Fake citations, real consequences: what Indian courts have actually done about AI-generated case law

On 2 July 2026, a two-judge bench of the Supreme Court of India set aside an insolvency order because the reasoning underneath it rested on judgments that had never been delivered. Justice P.S. Narasimha, writing with Justice Alok Aradhe, reached for an unusually violent image:

"The production of fake, non-existent, and hallucinated material and its utilisation as precedents in law, is like the release of methyl isocyanate in the province of law and justice: invisible, insidious, and catastrophic by the time anyone notices. It not only contaminates but takes away the very lifeblood of judicial determination."

Methyl isocyanate is the gas that leaked at Bhopal. It is not a metaphor an Indian judge selects casually.

The case was Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668 — a Section 7 IBC petition against Essel Infraprojects in which the NCLT had admitted the application relying on authorities that did not exist, and the NCLAT had affirmed without noticing. Both orders went. The matter was remanded. And the Bar Council of India was directed to constitute a committee to frame guiding principles and disciplinary consequences.

For advocates, the operative sentence is shorter and colder than the Bhopal line:

"It is a misconduct on the part of an advocate to cite such judgments without verification."

This article sets out what has actually happened in India — every documented matter, what the sanction was, and who bore it — and then what has happened everywhere else, because the gap between the two is the most useful thing an Indian advocate can know right now.

The Indian record, in full

Nine matters are documented on the public record as of 1 September 2026. They are worth reading as a list, because the pattern in the right-hand column is not the one most people expect.

One clarification, because it is widely reported wrongly. The Delhi High Court order in Greenopolis — the September 2025 matter — never mentions artificial intelligence. The operative text runs to two paragraphs and says only that "some of the judicial precedents cited on behalf of petitioner do not even exist and in some of the precedents, the quoted portions do not exist." The AI attribution comes from commentary, not from the court. It is a fair inference. It is not a finding, and anyone citing it as one is doing the thing the case is about.

The finding hiding in that table

Read the sanctions column again. In nine documented matters:

• No Indian advocate has been fined, suspended, struck off or held in contempt for filing AI-fabricated citations.
• The one financial penalty — the ₹50,000 in Bahry — fell on a company director appearing in person, not on a lawyer.
• No Bar Council disciplinary order on the subject is publicly documented. The Supreme Court directed the BCI to constitute a committee on 2 July 2026; there is no public confirmation that it has met or reported.

And in five of the nine, the fabricated citations were produced not by counsel but by a judge, a tribunal, or a tax officer. That is the reverse of the position in the United States and England, where the fabricator is almost always the lawyer who signed the filing. It is why the Supreme Court's remedy in Pooja Ramesh Singh was addressed to judges as well as advocates, and it is the single most distinctive feature of the Indian picture.

It is also why the High Courts have moved first on their own officers. Kerala issued the first binding Indian judicial AI policy in July 2025, barring district judges from using AI "to arrive at any findings, reliefs, order or judgment." Gujarat followed in April 2026, prohibiting judges and staff from using AI to author or substantially compose any judgment even if the judge later reviews it. Punjab and Haryana circulated a similar bar. Every one of those instruments regulates the bench. None of them regulates the bar.

What the law now requires of you

Two things changed in 2026, and they point in different directions.

The first is settled. Pooja Ramesh Singh states the rule: citing an AI-generated judgment without verification is misconduct. The Court went further on consequences than most advocates realise. A decision touched by fabricated material is to be set aside

"even if an iota of fake or hallucinated material enters the decision-making process, as it would violate the sanctity of adjudication"

and is "no decision in the eyes of the law, irrespective of whether such material had a direct or indirect bearing on the decision-making."

That is a deliberate correction. Five months earlier, the Andhra Pradesh High Court had taken the opposite view in the Gummadi Usha Rani revision: fake citations "would not vitiate" an order "when the law stated in the order and its application is correct." After 2 July 2026, that reasoning does not survive. Materiality is no longer a defence.

The second is unsettled. In June 2026 the Supreme Court's AI Committee published Draft Regulations for Use of Artificial Intelligence in Courts. Regulation 43(3) would require any party or legal representative using AI to prepare a document, pleading or evidence to disclose that fact at the time of filing; Regulation 43(6) places full responsibility on the person filing if the document turns out to be fabricated because of its AI-generated character.

Those regulations are in consultation, not in force. The Supreme Court Advocates-on-Record Association has opposed the disclosure rule as unworkable, arguing that advocates are already professionally responsible for accuracy and that a certificate of personal verification would do the same work more honestly. Justice Prathiba M. Singh of the Delhi High Court has warned publicly that a declaration requirement could become "another compliance ritual."

The practical position today: you have no disclosure duty, and complete responsibility.

What happens elsewhere, and why it matters here

India's leniency so far is not the global norm. It is worth knowing the range, because Indian benches are already citing foreign practice, and because the BCI committee will look at it.

The scale is not incidental. A public database maintained by Damien Charlotin, a research fellow at HEC Paris, had logged 1,598 court decisions worldwide involving AI-hallucinated material as of 9 June 2026 — against roughly 120 in May 2025. Charlotin counts only matters where a court expressly found or clearly implied reliance on hallucinated content, and describes the total as an undercount.

