A California appellate court fined a lawyer $10,000 for AI-fabricated citations. An Indian tax tribunal ruled on a Rs 669 crore dispute using four citations that did not exist. The pattern behind both cases explains why grounding in Indian law, not just AI capability, decides which legal tech actually works here.
"In December 2024, the Income Tax Appellate Tribunal’s Bengaluru bench recalled an order in Buckeye Trust v. PCIT, a dispute worth roughly Rs 669 crore. The department’s representative had relied on ChatGPT-generated case law. Three of the four citations were fake Supreme Court judgments. The fourth was a fabricated Madras High Court ruling. The bench had copied them into its own order without checking."
That case is not an outlier. It sits alongside at least nine other documented instances of Indian tribunals, trial courts, and High Courts encountering fabricated AI citations since 2023, a pattern that culminated in the Supreme Court’s own ruling in July 2026 declaring zero tolerance for unverified AI-generated material in judicial proceedings. Nearly every one of these incidents traces back to the same root cause: a general-purpose AI tool, built and trained largely on global and American legal content, being asked a question about Indian law it had no reliable way to answer.
This is the actual argument for homegrown legal technology in India, and it has very little to do with national sentiment or a preference for local vendors over international ones. It has to do with a specific, well-documented technical gap: global AI tools are trained on global legal data, and Indian law is a large, fast-changing, multilingual body of statutes and precedent that a model built primarily on American and English case law simply does not contain. This piece walks through the evidence for that gap, what it costs a firm that gets it wrong, and what to look for in a platform that is built for Indian law from the ground up rather than adapted to it after the fact.
This distinction has become sharper, not softer, over the past two years, for a specific reason. India has spent that period rewriting the foundations of its own criminal law, replacing the Indian Penal Code, the Code of Criminal Procedure, and the Indian Evidence Act with the Bharatiya Nyaya Sanhita, the Bharatiya Nagarik Suraksha Sanhita, and the Bharatiya Sakshya Adhiniyam. A tool trained before that transition, or trained on a dataset that was never updated to reflect it, is not simply missing a feature. It is confidently answering questions about a legal code that, for large parts of the country’s criminal justice system, no longer exists in the form the model learned it in.
The Pattern Behind India’s AI Hallucination Cases
The Buckeye Trust matter was not the first warning sign, and it was far from the last. Delhi High Court flagged the risk of “fictional case laws” surfacing through ChatGPT as early as 2023. Since then, the pattern has repeated across tax tribunals, company law tribunals, and High Courts, each time following the same shape: a fabricated citation, confidently generated, that nobody checked before it entered the official record.
Buckeye Trust v. PCIT, ITAT Bengaluru
A Rs 669 crore dispute decided using four fabricated citations, three fake Supreme Court judgments and one fake Madras High Court ruling, sourced from ChatGPT and never verified before the order was passed.
Bombay High Court, Rs 28 Crore Assessment
The Court quashed a tax assessment resting on three AI-invented precedents, treating the fabricated authorities as a serious lapse rather than a harmless drafting error.
Delhi High Court's Early Warning
The Court cautioned against “fictional case laws” generated by ChatGPT in a trademark matter, one of the first Indian judicial warnings about generative AI’s unreliability on Indian citations.
Supreme Court's Zero-Tolerance Ruling
Setting aside NCLT and NCLAT orders built on six defective, uncited precedents, the Supreme Court declared zero tolerance for unverified AI-generated material in any judicial proceeding.
A comprehensive international database now tracks more than 160 documented cases of AI hallucination in court filings across multiple jurisdictions, with the Indian cluster growing sharply over the past year. The mechanism behind nearly every one of them is identical: a general-purpose model answered a question about a specific legal system without having reliable access to that system’s actual source material.
What makes the Indian cluster distinctive is not the number of incidents on its own, but how quickly they moved from research curiosities to matters with real financial stakes. The Buckeye Trust dispute involved roughly Rs 669 crore. The Bombay High Court matter involved a Rs 28 crore tax assessment. These were not academic moot problems or low-stakes filings where a fabricated citation could be caught and corrected without consequence. They were live disputes with real money and real parties on both sides, decided, in part, on the strength of case law that a search of any actual Indian legal database would have shown never existed.
Why General-Purpose and Global Tools Struggle With Indian Law
None of this means AI is unreliable in general. Tools like ChatGPT and Claude are genuinely strong at language: summarising, restructuring, drafting, and translating. The problem is narrower and more specific, and it explains why the same tool that drafts fluent English prose invents an entire Supreme Court judgment when asked for a citation.
Built on Global, Largely Western Legal Text
General-purpose models are trained overwhelmingly on American and English case law and commentary, simply because that is where most digitised legal text online originates from.
