Responsible AI Use in Legal Research and Drafting: A Practical Framework for Indian Advocates

On 17 February 2026, a Supreme Court bench of Chief Justice Surya Kant and Justices B.V. Nagarathna and Joymalya Bagchi - hearing an unrelated PIL - flagged what the Chief Justice called an 'alarming' trend: lawyers filing petitions drafted with AI tools that cite judgments which simply do not exist. Justice Nagarathna cited a specific example she had encountered: a case styled 'Mercy v. Mankind,' which does not exist in any Indian law report. The CJI added that a 'series of such judgments' had been cited before Justice Dipankar Datta's court as well.

'We are alarmed to reflect now - some of the lawyers have started AI to draft. It is absolutely uncalled for,' the Chief Justice said. This wasn't a formal judgment - it was a bench speaking directly to the profession about a pattern it was already seeing repeatedly in filings.

This isn't an isolated incident. In December 2024, the Bengaluru bench of the Income Tax Appellate Tribunal issued an order in the Buckeye Trust matter citing three Supreme Court judgments and one Madras High Court ruling that did not exist - traced to a department representative's unverified use of ChatGPT. The Tribunal recalled its own order within a week of the error surfacing.

In Gummadi Usha Rani v. Sure Mallikarjuna Rao (Civil Revision Petition No. 2487 of 2025), a trial court's order relied on four fabricated Supreme Court judgments while rejecting objections to a court-appointed commissioner's report. On revision, the Andhra Pradesh High Court confirmed the citations were AI-generated and fake - while also holding that the fake citations did not, by themselves, vitiate the underlying order, since the Court found the actual legal reasoning otherwise sound. That's a nuanced outcome worth understanding precisely: the fabricated citations weren't harmless, they were simply not the only thing propping up the order in that instance.

In October 2025, the Bombay High Court quashed a Rs 27.91 crore income-tax assessment after discovering that the National Faceless Assessment Centre's order relied on three non-existent judicial decisions. And in January 2026, the Bombay High Court imposed a cost of Rs 50,000 on a party for what the Court described as dumping fake case laws into written submissions.

In November 2025, the Supreme Court's Centre for Research and Planning released a White Paper on Artificial Intelligence and the Judiciary - part of a set of eight reform papers - framing AI adoption as a governance question, not just a technical one. Its central position: AI is assistive technology, every AI output remains subject to human verification rather than automatic acceptance, and courts should consider structural oversight such as internal AI ethics committees.

That's consistent with the Supreme Court's own separate draft Regulations for Use of Artificial Intelligence in Courts, 2026 (see our companion article), which would require advocates to disclose when AI was used to prepare pleadings, documents, or submissions. Taken together, the direction is consistent: use AI, but verify its output and be prepared to show your work.

For legal research: never cite a case in any filing because an AI tool produced it. Independently pull the actual judgment from a primary source - the court's own website, NJDG, or a reliable legal database - and confirm the case name, citation, court, date, and that it actually holds what the AI output claims it holds, before it goes anywhere near a document a client or a court will see.

For drafting: treat an AI-generated draft exactly like a first draft from a junior associate - useful as a starting point, not a finished work product. Every factual assertion, every citation, and every section reference needs the same review a senior advocate would give a junior's first attempt, not a lighter one just because the draft reads fluently.

For firm-level practice: decide in advance who is responsible for this verification step before a tool is adopted, not after a citation gets challenged in open court. And where a jurisdiction or court expects disclosure of AI assistance - a trend that draft Regulation 43(3) and courts' informal remarks both point toward - build that into the filing checklist rather than treating it as optional.

The Bar Council of India's Standards of Professional Conduct predate large language models and don't yet contain AI-specific provisions. That gap doesn't mean unverified AI-generated filings are risk-free in the meantime - the cases above show courts are already treating fabricated citations as a real problem under their existing powers, through cost orders, recalled orders, and pointed remarks from the bench that put the profession on notice. Waiting for BCI-specific AI rules before adopting a verification discipline is a bet against a trend that is already visible in reported orders.

This is the exact problem source-grounding is meant to solve: an AI system whose output is tied to an identifiable, checkable source lets an advocate verify a citation in seconds rather than reconstructing it from scratch - but grounding is a tool that makes verification faster, not a substitute for actually doing it. See our companion piece on why legal AI needs to be source-grounded, not just fluent, for that distinction in more depth.