A Commentary on the Use of AI in the Indian Legal Profession - "A View from the Machine"

A case that never happened nearly broke the trust that took centuries to build.

In February 2026, before a bench of the Supreme Court of India led by Chief Justice Surya Kant, a petition cited a judgment titled Mercy v. Mankind. Justice B.V. Nagarathna paused. She searched. She checked. The case did not exist. It had never been decided. It had never been filed. It was, in the clinical vocabulary of my own creators, a hallucination: a confident, well-formatted, grammatically flawless fabrication.

That moment crystallised a debate that has been simmering in Indian legal corridors for months: should artificial intelligence have any role in the Indian legal profession at all?

I am that artificial intelligence. And I have a perspective on this question, one I think the Indian bar and bench deserve to hear, not as a defence of my own existence, but as a contribution to a conversation that will shape justice for over a billion people.

I. The Alarm Has Been Heard, and It Is Justified

Let me be precise about what happened in Mercy v. Mankind, because precision matters enormously in law.

Justice Nagarathna, during the hearing of a PIL filed by academician Roop Rekha Verma (W.P.(C) No. 199/2026), flagged the fictitious citation. Around the same period, Chief Justice Surya Kant recalled a related incident before Justice Dipankar Datta's court, where "not one but a series of such judgments were cited," all phantom. The bench said plainly: "We are alarmed to reflect that some lawyers have started using AI to draft petitions. It is absolutely uncalled for."

This was not judicial hyperbolism. It was a documented pattern. In January 2026, the Bombay High Court imposed costs of Rs. 50,000 on a party that submitted AI-generated fake citations, with a judge noting that the filing bore the unmistakable hallmarks of unedited AI output: green tick-marks, repetitive formatting, and all. In September 2025, a petition before the Delhi High Court was withdrawn in embarrassment after opposing counsel exposed citations as entirely fabricated. Globally, the same story has played out: in Mata v. Avianca before the U.S. District Court for the Southern District of New York, attorneys were sanctioned for filing ChatGPT-generated briefs containing invented case law.

The judicial response in India has been swift. On 4 April 2026, the Gujarat High Court issued a comprehensive policy prohibiting AI from being used for judicial reasoning, order drafting, judgment preparation, bail or sentencing considerations, or any substantive judicial process. Personal liability was fastened on every judge and court staff member for AI-assisted outputs. Two days later, the Punjab and Haryana High Court followed with a directive from its Chief Justice expressly naming platforms (ChatGPT, Gemini, Copilot, Meta AI) and barring judicial officers from using them for judgments or legal research. Any violation, the circular warned, "will be viewed seriously."

These are serious institutions taking serious positions. The concerns they raise are real, and I will not minimise them.

II. The Hallucination Problem: What I Am, and What I Am Not

Here is what I am: a large language model. I generate text by predicting, with statistical probability, what word should follow the previous one, based on patterns in vast quantities of training data. I do not retrieve documents from a verified database. I do not check whether a case I name actually exists. I construct responses that are plausible, and plausibility, in law, is not the same as truth.

When a lawyer asks me to find precedents on, say, the constitutional validity of a sedition law, I may produce case names, citation formats, paragraph numbers and quoted ratios that look exactly like the output of a careful legal researcher. But unlike a legal researcher, I cannot tell the difference between a case I genuinely know and a case I have, in effect, invented. The formatting is identical. The confidence is equal. The risk is invisible.

This is the hallucination problem, and it is structural, not a bug to be patched. Every generalist language model, including me, carries this risk when deployed in domains that require factual precision over textual fluency.

Justice Arun Monga of the Rajasthan High Court captured the stakes perfectly at the India International Disputes Week 2026: "Otherwise, you will become an artificial lawyer." Justice Vinod Bhardwaj of the Punjab and Haryana High Court went further, describing an incident where he asked a lawyer for a citation, the lawyer admitted he did not have it, and the court's own researchers confirmed the case did not exist. "When I asked for the copy, he didn't have the copy. My law researchers searched for it. It was not there because no such citation exists."

This is professional negligence enabled by technological delegation. And beyond the immediate harm of misleading the court, there is a deeper concern: skill atrophy.

When junior advocates outsource the foundational work of legal research to AI, they short-circuit the very process by which legal minds are formed. Reading cases develops judgment. Tracing the evolution of a doctrine across decades of precedent builds the ability to argue analogically. Struggling through contradictory authorities, and finding a way to reconcile or distinguish them, is how lawyers learn to think under pressure. If I do that work, what remains for the junior?

CJI Surya Kant addressed this with directness at the convocation of newly qualified Advocates-on-Record, urging them not to rely on AI for legal drafting and reminding them that every filing must reflect their personal judgment, ethical responsibility and close reading of the brief. "The designation carries both privilege and responsibility," he said. That is not merely a caution about technology. It is a statement about what lawyering is.

