There are over 5,180 vacancies in lower courts out of a sanctioned strength of 24,521, and 326 vacant posts in the High Courts out of a sanctioned strength of 1,108.

Indian courts are therefore forced to operate at a lower strength. As per the India Justice Report, 2019, India spends only 0.08 per cent of its GDP on the judiciary. Except for the Supreme Court, which is funded by the Central Government, there are no uniform guidelines on State expenditure, resulting in less than one per cent expenditure by the majority of States, except Delhi at 1.9 per cent.

National investment through a Centrally Sponsored Scheme, including an allocation of ₹810 crore for infrastructure development and ₹1,200 crore for digitisation of courts, is not enough to reduce pendency. The abuse of legal procedure, including delays caused by adjournments, inordinate delays and frivolous litigation by the Government in over 73 per cent of admitted matters before the Supreme Court, has led to an informational overload.

Such sustained pressure makes the use of AI systems in the judicial process increasingly necessary. However, efficiency cannot come at the cost of safeguards.

The Verification Problem

The Draft Regulations on AI in Courts lay down guidelines in an attempt to integrate and regulate the use of AI in courts. However, the verification architecture under Regulations 19(c), 20(e), 43(3) and 44 is structurally self-defeating.

The problem is that the Regulations presuppose a human verification capacity that does not exist at the scale that AI deployment itself could create. This could force verifiers to rely on the very AI tools they are expected to check, producing what may be described as a verification regress.

Regulation 19(c) lays down permissible uses of an AI system, providing for the translation of judgments, orders, pleadings and other legal documents, subject to human verification of their accuracy and fidelity to the original.

With hundreds, if not thousands, of cases disposed of each day, the sheer scale of information would make this difficult for AI systems to process, let alone for human verification committees to examine. Even if a large human verification team were established in each court, procedural compliance could take decades.

It would also require large-scale infrastructure development in each State to store court data, which could be particularly difficult in remote areas because of geographic restrictions. The Regulations do not adequately account for the difference between the working capacity of humans and machines. A task of this scale cannot simply be delegated to humans and expected to be completed within court working hours.

SUVAS and the Scale of Translation

The experience of SUVAS (Supreme Court Vidhik Anuvaad Software) illustrates the scale of the challenge. SUVAS, a domain-specific Neural Machine Translation system built on the open-sourced Anuvaad engine, was launched in 2019 to translate Supreme Court judgments into vernacular languages.

As per the Ministry of Law and Justice’s 2025 report, over 83,783 judgments had been translated, compared with just 31 between March 2020 and October 2021. The question, therefore, is how a human-verification mechanism is expected to keep pace with a system that can move from dormant operation to translating tens of thousands of judgments.

The translations have also been disproportionately higher for large-population languages such as Hindi, with 36,344 judgments, and Punjabi, with 25,004. In contrast, smaller or less-digitised language communities such as Kashmiri, Khasi and Garo accounted for only 12 judgments combined.

This uneven absorption of languages points to the lack of translation infrastructure required for implementing AI systems across India.

The Black-Box and Explainability Dilemma

Regulation 2 governs the use, deployment and integration of AI systems in judicial, administrative or adjudicatory functions in all courts, tribunals and statutory commissions performing adjudicatory functions within India.

Regulation 4(2) entrusts courts to use AI only in an assistive capacity and attempts to provide procedural safeguards under Regulation 20, which lays down an absolute and non-derogable list prohibiting certain uses of AI in court processes.

Regulation 20(e) states that no undisclosed, opaque or unexplainable AI system shall be used in any court process that may materially affect the lawful rights or personal liberty of any party.

The definition of a ‘black box’ under Section 3(n) applies to an AI system employing deep-learning techniques whose internal processes and decision-making logic are not transparent and cannot be explained to a person materially affected by its output or decision.

The definition, however, restricts opacity in court processes involving high-risk or adjudicatory uses rather than imposing a blanket ban. Under Regulation 3(w), an AI system can fulfil the definition of ‘explainability’ by providing a comprehensible account of the reasoning used to arrive at an output upon request, without requiring specialist technical knowledge.

This creates a potential problem. AI systems can generate a plausible-sounding justification for an output without that explanation necessarily representing the actual computational process that produced it. An AI system could therefore remain a black box while simultaneously satisfying a formal requirement of explainability.

The absence of a sovereign Indian model, together with the lack of an AI liability framework, further weakens these protections and risks leaving the protection of material rights and personal liberty largely on paper.

Hallucinations and the Limits of Self-Certification

Section 3(z) defines ‘hallucinations’ as the phenomenon of AI systems generating outputs that are factually incorrect, fabricated and misleading. This includes fabricated case precedents.

