E-Paper
Why India urgently needs practice directions on AI legal drafting
Generative AI does not “know” the law. It is designed to produce plausible, helpful-sounding text, and sometimes that text is fiction
Generative artificial intelligence (AI) has become a common drafting aid for lawyers, and the efficiency gains are real. But the technology has a well-documented weakness: it hallucinates. It produces case names, citations, and quotations that look authentic but do not exist. This is happening with increasing frequency in Indian courts. A bench led by the Chief Justice of India, Justice Surya Kant, recently flagged the problem directly, noting that even where citations are real, fabricated quotations are being attributed to judgments placing an additional verification burden on judges already pressed for time. Similar concerns have been voiced across other judicial fora.
One incident deserves special attention. The Bengaluru Bench of the Income Tax Appellate Tribunal issued an order in Buckeye Trust v. PCIT citing four judgments. Three were invented; the fourth existed but was irrelevant. The order was withdrawn within a week. But the deeper concern is not the error itself, it is who made it. When a judicial officer uses AI to author a judgment or worse, adjudicate a dispute it strikes at the trust litigants place in the adjudication process itself.
The Vacuum Is the Problem
Generative AI does not “know” the law. It is designed to produce plausible, helpful-sounding text, and sometimes that text is fiction. When hallucinated material enters court records, whether through a filing or a judicial order, it makes a mockery of stare decisis by replacing binding precedent with invented authority. A court can reject a defective filing, but it cannot easily undo an order that has been published, relied upon, and perhaps appealed while carrying the imprimatur of judicial reasoning. None of this means AI should be banned. But its use in legal proceedings cannot rest on goodwill alone.
What Other Jurisdictions Have Done
Several jurisdictions have responded to AI-assisted legal drafting with concrete guidelines anchored in familiar principles: educating users about AI’s pitfalls on one hand and mandating verification of authorities, accountability for filings on the other.
Singapore’s “Guide on the Use of Generative Artificial Intelligence Tools by Court Users,” applicable across its court system, does not prohibit AI use but requires that materials be independently verified, accurate, and appropriate, with consequences ranging from costs to disciplinary action for non-compliance.
In the United States, courts such as the Eastern District of Texas have adopted disclosure and certification frameworks requiring lawyers to certify personal verification of authorities, acknowledging that generative AI can produce fake legal authority and insisting on verification and responsibility.
On AI use on the Judicial side, the Federal Court of Canada has published principles limiting AI use in judgments and orders without public consultation, reflecting the higher legitimacy threshold when AI touches adjudication itself. The United Kingdom has gone further, issuing dedicated guidance for judicial office holders that anchors any AI use to the obligation to protect the integrity of the administration of justice, and categorically states that AI is not recommended for legal research or legal analysis.
India, by contrast, has seen only scattered, case-by-case reactions. AI is being used widely, often without a proper understanding of its limits, and without system-level safeguards to prevent hallucinated authorities from entering the record.
A Rulebook Needed
The practice directions India needs do not have to be complicated. Four principles, drawn from international experience, would suffice.
First, reaffirm non-delegable responsibility. AI does not displace professional duties. The filing advocate remains responsible and bound by the duty of candour for what is placed before the court. “AI did it” must never become a defence.
Second, mandate verification through a short certification by filing counsel that every citation and quotation has been personally verified from authentic sources. This draws a workable balance between the convenience of AI and professional responsibility.
Third, set a clear enforcement ladder. Inadvertent errors may warrant correction and costs; recklessness should attract exemplary costs and, where appropriate, disciplinary referral. India should standardise consequences rather than rely on ad hoc judicial displeasure.
Fourth, and most urgently given recent developments, address judicial and tribunal use of AI. The Canadian and UK approaches reflect a foundational point: AI cannot adjudicate. If any AI tool is used in drafting, full responsibility remains with the judicial officer. The reasoning must be independent, traceable to the record, and defensible without reference to any tool.
The Bottom Line
The legal system runs on trust. Lawyers are trusted to place accurate material before courts, and judges are trusted to decide on what the record shows. Generative AI puts both forms of trust at risk. The Supreme Court has signalled alarm. That alarm must now become action. Clear practice directions, short and consistently applied, will give the Bar certainty, protect judicial time, and reassure litigants that the administration of justice is not being quietly rewritten by software that can invent a case as easily as it formats a paragraph. Thoughtful directions, arrived at now and in good faith, will allow India to take the benefits of AI without importing its most dangerous defect into the heart of adjudication.
