Written by Chirag Garg, an LL.B. (Hons.) graduate from Amity Law School, Amity University, Noida.
As part of its digital mission to monitor and conduct research in a more efficient manner, the Supreme Court of India introduced SUPACE (Artificial Intelligence Research Assistant) in the year 2021. It makes no decisions. It comes in the form of presenting facts before a human trier of fact. Most conversations about judicial AI end there: earlier at the thought of an AI that aids, later at the thought of an AI that determines. It’s also where the actual legal issues start. Where a court begins to use a machine to provide it with information regarding the pertinent facts or precedents, an incorrect result from that machine is an incorrect result from the court.
Although it is not a hypothetical case yet, the situation is not just what could happen. Imagine if the research tool returned a superseded precedent, or the translation gave an inaccurate witness translation, or the risk assessment tool applied at the bail stage measured a factor which none of the statutes give it the power to measure. The order is signed by the judge. The litigant pays the sentence (or remains subject to it), or his case is dropped. What accountabilities are there for the judge to fulfil (at the same time as he/she relied on the output)? Who built the tool, or from whom did the court procure it? The law offers no clear guidance in this regard, so it is not too late to consider what it ought to look like before judicial use of AI becomes the norm.
Such a development has already been provided for, if not in the form stated, with the specific example of the Munich Olympics in mind. As of early 2026, benches of the Supreme Court, comprising the Chief Justice of the Court, Surya Kant, and Justices P S Narasimha and Alok Aradhe, marked petitions containing AI-generated citations to cases that are non-existent, stating that the order by justices that ‘misses’ such ‘case law’ would be a ‘work of misconduct’ rather than a simple ‘error’. In both cases, the drafting was done by the lawyer with the help of AI, not the court. There was no “winner” and “loser” to be determined with a phoney citation drawn up by either side of a bench, not a lawyer’s hand, but instead on a research tool the court has approved instead. That trickier question that’s made up in its own court is at large as well.
1. The judge: immunity meets over-reliance
The principle of judicial immunity in India means that a judge cannot be held liable for an action taken while performing his/her judicial duty, because the judge deciding a case in good faith should not be afraid of a lawsuit if it is done wrongly. That protection is for human interaction and human decisions. It did not necessarily need to be constructed for a judge who accepts a black box recommendation without investigating it for himself. Some judges mistakenly apply a wrong precedent, while others incorrectly believe an AI summary created from a precedent is the accurate precedent. The first is a normal mistake of the court, the type immunity serves to ward off. The second is more reminiscent of an abdication of the judicial function itself, and it is not clear into which old immunity doctrine, covering the notion of discretion in good faith, it is supposed to fit.
2. The developer: a product liability regime that was not built for this
India was also subjecting a new law that imposes product liability on manufacturers in the event of a faulty product/service or inadequate warning under the Consumer Protection Act, 2019. In theory, software that is being sold as a ‘court research’ or ‘assistance’ tool could be within this scope. The Act, however, is drafted with a realistic concept of ‘defect’: it would have been hard to draft a system where the ‘defect’ appears only after some interpretations of the system were used thousands of times. While establishing that a particular mistake made in a particular instance was due to a specific deficiency of the car is one evidentiary problem, establishing that the same mistake was made by the particular judge (or judges) was another. The Indian product liability law, which hasn’t yet been litigated against such claims, would stand the test during such contact, I expect, less well.
3. The state: procurement without accountability
If, like SUPACE, a court adopts an AI tool in the context of a government procurement process, then the state itself can become a potential defendant. Historically, the state has been immune from liability with respect to anything that is done in a judicial setting, but procurement and deployment of a technology tool is an administrative decision, which does not reflect a judicial decision and, therefore, is usually subject to challenge. But the real question is that if a litigant is wronged by mistake in an AI-assisted judicial process, will he even be aware of the use of AI in the judicial system? The use of AI in judicial affairs and processes is not currently mandatory in India. If the litigant never becomes aware that there’s something to challenge, the right to contest a decision is worth very little.
