Why Indian Lawyers Need Adversarial AI, Not Just Research AI

Opinion | June 2026

A judge in a trial court cited Supreme Court judgments that did not exist. In February 2026, the Karnataka High Court ordered a probe against that judge.¹ The judgments were almost certainly AI-generated — plausible-sounding, formatted correctly, carrying docket numbers and dates, and entirely fabricated. A judge, with far more checking rigour than the average junior associate, missed them.

This is the hallucination problem arriving not as a distant American case study but inside the Indian courtroom system. And it is a symptom of a larger dysfunction in how the profession is thinking about AI. We have been building — and celebrating — research AI. What we actually need is adversarial AI. Until we are honest about the difference, the legal profession is building impressive tools that address one dimension of practice while leaving the more consequential dimension almost entirely unaddressed.

What Research AI Is Doing and What It Cannot Do

Research AI is a retrieval and synthesis engine. It finds relevant judgments faster. It summarizes hundred-page orders. It extracts key ratios. SCC Online's conversational assistant, Manupatra's native legal AI, LegRAA from NIC Pune — these platforms have made genuine inroads into India's research bottleneck.

But research AI does not challenge you. It does not ask: "Have you considered that the Supreme Court's 2022 constitution bench ruling is in tension with the ratio you just cited, and that your opponent will exploit that tension in the first paragraph of their reply?" It does not probe whether your distinction between your facts and the binding precedent will survive cross-examination from the bench. It does not red-team your written submissions before your opponent does.

It finds answers to the questions you ask. It does not ask whether you are asking the right questions.

The Hallucination Problem Is Worse Than You Think

A Stanford Human-Centered AI study found that legal AI tools — not generic chatbots, but platforms specifically built for legal research — hallucinate in more than one out of every six benchmarking queries.² One in six. That is not an acceptable false-positive rate for any professional application, let alone one where the output ends up in court filings.

Former Chief Justice B.R. Gavai publicly warned the Indian bar about AI tools generating fake citations.¹ In the United States, United Kingdom, Canada, and Australia, lawyers have already faced sanctions and disciplinary action for submitting AI-generated citations they did not verify.¹ A January 2025 empirical study from NLSIU found the same pattern: strong performance on drafting and issue spotting, but "frequently generating hallucinations" on specialised legal research.³

The hallucination rate is not a bug that more training data will fix. It is a structural feature of how large language models work: they generate the most statistically plausible continuation of a prompt, not the most factually accurate one. In the domain of Indian case law, plausible and accurate can diverge catastrophically.

What Adversarial AI Would Actually Look Like

Adversarial AI stress-tests legal positions the way a skilled opposing counsel would. It takes your draft submissions and identifies the three most vulnerable arguments the ones you have convinced yourself are strong. It scans the legal landscape for judgments most likely to be cited against you. It models how a bench with a particular interpretive history is likely to receive your argument.

This is not science fiction. It is the natural computational extension of what the best senior advocates do when reviewing a junior's draft. The difference is that a senior advocate does it from memory and experience, constrained by time, for the matters she actively reviews. Adversarial AI would do it systematically, at scale, for every matter in the practice — including the ones that never reach a senior partner before the filing deadline.

Why India Needs It More Than Most Jurisdictions

India's multi-tier judicial architecture — Supreme Court, High Courts with divergent positions on the same questions, specialised tribunals, state legislation alongside central law — creates a genuinely complex precedential landscape. High Courts across different states take different views on the same legal question. Knowing which way a bench is likely to rule requires knowing not just the relevant precedents but which conflicts exist and what the Supreme Court's most recent treatment implies.

Deepak Kapoor of Manupatra identified this precisely: generic AI consistently fails to "distinguish binding vs persuasive precedent" and misinterprets "legal context."⁴ For research AI, this failure is a problem. For adversarial AI, it is a disqualifying defect — because the value of adversarial AI lies entirely in accurately characterising the strength of the threats it identifies. This creates a sequencing requirement: native legal AI is the prerequisite for adversarial AI. You cannot build a reliable challenge engine without first building a reliable knowledge engine.

The Disclosure Regime Creates the Adversarial Dynamic Anyway

The Supreme Court's draft AI Regulations, released June 2026, introduce a disclosure obligation that will create adversarial AI dynamics whether or not the profession explicitly builds for them. When your pleadings disclose AI assistance, your opponent's instructions will shift: find the AI errors in their submissions. Sharp opponents with good research AI will use it to probe the gaps your research AI missed.

If both sides are running AI research, the lawyers who win will be the ones whose AI also helped them find their own vulnerabilities first. Defensive adversarial AI — red-teaming your own submissions before opposing counsel does — becomes a competitive necessity the moment disclosure makes AI use visible.

The Question the Profession Is Not Yet Asking

India's legal AI conversation has been dominated by the productivity narrative: faster research, cheaper drafting, reduced adjournments. These are real and important gains. But the more consequential shift — the one that will change how disputes are decided — is the adversarial shift. AI will change what constitutes thorough preparation.

The profession has 50 million pending cases as a permanent backdrop, a judiciary building native AI infrastructure, and a disclosure regime that makes AI use visible to opponents. The tools being built now are largely in the research-and-summarisation category. The gap between where investment is going and where competitive advantage will actually be won is the adversarial gap. Research AI is the starting point. Adversarial AI is the destination. Building from one to the other deliberately is the profession's real task.