Answers From Memory, Not From a Source
Without a live, curated Indian case law corpus to search, a general-purpose model predicts what a citation should plausibly look like rather than retrieving one that actually exists.
A Codebase That Changed Underneath Them
The shift from the Indian Penal Code to the Bharatiya Nyaya Sanhita, and equivalent changes to criminal procedure and evidence law, means models trained before the transition carry outdated section numbers by default.
Blind to a Large Share of the Record
A meaningful and growing share of Indian judgments and orders now exist in Hindi and regional languages. Models built for English-first markets have limited reliable access to that material.
The common thread is grounding, not raw capability. A model that can write elegant legal English is not the same as a model that can correctly tell you what Section 138 of the Negotiable Instruments Act currently says, whether a given precedent has been overruled by a later Bharatiya Nyaya Sanhita-era ruling, or what a specific Bombay High Court order from last month actually held. Those answers require a live, current, India-specific source to check against, and a general-purpose global tool was never built with one attached by default.
Grounding & Compliance Under DPDP Act
| Consideration | Typical Global Platform | Typical Homegrown Platform |
|---|---|---|
| Primary training corpus | Global, English-dominant legal text | Indian statutes, judgments, and procedure by default |
| Citation grounding | Often generated from memory | Retrieved from a live Indian legal database |
| Regulatory transition coverage | May lag behind BNS, BNSS, BSA updates | Built to track current Indian statutory codes |
| Vernacular language support | Limited, often a later add-on | Frequently core to the product from the start |
| Data residency under DPDP | May require separate compliance review | Often built around Indian hosting by default |
Why Retrieval-Grounded Tools Behave Differently
This is where India’s homegrown legal AI platforms diverge from the general-purpose tools that keep making headlines for the wrong reasons. Products like Manupatra’s AI assistant, SCC Online’s research tools, Indian Kanoon, and CaseMine’s AMICUS are built around retrieval-augmented generation: rather than generating an answer purely from what the model remembers, the system first searches a real, curated corpus of Indian judgments and statutes, then builds its answer from what it actually finds there.
"The presence or absence of a real, searchable source is what separates a citation you can trust from one you have to hope is real."
That single architectural choice, search first, answer second, is why retrieval-grounded Indian tools sharply reduce the fabrication problem, even though no system eliminates it entirely and human verification remains essential regardless of which tool produced the first draft. It is also, telling, the same lesson global players have started acting on rather than disputing.
A Practical Checklist: Grounded Indian Foundations
A citation you can actually trace back to a real source.
Retrieval-grounded search against a live Indian corpus is the specific technical choice that separates a verifiable citation from a plausible-sounding fabrication.
Currency with India’s fast-moving statutory transitions.
The shift from the IPC, CrPC, and Evidence Act to the BNS, BNSS, and BSA is exactly the kind of change a platform maintained specifically for Indian law tracks as a priority, not an afterthought.
Native handling of vernacular material.
As courts issue more judgments and orders in regional languages, a platform built around that reality from the start avoids the translation-layer quality loss a retrofitted global tool often carries.
A clearer, more direct answer on data residency.
Under the DPDP Act, a platform built around Indian hosting and Indian compliance requirements from day one gives a firm a shorter, more direct answer than one built for a different jurisdiction's rules first.
Where LegalOS Fits: Built Natively for Indian Law, Not Adapted to It
This is the specific standard DhiTantra’s LegalOS is built around, and it is worth being direct about it rather than leaving it implied. LegalOS is grounded in Indian statutes, judgments, and procedure from the outset, not a global model with an Indian layer added afterward. Every citation it returns is checked against the actual source it claims to come from, the same retrieval-first discipline that separates SCC Online, Manupatra, and CaseMine’s approach from the failure mode behind the Buckeye Trust and Essel Infraprojects incidents.
Indian Law as the Foundation, Not an Add-On
Research is grounded in Indian statutes and case law from the start, including current coverage of the BNS, BNSS, and BSA transition, rather than retrofitted onto a global model.
Every Answer Traces to a Real Source
Citations are retrieved from an actual, checkable record, the same grounding discipline behind every documented case where a general-purpose tool instead invented one.
Designed Around DPDP From the Start
Data handling is built around Indian regulatory requirements as a starting assumption, not a compliance layer bolted on for a later market entry.
Research, Drafting, and Matter Memory Together
Grounding is only useful if it reaches the document being filed. LegalOS keeps verified research, drafting, and matter history on one shared record per case.
None of this is a claim that global platforms have nothing to offer Indian firms, particularly those handling genuinely cross-border matters where a global research base is directly relevant. The claim is narrower: for the overwhelming majority of Indian legal work, research grounded natively in Indian sources, current with Indian statutory transitions, and built around Indian compliance requirements starts from a stronger position than a global tool retrofitted to reach the same place.