Can lawyers rely on AI for case citations? The Supreme Court's red line

On 2 July 2026, the Supreme Court of India set aside orders of the NCLT and NCLAT in the Essel Infraprojects insolvency — not because the law was misapplied, but because the reasoning rested on precedents that do not exist.

Six citations, the Court found, were fake, fabricated or wrongly cited. They had been generated by AI, filed by counsel, relied upon by a tribunal, and survived an appeal — and nobody caught them until the Supreme Court did.

The Court declared zero tolerance for AI-generated fake judgments and directed the Bar Council of India to frame norms for AI use in practice. It was not an isolated incident: legal-news trackers have counted around ten such episodes in Indian courts and tribunals in recent months. Two judicial warnings had already preceded this one. This was the case where warning became consequence.

What actually went wrong in the Essel Infraprojects matter?

Strip the technology out for a moment and look at the chain of events.

A submission was filed containing authorities no one had read. A tribunal adopted them. An appellate tribunal let them pass. At every stage, the system assumed someone else had checked.

The AI is the newest link in that chain — but it is the weakest link only because it is the most trusted. Which is exactly what the Court found alarming: fabricated precedent doesn't announce itself. It arrives formatted correctly, cited confidently, and sounding exactly like the real thing.

Why does AI invent case law?

There is no mystery here, and no malice in the machine.

Large language models generate plausible text, not retrieved truth. When you ask one for case law, it produces what a citation should look like — party names that sound right, a year, a court, a proposition that fits your argument. Sometimes that matches a real judgment. Sometimes the system assembles a citation that has never existed, with exactly the same confidence and exactly the same formatting.

A fabricated citation looks perfect because looking right is the only thing the system is built to do. It is not lying; it is autocompleting. The danger is not that the tool is bad at its job — it is that its job is not the job lawyers assumed it was.

Who answers for an AI-drafted filing?

Here is the principle that survives every technology cycle:

A citation in a filing is not a research note. It is a professional warranty.

When an advocate cites a case to a court, they are personally vouching that it exists, says what they claim, and stands un-overruled. That warranty cannot be delegated — not to a junior, not to a database, and not to a language model. The tool never signs the vakalatnama (the document by which an advocate takes responsibility for a matter). The signer owns every sentence.

None of this is new law. An unchecked citation from a junior's research memo in 1995 carried the same professional risk. What has changed is scale and polish: the errors are faster to produce and harder to spot.

What does a verification-first workflow look like?

The constructive answer is not "don't use AI." Used well, these tools are excellent at structure, summarisation, first drafts, translation, and finding the question you should be asking. The answer is a workflow rule:

1. AI is never a source of authority. It can suggest that a line of cases might exist; it cannot establish that one does.
2. Every citation is verified against a primary database — SCC Online, Manupatra, or the court's own records — before it enters a draft, let alone a filing. Verified means the judgment was opened and the proposition checked, not that the name searched clean.
3. No citation from memory — the machine's or the lawyer's. Human memory hallucinates too; it just does it more slowly.
4. The verification happens at the point of filing, by the person signing. Not upstream, not on trust.

The honest-limits version of the same point: AI in a legal practice is a drafting engine, not a knowledge engine. The moment it is treated as the second, someone eventually explains a nonexistent precedent to a Bench.

What happens next?

The BCI has been asked to frame norms, so Indian practice rules on AI use are now a matter of when, not if. Courts in the US and UK have already sanctioned lawyers for identical failures — this is a global correction, not an Indian embarrassment.

The choice in front of the profession was never AI versus no-AI. That debate is over; the tools are already in every chamber. The real divide — the one the Supreme Court just made consequential — is verified versus unverified.

The lawyers who thrive with these tools will be the ones who understood that the machine changed how fast a draft gets written, and changed nothing about who answers for it.

This area is moving fast — corrections and additions welcome.