May 2025120

January 2026719

April 20261,227

June 20261,598

AI Hallucination Cases Database, maintained by Damien Charlotin (HEC Paris), figures as of 9 June 2026. Counts only decisions where a court expressly found or clearly implied reliance on hallucinated material; the maintainer describes the total as an undercount. Coverage skews towards the United States because of public docket access.

The duty, as courts have actually described it

The most complete statement anywhere is not Indian. It is the English Divisional Court in Ayinde v Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), decided by the President of the King's Bench Division. It is worth quoting at length because it disposes of the three excuses advocates reach for.

On what these tools are:

"Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research... They may cite sources that do not exist. They may purport to quote passages from a genuine source that do not appear in that source."

On the duty that follows:

"Those who use artificial intelligence to conduct legal research notwithstanding these risks have a professional duty therefore to check the accuracy of such research by reference to authoritative sources, before using it in the course of their professional work."

On the "my junior did it" defence:

"This duty rests on lawyers who use artificial intelligence to conduct research themselves or rely on the work of others who have done so. This is no different from the responsibility of a lawyer who relies on the work of a trainee solicitor or a pupil barrister."

And on the "my client gave me the citations" defence, which was the actual facts of Al-Haroun:

"A lawyer is not entitled to rely on their lay client for the accuracy of citations of authority or quotations that are contained in documents put before the court by the lawyer. It is the lawyer's professional responsibility to ensure the accuracy of such material."

Justice Myers of the Ontario Superior Court reduced it to a list in Ko v Li, 2025 ONSC 2766, which the English court reproduced in full. Two lines carry the whole thing:

"It is the lawyer's duty to ensure human review of materials prepared by non-human technology such as generative artificial intelligence."

"It should go without saying that it is the lawyer's duty to read cases before submitting them to a court as precedential authorities."

The Bombay High Court's formulation in Bahry is the Indian equivalent, and it is notably not anti-technology:

"If an AI tool is used in aid of research, it is welcome; however, there is great responsibility upon the party, even an advocate using such tools, to cross verify the references and make sure that the material generated by the machine/computer is really relevant, genuine and in existence."

Where the rule actually bites

The Ninth Circuit made the sharpest point of all in June 2026, and it is the one to remember:

"The rules are not violated at the point of research and drafting, but at the point of signing and filing. If an attorney files a brief with cases or quotations that do not exist... it generally does not matter if he pulled the hallucination or misrepresentation from the output of an artificial intelligence tool or from his own natural intelligence."

Using AI is not the offence. Filing unverified work is. That distinction is what makes this a workflow problem rather than a question of whether to adopt technology at all.

The same court was equally clear about what turns a mistake into a catastrophe. The six-month suspension was not for the fabricated cases; it was for what came after:

"If, in the Motion to Correct, the attorneys had disclosed that AI was used in the opening brief against firm policy and apologized for failing to check the brief, lesser sanctions may have been warranted. But that is not what they did."

There is a matching data point on the other side. In United States v. Cohen, a lawyer who promptly disclosed that she had been unable to verify false citations in a colleague's brief was found not to have acted in bad faith. Prompt candour changed the outcome.

What this means for how you work

The honest summary of the Indian position, as of today:

1. The rule is stated. Unverified AI citations are misconduct, per the Supreme Court.
2. Nobody at the Indian bar has yet paid for breaking it.
3. A committee has been directed to decide what the consequences should be.
4. The draft court regulations would add a disclosure duty on top.

Anyone reading that as breathing room is reading it wrongly. The enforcement gap is the gap between a rule being announced and the first advocate becoming the example. Ayinde declined to start contempt proceedings and said so in terms that should be read twice:

"This court's decision not to initiate contempt proceedings in respect of Ms Forey is not a precedent. Lawyers who do not comply with their professional obligations in this respect risk severe sanction."

The practical response is not to avoid AI. It is to stop using tools that can invent an authority in the first place.

The distinction is architectural, not cosmetic. A general-purpose language model produces a citation the same way it produces any other sequence of words — by generating what a citation for this proposition would plausibly look like. It has no index of judgments to consult and no way to report that nothing matched. That is why the failures look the way they do in the record above: not garbled text, but confident, correctly-formatted, entirely fake case names, and real judgments quoted at paragraphs they do not contain.

A retrieval-grounded system fails differently. It searches an actual corpus, and when nothing matches it returns nothing. It cannot cite El Gendi v Camden LBC because El Gendi v Camden LBC is not in the database, because it does not exist.

How Synapse handles this

We built Synapse around this failure mode, because it is the one that ends careers.

Every authority in a Synapse draft is fetched from an indexed corpus before the sentence is written — 12,305 sections of Indian statute and judgments retrieved from a real database, not recalled from a model's weights. The draft is assembled from what was retrieved. If nothing is retrieved, nothing is cited.

Every citation opens. Hover any line in a Synapse draft and the source appears: the provision, the paragraph of the judgment, or the moment in the consultation recording where your client said it. You are not asked to trust the output. You are shown what it was built from, so verification takes seconds instead of an afternoon.

Sections map both ways. Search by the numbering you know — the 2023 codes and the provisions they replaced are indexed against each other, so an IPC number still finds you the BNS section.

None of this removes your duty to read what you file. It is not designed to. What it removes is the category of error where the thing you are checking was never real to begin with.

---

Sources for every case and quotation in this article are linked inline. The Supreme Court judgment discussed is Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668. Paragraph numbers reported in secondary sources should be confirmed against the judgment text before being relied on in a filing.