When the Supreme Court of India released the Draft Regulations for the Use of Artificial Intelligence in Courts, 2026 for public consultation, it offered a clear answer to a question that has unsettled courts everywhere AI has entered sentencing, bail, or case management. The question is deceptively simple: who is responsible when AI gets it wrong? The regulations say that the question barely needs asking. AI is advisory, It is the judge who decides. And critically, the draft states that accountability for AI-assisted decisions rests entirely with the judicial officer using the technology (Regulation 8) . Neither software failure nor algorithmic bias may be invoked as an excuse for an erroneous outcome.

It is a satisfying draft to read. It is doctrinally tidy, constitutionally defensible, and easy to communicate to a sceptical public. On close inspection, it also solves the accountability problem on paper. The regulations correctly identify who must answer for an AI-assisted judicial error. But they say almost nothing about whether that person is actually positioned to prevent the error in the first place, or about everyone else standing between the algorithm and the judge’s signature who currently answers for nothing at all.

The absence of a clearly articulated liability framework for AI-related harm tends to be treated as a peripheral concern, often receiving limited attention within broader discussions of the regulation’s shortcomings. This piece argues that it deserves more sustained attention. Without resolving who bears liability when an AI-assisted judicial decision causes harm, the regulation’s broader commitment to accountability remains structurally incomplete.

What the Regulations Actually Say

The architecture rests on a principle that the draft frames as “human primacy”, set out in Chapter II alongside judicial independence, fairness, transparency, accountability, and auditability. Regulation 3(1)(m) defines AI narrowly enough to exclude general-purpose software, capturing only systems that infer, learn, and generate decisions, predictions, or recommendations for court processes. Chapter III draws the now-familiar line between permitted uses such as case management, legal research, translation, transcription, document summarisation, judicial analytics and prohibited uses, chiefly, any system that independently adjudicates disputes, scores recidivism risk, assesses bail eligibility, or evaluates witness credibility.

Every permitted use carries the same condition that a human being must remain in the loop. AI output is advisory only and the judge bears full responsibility for what is ultimately decided.      Before deployment, tools must clear a Technical and Ethical Impact Assessment. This assessment examines training data quality, bias risk, cybersecurity vulnerabilities, and adherence to explainability standards. It may be preceded by Controlled Environment Testing in an isolated setting. 

The regulations also establish an elaborate institutional structure to oversee this process. This structure is headed by the Apex Body at the Supreme Court level.  AI Committees and Secretariats are established at the Supreme Court and High Court level, responsible for vetting tools, conducting audits, and investigating incidents. A proposed Centre of Research and Excellence on Artificial Intelligence is envisioned for long-term policy development in the field.

On paper, this is a fairly complete lifecycle: assess, test, deploy with restrictions, audit continuously, and hold the human user accountable at the point of decision. What it does not contain anywhere is a liability framework that distinguishes between the different ways an AI-assisted judicial error can actually occur.

The Logic of Concentrating Accountability in the Judge

The choice to fix accountability on the judge, rather than distribute it across vendors, committees, and code, rests on a defensible rationale. Constitutional doctrine on personal liberty has long insisted that decisions affecting it be made by an identifiable, reasoning human actor: the due process logic from Article 21 jurisprudence on fairness and natural justice in administrative and judicial action. If responsibility for a judicial outcome could be shared with, or shifted onto, a software vendor or an institutional committee, the chain of constitutional accountability would blur exactly where it needs to stay sharp. 

Keeping the judge as the singular, irreducible point of responsibility is also administratively simple. It avoids years of litigation over apportionment between developer negligence, institutional approval failure, and judicial misuse. Multi-party liability disputes of this kind tend to slow down accountability rather than deliver it. The 2018 Uber self-driving fatality shows why that simplicity can be misleading. In the Arizona collision that killed Elaine Herzberg, Uber’s automated driving system failed to correctly identify the pedestrian and predict her path, while the human safety driver was also inattentive. The ensuing accountability question did not map neatly onto a single actor: the software system, its developer, and the human operator each formed a part of the causal chain, raising what legal scholars have described as the “hardly solvable question of apportionment of liability” in accidents involving AI. Vasquez, the human safety driver, was charged with criminal negligence. Uber faced no criminal charges despite the software failure originating in its system. The trouble is that a bright-line rule like this only does the work it is designed for when the party who bears the liability also has a realistic ability to detect and prevent the harm. That assumption is where the draft regulations are the weakest.