The Draft Regulations establish a transparency and disclosure mechanism through Section 43(3). Where an AI tool is used by a party or legal representative in preparing or submitting a document, pleading or evidence, its AI-assisted character must be disclosed to the court through a prescribed declaration or certificate. Court-initiated AI use must similarly be declared.

The Supreme Court’s recent decision in Pooja Ramesh Singh vs Jammu & Kashmir Bank Ltd & Anr, 2026 LiveLaw (SC) 653 demonstrates the difficulty. The Court set aside an NCLT order after finding that the judgments and citations relied upon by the Tribunal were fake and AI-generated. It called for a zero-tolerance policy towards citing, referring to or relying upon AI-generated precedents without verification.

The case also exposed a significant weakness in the proposed disclosure mechanism. According to the Bank’s affidavit, the fabricated judgments were not cited before the NCLT by its counsel. Instead, the Tribunal had obtained them through its own research. The failure therefore came to light only through an appeal filed by the opposite counsel and not through voluntary disclosure under Regulation 43(3).

The incident shows that procedural paperwork can be evaded and that self-certification does not establish an independent verification mechanism. Instead, it could increase the burden on courts to verify the authenticity of every judgment cited by advocates.

The Missing Link in Human Verification

Regulation 19(c) establishes a Human-in-the-Loop (HITL) framework under which each officer is required to independently verify AI output for accuracy and fidelity to the original.

Regulation 44, meanwhile, establishes an institutional AI Content Verification Authority, mandated with the oversight, operation and continuous updating of verification standards, tools and protocols applicable to GenAI-generated content in court processes.

The difficulty lies in the absence of a clear link between these two provisions.

Regulation 44 establishes centralised standards for verification, but the Draft Regulations do not clearly state that the independent officers performing the verification duty under Regulation 19(c) must comply with those standards.

The result is a decentralised duty-bearer operating without an explicit statutory connection to the centralised standard-setter. An officer could therefore perform the verification duty without being expressly required to follow the protocols developed by the Authority under Regulation 44.

This is particularly striking because the Regulations provide detailed provisions concerning the composition, powers and functions of the AI Committees under Regulation 33, AI Secretariat under Regulation 34, and Apex Body under Regulations 22, 23 and 24. By contrast, the Content Verification Authority under Regulation 44 is dealt with in only a single sentence.

The Regulations therefore delegate the continuous updating of verification standards to an Authority without laying down a detailed procedure for handling the thousands of judgments and other materials produced by courts annually.

The institution tasked with identifying fabricated legal material itself risks suffering from insufficient institutional provisions.

Innovation Versus Restraint

There is, however, a broader policy rationale behind the structure of the Draft Regulations.

Regulation 16 creates a ‘Presumption in favour of responsible AI adoption’ by permitting AI for court assistance rather than dispute-outcome prediction. Regulation 17 promotes ‘Innovation over Restraint’ by permitting integration of AI systems and tools to improve judicial efficiency.

This is reinforced by Regulation 4, which makes AI use in court processes strictly subservient to human judgment.

The Draft Regulations appear to recognise the need to balance innovation with restraint. A principle-based regulatory framework can allow regulators to gather deployment data before finalising detailed rules, whereas excessive procedural prescription could risk stifling innovation.

Regulation 1(2), which allows the Chief Justice of India and Chief Justices of respective High Courts to notify the Regulations in official gazettes on different dates, also creates a decentralised commencement mechanism.

This phased approach could allow detailed procedures for bodies such as the Content Verification Authority to be added through subsequent notifications rather than requiring the first public draft to contain every institutional detail.

Justice PS Narasimha, chairman of the AI Committee that oversaw these Regulations and a member of the Bench in Pooja Ramesh Singh, encountered the very verification failure examined in this article. This suggests that the gap identified here is not merely an interpretive dispute but reflects a practical weakness that has already surfaced in judicial proceedings.

Building the Missing Connective Tissue

The safeguards under Regulations 19(c), 20(e), 43(3) and 44 currently exist on paper without sufficient institutional connective tissue to make them function effectively.

One solution would be to introduce an independent audit trigger to supplement the self-certification mechanism under Section 43(3). This would reduce dependence on litigants or opposing counsel to disclose AI use, as happened in Pooja Ramesh Singh.

A mandatory cross-reference between Regulation 19(c) and Regulation 44 would also ensure that verification undertaken by independent officers follows a standard protocol established by the institutional authority. Crucially, the verification process itself should not simply be delegated back to AI.

The Regulations should also amend Section 3(w) so that the requirement of explainability corresponds faithfully to the actual decision-making process of the AI model that produced the output, rather than merely requiring a comprehensible explanation to be made available upon request.

Until this connective tissue exists, the Draft Regulations risk regulating AI in name while leaving courts to verify machines with machines.

That is the verification regress that the regulatory framework must address before AI becomes deeply embedded in India's judicial process.

(The writer is a law student based in New Delhi, India)