct risk-scoring on recidivism or flight risk, evaluate the credibility of a witness, or influence judicial deliberations in any manner.
“The draft does not proceed from a position of fear or resistance to AI,” says Ankit Sahni, partner at Ajay Sahni & Associates. “It recognises that AI can improve access, efficiency and administration, but draws clear boundaries where liberty, adjudication, sentencing, risk scoring and outcome prediction are concerned. That is the right balance for a justice system.”
“The draft highlights a bias towards innovation and action, balanced with hard guardrails around use cases which may have significant impact on individual rights, and is a reasonable framework for AI which may have such deep and meaningful impact on the administration of justice,” said Arun Prabhu, partner & co- head, digital +, TMT, Cyril Amarchand Mangaldas.
Visible Hurdles:
Uneven Capacity and Lack of Infrastructure
As of February 2026, India’s High Courts were functioning with 813 judges against a sanctioned strength of 1,122, leaving 309 vacancies or a 27.5 percent shortfall. District and subordinate courts carried 4,848 vacancies out of a sanctioned strength of 25,894.
On digital infrastructure, a total of 22,411 district and subordinate courts across India are covered under the eCourts project, with Uttar Pradesh (3,321 courts) and Tamil Nadu (1,317 courts) accounting for the largest shares. The e-Committee of the Supreme Court has conducted 966 training programmes and trained 3.22 lakh stakeholders including judges, lawyers, litigants, and court staff, under a six-tier national, state, and regional model.
Twenty-nine virtual courts are now functioning across 21 states and union territories, currently limited to traffic-related offences. The foundation is being built, but it is being built unevenly, and the draft AI policy assumes a level of digital readiness at the base of the pyramid that the data does not yet support.
Legal experts point to lack of infrastructure as one of the core weaknesses.
“High courts are more likely to adopt these systems safely first, because they generally have better administrative capacity, more experienced technical staff, and stronger institutional control over training and standard-setting,” said Hersh Desai, partner, Chitnis and Desai. “Infrastructure gaps, uneven staffing, limited digital literacy, and variable internet or hardware quality will make uniform adoption difficult, even if the rules are good on paper.”
Dr. Sudhir Raja Ravindran, Attorney at Altacit Global and Solicitor in England & Wales, said, “Safe adoption will require a phased model. Courts with weaker infrastructure should not be pushed into complex AI deployment before basic digitisation, cybersecurity, training, and grievance redressal systems are in place.”
“Uneven infrastructure, staff shortages and variable digital readiness means that poorly implemented AI may not reduce error but simply move it somewhere less visible. A phased model with minimum infrastructure thresholds and mandatory training before deployment is essential,” said Anshul Verma, partner SKV Law Offices.
‘Subservient to human judgment’ – In letter and spirit
The draft rules, under permissible use of AI permit use of AI for ‘legal research, precedent retrieval, citation verification and document summarisation’. One of the core principles of the draft rules include, ‘human primacy and judicial independence’ which says, “the use of Artificial Intelligence in Court processes shall at all times remain strictly subservient to human judgment and judicial authority.”
Legal experts say that problem is not merely one of inadequate human oversight but of automation bias, and the tendency of reviewers to anchor to machine-generated output, meaning the “human-in-the-loop” can inadvertently launder errors rather than catch them.
“In practice, the line should be enforced through process, not slogans,” said Desai. “If the AI output is material to a judicial decision, the file should show what human review was done and what source material was consulted. Without that, “human oversight” becomes a formality rather than a safeguard.”
“AI must not be used in a way that creates automation bias. Judges and officers should be trained to question AI outputs, not merely approve them. The real test of “human-in-the-loop” is whether the human has meaningful power, time, information, and institutional confidence to disagree with the machine,” said Ravindran.
“This will be enforced by ensuring that human judgement is interspersed at the appropriate gates. Human in the loop requirements, to be implemented before key decision points, as well as mandatory human review and grievance redressal, can help implement this in practice,” said Prabhu.