Three Places the Framework Goes Silent

First, the regulations assume judges will catch what the AI gets wrong, but the psychology of human oversight rarely cooperates with that assumption. “Human in the loop” is a procedural guarantee, not a substantive one. A judge formally retains the power to override an AI-generated summary, translation, or research output, but a substantial body of research on automation bias shows that people, including trained professionals, tend to defer to algorithmic outputs even when explicitly authorised to disagree, particularly under time pressure. A survey of Canadian judges and lawyers found that many did not regard risk-assessment tools as particularly reliable yet preferred using them because doing so reduced the personal and professional risks associated with their own decisions, as documented in research on judicial attitudes to risk tools. That is precisely the structural dynamic the Indian framework will create.

 The justification for AI adoption in the first place is the backlog: as of March 2026, India’s courts carry over 55.8 million pending cases, with more than 85% concentrated in districts and subordinate courts. The same overload that makes AI assistance attractive is the overload most likely to erode the unhurried, sentence-by-sentence scrutiny that “advisory only” presumes. Providing that an overburdened judge remains fully liable for catching an AI’s hallucinated citation is not the same as giving them the time or institutional support to actually catch it.

Second, the bodies that approve AI tools bear no adverse consequences should their approval subsequently prove to be erroneous. The Apex Body and AI Committees, as the designated Appropriate Authorities, oversee the approval of AI Systems following a Technical and Ethical Impact Assessment and clear a tool for use under Regulations 3(1)(l), 33(3)(a) and 35(1). If that assessment misses a systemic flaw such as biased training data, an undetected hallucination tendency, a vulnerability that surfaces only at scale, the regulations are silent on whether the approving institution bears any consequence. However, this criticism needs qualification. Regulation 53 expressly preserves existing legal remedies, so the draft does not immunise approving bodies or other actors from tort or other legal claims. The problem is practical rather than absolute: preserving remedies in the abstract does not create a structured mechanism for identifying the responsible institutional actor, allocating fault, or making that remedy readily usable by an aggrieved litigant. A person harmed by an AI-assisted decision may have to initiate proceedings without clear guidance on whether the approving body, the vendor, or another actor is the proper defendant, and without access to the internal assessment or procurement records needed to establish causation and negligence. The residual burden can therefore still flow downward to the individual judge who relied on a tool the system had certified as safe.

Third, the regulations recognise vendor liability but leave it largely within a contractual framework. Chapter VI anticipates engagement with private entities, and Regulation 46(4)(f) and (l) requires agreements with AI developers to contain clauses attributing liability for harm or an AI incident. This is an important safeguard, but contractual liability provisions in government procurement agreements do not, by themselves, create a transparent or readily accessible remedy for a litigant harmed by an AI-assisted judicial process. The affected litigant ordinarily is situated outside the contractual relationship between the procuring authority and the vendor and may have no direct access to the terms governing liability, indemnification, or recourse. Nor do the Regulations establish a clear mechanism for determining when a vendor’s contractual liability is triggered, how it should be apportioned against institutional or judicial responsibility, or how an aggrieved person may pursue it. Therefore, vendor liability exists in the regulatory architecture, but remains contractually mediated, while judicial accountability is immediate, individual, and expressly placed on the officer “using” the technology.

This Gap Has a Name: The Loomis Problem

This is not a hypothetical gap; it is the same one that shaped the most cited cautionary tale in this field. In State v. Loomis, decided by the Wisconsin Supreme Court in 2016, the due process challenge ran against the sentencing court’s reliance on the COMPAS risk-assessment tool, not against Northpointe Inc., the company that built it. The court’s eventual answer was to require cautionary disclosure about COMPAS’s limitations in presentence reports, rather than resolve who answers when the tools itself is flawed, and critics at the time noted that such a caution is unlikely to produce real scrutiny if it says nothing about a judge’s actual capacity to evaluate a proprietary tool, or the institutional pressure to rely on it. 