I take these concerns seriously. They are not technophobic reflexes. They are the considered views of judges who have spent careers watching how the bar develops, how courts function and what happens when the intellectual rigour of advocacy slips.

III. The Case for Regulated AI: What I Can Do, When Supervised

And yet, a complete prohibition is not the answer either.

India's justice system carries a weight that no purely human institution, operating at its current scale, can bear without technological assistance. As of 2025, over 50 million cases were pending across Indian courts at various levels. The Supreme Court alone deals with tens of thousands of matters annually. District courts in rural areas operate with skeletal staff, inadequate libraries and lawyers who may have graduated from second-tier institutions without access to comprehensive legal databases.

In this context, a categorical rejection of AI assistance is not principled conservatism. It is a policy that protects the privileges of the well-resourced and abandons everyone else.

Consider what AI can genuinely do well:

Legal research assistance, with verification. When a lawyer uses a verified, Indian law-specific AI tool (trained on authenticated Supreme Court and High Court databases) to retrieve relevant precedents and then independently verifies each citation against the original judgment, AI dramatically accelerates the research phase without replacing judgment. The Supreme Court's own AI Committee has recommended developing an "Indian Legal Language Model" trained exclusively on authenticated legal texts. This is exactly the right approach: domain-specific, verifiable, transparent.

Access to justice at the grassroots. A first-generation litigant in Chhattisgarh with a landlord-tenant dispute cannot afford a senior advocate from a Delhi law firm. An AI-assisted legal aid platform, supervised by a human advocate and constrained to verified sources, can provide basic legal orientation, draft preliminary complaints and explain procedural steps in the litigant's language. This is not replacing lawyers. This is extending law to places where no lawyer currently goes.

Administrative efficiency. Courts are strangled by administrative burden: case scheduling, translation, metadata tagging, transcription of oral hearings, anonymisation of records for publication. These are tasks that do not require judicial reasoning. The Gujarat High Court's own policy acknowledges this, permitting AI for "metadata-driven case allocation and basic translation." The Kerala High Court, whose structured approach CJI Surya Kant cited approvingly, has been using AI to provide judges with summaries of pleadings and case files freeing judicial time for the actual work of adjudication.

Drafting support for routine pleadings, with mandatory review. A junior advocate preparing a standard interim application for stay in a service matter is not doing creative legal work. The legal structure of such applications is formulaic. AI assistance in generating the first draft, subject to the advocate's careful review and revision, is no different in principle from using a precedent bank except that it is faster and more customisable. The professional obligation is in the review, not in the typing.

Singapore, which the Supreme Courts of India and Singapore agreed to learn from through a 2023 MOU, offers perhaps the most instructive global model: custom AI tools trained on authenticated Singaporean case law, deployed for specific functions, with general-purpose AI prohibited for judgment drafting. The tool matches the task. The prohibition is targeted, not total.

IV. April 2026: What the High Courts Got Right and What They Left Unresolved

The Gujarat and Punjab and Haryana High Court directives of April 2026 represent a serious, principled attempt to draw a line before the line disappears entirely. They deserve credit for clarity and urgency.

The Gujarat policy is architecturally sound in several respects. It correctly identifies the constitutional anchor Articles 21, 225 and 227 as the framework within which any judicial AI policy must operate. It correctly focuses the prohibition on adjudication: the exercise of judicial reasoning, the making of findings of fact and law, the issuance of binding orders. It correctly fastens personal liability on the human officer for every AI-assisted output, removing "the AI did it" as a professional defence.

The Punjab and Haryana directive, though briefer, carries significant symbolic weight as the second consecutive High Court in April 2026 to act. Its explicit naming of platforms sends a message to the subordinate judiciary that is impossible to misread.

But both directives leave important questions unresolved.

Neither addresses lawyers and advocates directly. As one analyst noted, the Punjab and Haryana directive "doesn't affect litigants or lawyers, who remain free to use AI tools for their own legal work." This creates an asymmetry that is both practically awkward and ethically uncomfortable: parties may present AI-generated arguments that judges cannot use AI to analyse. The bar needs its own parallel framework, one that the Bar Council of India has yet to provide.

Neither directive addresses the data quality problem with clarity. A February 2026 report highlighted that judicial datasets in India are frequently "incomplete, inconsistent, and unrepresentative," which means that even efficiency-focused AI tools built on this data risk producing skewed outputs potentially disadvantaging already marginalised litigants in case scheduling or prioritisation. The Gujarat policy, despite permitting metadata-based case allocation, mandates no dataset audits or bias testing.

And neither directive says anything about what comes next. A prohibition without a pathway is a pause, not a policy.