4. What the statute books do not cover
The Digital Personal Data Protection Act, 2023, is applicable to the processing of personal data, not if a decision based on that processing is wrong. The framework of intermediary liability provided under the Information Technology Act, 2000, provides for intermediaries who host third-party content but are not there to actively influence a judicial decision on a particular matter. None of the statutes were created to feature courtroom artificial intelligence, nor does either cover the void.
There was an improvement on paper on 3 June 2026 with the release of the AI Committee of the Supreme Court’s draft Regulations for the Use of Artificial Intelligence in Courts, 2026, for public comments, presiding over this Committee being Justice P. S. Narasimha. The draft would prohibit AI from making decisions in court matters ranging from research, drafting, scheduling, transcription, translation, and case management but prevent its use for deciding cases, determining bail eligibility, evaluating witness credibility, and even affecting deliberation. It will also mandate that lawyers disclose information regarding the use of AI when submitting pleadings or evidence and suggest the creation of a permanent top-tier AI committee that will approve AI systems used in court proceedings. That disclosure requirement is a genuine step towards the second principle advocated below, and a disclosure were made independent of this writing. However, the draft regulations apply to use, not consequence: they outline what AI can and cannot do and what must be disclosed, but do not address who is held responsible if it does something wrong and it produces an error. The liability question this piece is concerned with sits just outside their scope, and remains unresolved even if the regulations are adopted as drafted.
5. Toward a workable framework
Responsibility should vary in a judicial AI system based on how much decision-making is given to the tool; it’s probably not a good idea to single out one particular individual to be held responsible. As a starting point, three principles appear reasonable.
First, tiered accountability. When it comes to scrutiny, an AI tool which provides known information should be treated differently from one which rates the risk or ranks outlooks. If a tool is used instead of judgement, its effect should be viewed as an input into a judge’s decision that the tool must go through a separate liability analysis and not as a neutral input that a judge simply relied on.
Second, mandatory disclosure. It should be permissible to inform the litigant that an AI tool has been used in their case and at a reasonable degree of specificity why it has been used. Otherwise, the right to beings is an imperfect process and academic. This is not an obligation to show legitimate source code and the extent of use, but just the fact and usage scope.
Thirdly, human-in-the-loop legally, not as a standard. Indian courts already call for reasoning in their decisions. The requirement should also be extended to where a judge uses an AI-generated finding, that he or she should document their reliance and, on their own, verify the same. This makes a good intention to be diligent a step that can be documented and can be reviewed.
6. Why this is a due process question, not just a technology one
Article 21 of the Constitution and the line of cases starting with Maneka Gandhi v Union of India (1978) 1 SCC 248 confer ‘personal liberty’, not just a procedure that exists in paper or on the books. Justice K S Puttaswamy v Union of India (2017) 10 SCC 1, AIR 2017 SC 4161, extended that reasoning into the digital sphere by recognising privacy as intrinsic to liberty under Article 21. Any efficient process that fails to produce an error that can be traced, corrected, and moralised upon by anyone is not a fair, just and reasonable process in any sense of the word. What you have to compare to is not a search engine that doesn’t find a result. It is the UK Post Office Horizon scandal: where a defective accounting machine’s output, which was deemed credible as it was produced by a machine, was used to secure incorrect convictions, taking many years to be unravelled. Even if a jurisdiction is aware of the risk, Wisconsin’s own case in State v Loomis, 2016 WI 68, 881 NW 2d 749 (Wis 2016), demonstrates this, as the court upheld the use of a proprietary risk-assessment tool at sentencing but issued a caution to judges to not blindly follow their recommendations. The lessons of both episodes were not about the unique badness of the technology, but just that. Well, of course, no one had had a device that could ask the machine any questions before the harm was now irreversible.
However, one of the advantages of building the judicial AI liability rules before a similar debacle occurs is that India is not facing one. Today, SUPACE and similar technologies remain enablement tools, and new Regulations of 2026 (if enacted) would simply formalise a boundary and engrain the disclosure requirement. The formalisation of the use of AI does not automatically resolve who is responsible when there is a breach of the permitted use. That narrower issue, the centre’s own field of playing, the area in which AI governance meets the realm of cyber law, still has genuine space for a push for an answer, whilst the wider playing field is still in draft: a liable actor mirrors discretion, and no AI tool goes to court without one at the back of it.


Leave a Reply