COMPAS faced a separate, more public reckoning months later, when a May 2016 ProPublica investigation alleged the tool disproportionately flagged Black defendants as high-risk, an allegation Northpointe disputed and researchers still debate on technical grounds. But the controversy made one thing clear regardless of where one lands on the statistical dispute: when an opaque, vendor-built tool produces a questionable outcome, the accountability framework that exists determines who gets asked hard questions. In Loomis, that accountability fell on the sentencing judge alone. Northpointe, the tool’s developer, answered to no court. India’s draft regulations reflect an awareness of this lineage as a cautionary one. What they have not done is close the structural gap that made Loomis cautionary in the first place.

Why the Indian Context Makes This Urgent

The impact of these gaps will not be uniform across the judicial hierarchy. The regulations extend to the Supreme Court, High Courts, subordinate courts, tribunals, and statutory commissions exercising adjudicatory power. These are a vast range of institutions with very different capacities. A High Court bench with research staff and institutional support is far better positioned to interrogate an AI-generated case summary than a subordinate court judge managing an extreme caseload with limited administrative backing. DAKSH has documented that subordinate court judges spend substantial portions of court time on administrative and listing functions and has identified the lack of supporting administrative capacity as a structural constraint on judicial efficiency. DAKSH’s analysis supports treating institutional capacity as a genuine component of the accountability problem, rather than assuming that “human oversight” operates identically across court levels. The linguistic and geographical access barriers the regulations cite as part of the case for adopting AI only reinforces this disparity.

The same “human primacy” rule will mean something quite different depending on where in that hierarchy it is applied, yet the draft regulations describe a single accountability standard regardless of institutional capacity. A liability framework that does not account for this variation risks becoming harshest in practice precisely where judicial bandwidth to exercise genuine oversight is thinnest.

The Case Against Waiting

A natural objection is that this is premature, that liability allocation in any new technology tends to get settled through disputes, grievance redressal, and eventually case law, rather than written into a regulation before a complex case has arisen. The draft does establish grievance redressal mechanisms and incident-reporting obligations in Chapters V through IX, and these will generate practical experience over time. But waiting carries a specific cost here that it does not carry in most settings. The e-Courts ecosystem is already scaling AI tools such as SUPACE across hundreds of courts, and the draft regulations are built to govern deployment at that scale, not a single pilot.

 By the time a hard case forces clarity on vendor or institutional liability, the reliance on AI advisory output will already be entrenched across the subordinate judiciary. Such a case will most likely resemble Loomis, where a litigant challenges an order shaped by a flawed tool. That entrenchment will be deepest where institutional capacity is weak. Addressing the liability framework at the draft stage is therefore not premature caution; it is the point at which the cost of getting it wrong remains low.

Conclusion

None of this argues against human primacy as a starting principle. It is the right one, and an improvement over jurisdictions that have not drawn that line at all. The argument is that a principle is not yet a liability framework. A workable one would need to distinguish between three failure modes: negligent over-reliance by a judge who had reasonable means to catch an error, systemic tool failure despite proper use, and institutional failure in the assessment and approval process itself. It would need vendor obligations tied to the Technical and Ethical Impact Assessment that persists past deployment, not a one-time compliance. And the audit and incident-reporting mechanisms the regulations already established would need to evaluate the approving body’s judgement, not only the individual judge’s conduct.

The draft gets the hardest question right: AI does not get to decide, and the judge’s signature remains the constitutionally meaningful act. But accountability cannot be made meaningful by locating the entire burden at that signature. A rule that places an extra duty of care on judges without ensuring the time, training, administrative support, and institutional safeguards necessary to satisfy it does not fully resolve the liability problem. The better approach is not to dilute judicial accountability, but to distribute responsibility across the chain that makes AI-assisted decision-making possible. Judges could remain accountable for negligent use where genuine oversight was reasonably available. Approving institutions could be held responsible for failures in assessment and governance. Vendors may remain liable where contractual or other legal duties are breached. Regulation 53 preserves the possibility of reaching these actors; the next step is to make that possibility structured, transparent, and practically accessible. Otherwise, the draft risks moving the accountability gap one level down, onto the desk of the person least equipped to absorb it.

The author is a third-year B.A. LL.B. (Hons.) student at Gujarat National Law University, Silvassa.