Non-existent citations, real consequences
The NCLT's imaginary precedents reached the Supreme Court. Here's what that tells us about AI in adjudication.
By Natasha Aggarwal, Amrutha Desikan and Bhavin Patel
Welcome to The Bridge! Regulatory orders shape markets, govern disputes, and test the rule of law, yet their quality rarely gets the scrutiny it deserves. The Bridge is TrustBridge’s window into the quasi-judicial work of India’s regulators: how they decide, and how they could decide better.
Introduction
On 2 July 2026, the Supreme Court remanded an order to the National Company Law Tribunal (“NCLT”)1 because it relied on hallucinated citations to arrive at its conclusion. The impugned decision admitted a petition filed by Jammu and Kashmir Bank Limited (“J&K Bank”) to initiate corporate insolvency resolution proceedings (“CIRP”) in respect of Essel Infraprojects Limited (“Essel”) under the Insolvency and Bankruptcy Code, 2016. Some citations in the order referred to judgments that do not exist, while others correctly identified reported decisions but attributed non-existent paragraphs to them: there were three non-existent citations, two correct citations with non-existent paragraphs, and one misattributed citation with a non-existent paragraph.
One of Essel’s directors filed an appeal before the National Company Law Appellate Tribunal (“NCLAT”). The NCLAT stated “we do not find error in the impugned order”, and dismissed the appeal.2 Notably, neither side’s lawyers nor the NCLAT detected the imaginary citations in the impugned decision.
The fact that the NCLT relied on hallucinated citations was finally raised in second appeal before the Supreme Court. In response, the Supreme Court recorded zero-tolerance for “producing, citing or using AI-generated precedents without verification”, and stated that any decisions that rely on such material are to be set aside.
In this post, we explore why hallucinated citations went undetected across three levels of adjudication, and argue that the solutions lie in awareness, training, and process and technological safeguards. These include: training adjudicators on the limitations of LLMs, developing clear institutional guidance on appropriate use, and deploying technical safeguards.
Why did this happen?
It’s tempting to blame technology for this mess, but hallucinations are a risk inherent to large language models (“LLMs”) - and users of this technology need to be aware of this ex ante.3 LLMs are not built to retrieve authoritative legal materials from verified databases. They are designed to generate text by predicting the most statistically likely sequence of words. As a result, it is highly probable that they will produce citations that appear perfectly plausible but do not exist.
The real problem is that adjudicators and other actors in this ecosystem are increasingly relying on LLM-generated citations without independent verification.4 Large pendency volumes and crushing workloads create incentives to let through research that looks half-decent without strict scrutiny.
This has consequences: in the instant case, J&K Bank filed the CIRP application in 2023, the NCLT passed the impugned decision in 2024 and now, after the parties have made their way through three levels of adjudication, the matter must be considered afresh by the NCLT. While these delays are a worry, the greater concern is jurisprudence developed off imaginary precedent.
The root cause of this problem is a lack of awareness of how these systems work, and lack of training in verification processes. This is dangerous, because these systems and tools are remarkably useful and produce responses that are polished and perfectly plausible, and have the potential to reduce the time taken to adjudicate a matter, which is tempting for an NCLT adjudicator burdened with an overflowing case docket.
How do we resolve this?
We can think of three ways:
First, we suggest that adjudicators and other actors involved in the adjudicatory workflow be trained on the use of LLMs. Such training should cover LLMs’ inherent limitations, methods to mitigate the impact of such limitations, and appropriate use cases in the adjudicatory workflow.
Second, adjudicatory bodies should develop process manuals that govern how AI is to be used in this ecosystem. Such manuals must be foregrounded in design principles such as human oversight and transparency and explainability,5 and could include material such as verification checklists.
Third, adjudicatory workflows will benefit from a number of technical solutions. Examples of such solutions include the use of retrieval augmented generation, integration of legal knowledge graphs, implementation of processes for post-generation verifications, and mandatory reviewer verification.6 Such solutions will augment an adjudicator’s capacity to use LLMs responsibly. Some of these technical solutions and safeguards are not necessarily complex; one can imagine simple solutions such as mandating that all filings include a list of citations. Once citations are presented in a standardised format, simple deterministic programs, such as Python scripts, can automatically check whether the cited cases actually exist in a verified database, and whether quoted extracts correspond to the correct judgments. Any discrepancy can be flagged for human review.
Conclusion
Hallucinations are not a bug waiting to be fixed, they are an inherent feature of how LLMs work. This means that no future version of this technology will make the problem disappear, and that the burden of managing it falls on institutions. LLMs are genuinely useful tools for overburdened adjudicators. That utility creates the precise conditions that make them so tempting to use without safeguards. The appropriate response is not to retreat from AI, but to build workforces and institutions capable of using it responsibly.
- The authors are researchers at TrustBridge, and would like to thank Renuka Sane for her feedback.
CITATION
Natasha Aggarwal, Amrutha Desikan and Bhavin Patel, 2026. “Non-existent citations, real consequences”, The Bridge, TrustBridge Rule of Law Foundation
Thanks for reading The Bridge! Subscribe for free to receive new posts and support our work.
References
Aakriti Bansal, ‘10 Cases that Show Indian Courts Have an AI Hallucination Problem’ (MediaNama, 3 July 2026) <https://www.medianama.com/2026/07/223-10-cases-ai-hallucination-cases-in-indian-courts/>
Himanshu Mishra, ‘Phantom Precedents: The Rise of AI-Generated Case Law in Indian Courts’ (Live Law, 17 March 2026) <https://www.livelaw.in/articles/phantom-precedents-ai-generated-case-law-indian-courts-526665>.
Pooja Ramesh Singh v. Jammu and Kashmir Bank Limited, Civil Appeal No. 11950 of 2025.
Pooja Ramesh Singh v. Jammu and Kashmir Bank Limited, Order dated 11 September 2025 in Comp. App. (AT) (Ins) No. 1808 of 2024.
Natasha Aggarwal and others, ‘Can Technology Augment Order Writing Capacity at Regulators?’ (2026) 22(1) Indian JL & Tech 2.
Pooja Ramesh Singh v. Jammu and Kashmir Bank Limited, Civil Appeal No. 11950 of 2025.
Pooja Ramesh Singh v. Jammu and Kashmir Bank Limited, Order dated 11 September 2025 in Comp. App. (AT) (Ins) No. 1808 of 2024.
Natasha Aggarwal and others, ‘Can Technology Augment Order Writing Capacity at Regulators?’ (2026) 22(1) Indian JL & Tech 2, 16.
Aakriti Bansal, ‘10 Cases that Show Indian Courts Have an AI Hallucination Problem’ (MediaNama, 3 July 2026) <https://www.medianama.com/2026/07/223-10-cases-ai-hallucination-cases-in-indian-courts/>; Himanshu Mishra, ‘Phantom Precedents: The Rise of AI-Generated Case Law in Indian Courts’ (Live Law, 17 March 2026) <https://www.livelaw.in/articles/phantom-precedents-ai-generated-case-law-indian-courts-526665>.
Aggarwal and others (n 3) 26.
ibid Annexure C.