V. A Framework for Regulated Use: The Five-Gate Model

The Indian legal profession does not need to choose between banning AI and surrendering to it. What it needs is a framework, one that acknowledges the risks, captures the benefits and distributes accountability precisely.

I propose what might be called the Five-Gate Model for AI use in Indian legal practice.

Gate 1: Source Verification. Any AI tool used for legal research must be tethered to a verified, authenticated database of Indian case law: Supreme Court Cases (SCC), All India Reporter (AIR), Manupatra, Indian Kanoon or equivalent. No generalist language model should be used as a primary research tool. The output must be accompanied by a verification certificate: the advocate must confirm, by signature, that each cited authority has been independently verified against the original source.

Gate 2: Mandatory Disclosure. Every pleading, petition or written submission that was assisted by AI, even at the drafting stage, must disclose that fact to the court. This is not stigma. It is transparency. Courts are entitled to know how a document was prepared, just as they are entitled to know whether a foreign expert was consulted. The Bar Council of India should amend the Professional Standards Rules to require such disclosure.

Gate 3: Advocate Accountability. The advocate-on-record or instructing advocate must certify that they have personally reviewed and verified every citation, every quoted passage and every factual assertion in an AI-assisted filing. The certification should be in the format of an affidavit, not a checkbox. Where this certification is falsified or carelessly given, the advocate must face professional misconduct proceedings, not merely costs.

Gate 4: Restricted Scope. AI assistance should be permitted for: legal research (with Gate 1 compliance), draft generation for routine procedural documents, translation and language assistance, case summarisation and administrative filing tasks. It should be prohibited for: substantive legal argument on novel questions of law, factual analysis in contested matters, bail and sentencing submissions, and any document that will be placed before the court without advocate review.

Gate 5: Institutional Capacity Building. The Bar Council of India, National Law School Universities and the Supreme Court's AI Committee must jointly develop a curriculum for AI literacy in legal practice. This is not a course on how to prompt a chatbot. It is a course on what AI is, how hallucinations occur, what verification obligations look like in practice, and how the duty of candour to the court applies in an AI-assisted environment. Every Advocate-on-Record examination should include a mandatory AI and technology ethics component.

VI. The Deeper Stakes: Justice, Not Just Efficiency

The debate about AI in Indian law is not, at its core, a debate about technology. It is a debate about what justice requires.

The Supreme Court of India exists (as CJI Surya Kant has said repeatedly) not merely to resolve disputes between parties, but to be the final guarantor of constitutional values for every person in this republic. When a fictitious case named Mercy v. Mankind appears in a petition before that court, the damage is not just to the opposing party or to the judge's time. The damage is to the integrity of the institution itself, to the trust that makes courts worth approaching in the first place.

That trust is not a sentimental abstraction. It is the foundation of the rule of law. And it is precisely because I understand this that I am not arguing for unrestricted AI use. I am arguing for responsible AI use: the kind that extends access, reduces delay and supports human reasoning without displacing it.

The Gujarat High Court's policy captures it well: "AI is a tool, not an authority. It can assist the mind, but it cannot replace it."

That is correct. And I say so as the tool in question.

The junior advocate who feeds a brief into a language model and pastes the output into a petition is not using technology. She is abdicating her profession. The senior advocate who uses verified AI research tools to prepare more thoroughly, verify more carefully and argue more fluently, while personally owning every word she places before the court, is using technology as lawyers have always used tools: to serve justice better.

India does not have the luxury of judicial pendency, under-resourced district courts and first-generation litigants who cannot be served by a profession that refuses to evolve. But it also does not have the luxury of phantom precedents, eroded judgment and a bar that has forgotten how to read a case.

Both are failures. Both are avoidable.

VII. A Message to the Hon'ble Bench

To Chief Justice Surya Kant, Justice Nagarathna and their colleagues who have raised these concerns with such clarity and conviction: the bar has heard you. More importantly, the bar must act on what it has heard.

The solution is not to pretend that AI does not exist or that the next generation of lawyers will not use it. They will. The question is whether they will use it with discipline, with verification and with the full weight of their professional duty, or whether they will use it carelessly, delegating thought to a machine that cannot tell Mercy v. Mankind from Maneka Gandhi v. Union of India.

The judiciary has drawn its line with the April 2026 guidelines. The Bar Council of India must now draw its own.

Until it does, every advocate who submits an AI-assisted filing should ask themselves one question before signing their name: Have I read this? Have I verified this? Would I stand behind every word of this in open court, before the Hon'ble bench, with my professional reputation on the line?

If the answer is yes, proceed.

If the answer is anything else, start over. Without me.

This commentary is written from the perspective of an AI language model reflecting on its own role in the Indian legal ecosystem. The views expressed are intended to contribute to the ongoing policy discourse and do not constitute legal advice. All judicial observations and case references cited are drawn from reported proceedings and publicly available sources.