The story we have been telling ourselves about AI hallucinations in court goes something like this: a junior lawyer, under time pressure, runs a query through a general-purpose AI tool, pastes the output into submissions, and files it without checking. The judge catches it. Sanctions follow. Everyone learns a lesson about professional responsibility, and careful practitioners take comfort that the problem belongs to the careless ones.
On 2 July 2026, India's Supreme Court rendered that story obsolete.
The matter was Essel Infraprojects. The fabricated precedents did not come from a careless lawyer. They came from the tribunal's own research. The tribunal sourced them, relied on them, and built a decision on them. The matter went up on appeal, and the hallucinations sailed through undetected. India's Supreme Court eventually caught it, declared "zero tolerance" for AI-generated fake precedents, and in doing so sent a signal to every common law jurisdiction paying attention. MediaNama reported the judgment on 3 July 2026.
There is a reason this matters differently from every other hallucination story published in the past two years. When counsel files fake cases, the framing is discipline: a professional duty was breached, sanctions are available, the system corrected itself. The implied message is that careful lawyers are safe. When a tribunal generates fake cases from its own AI-assisted research, and they survive a full appellate tier, the discipline framing collapses. The exposure is ungrounded retrieval anywhere in the chain, regardless of which party did the retrieving.
Australian practitioners might read the Essel Infraprojects ruling with a certain detachment. This happened in India. We have the Federal Court's GPN-AI, effective from 16 April 2026. We have our own sanction cases. We have bar associations with specific guidance. The regulatory framework is taking shape, and responsible practitioners are already inside it.
That detachment is worth examining carefully. The GPN-AI and its state equivalents are correctly designed around practitioner responsibility: verify what you file, disclose what you used, accept professional accountability for everything that bears your signature. Those are enforceable obligations and they create the right incentives for counsel. They do not, and cannot, govern how research gets done before a submission reaches a desk. They do not govern a judicial officer who opens an AI tool for background research. They do not govern the associate assisting with a bench memo.
The GPN-AI, carefully drafted as it is, governs practitioners appearing before the Federal Court. It does not extend to judicial research workflows, bench memoranda, or the research habits of registrars and associates. That is not a criticism of the practice note's drafters; regulating how a judicial officer researches the law is a different and constitutionally more complex exercise. The perimeter of the current compliance framework stops at the practitioner's signature, and that is worth being clear about.
The Essel Infraprojects matter is a stress test of the disclosure-and-verification model. It shows what happens when the reliance point sits upstream of the signature obligation, and when the person relying has no external review mechanism checking their work. Counsel did not put the fake precedents before the tribunal. The tribunal found them itself. There was nobody with a professional duty to check.
We should be honest about what that means. The compliance posture that has developed in Australian courts over the past 18 months is the right posture: verify, disclose, accept responsibility. The Essel Infraprojects ruling adds a necessary extension to it. Verification has to be architecturally possible. The tools generating the authorities have to be built so that tracing them is a structural feature, not an afterthought.
The Supreme Court of India's zero-tolerance declaration is worth reading closely, because it is not a prohibition. Nobody credible is saying "do not use AI" anymore. The Court is saying that a legal authority placed before a court must be real: it must exist, it must say what the citing document says it says, and someone must have confirmed that before reliance.
That is a grounding standard. The question it puts to every legal process, judicial and professional alike, is whether the authority being cited traces to a verified source, or whether it emerged from a retrieval system optimised for fluency rather than accuracy.
Fluency is the trap. A well-written hallucination looks like a case. The citation format is plausible, the legal reasoning is coherent, the style matches the jurisdiction. A model trained to produce text that resembles legal output will produce text that resembles legal output, whether or not the underlying authority exists. The Essel Infraprojects tribunal relied on something that felt like a case. So did the appellate tier that reviewed it. A trained judicial mind, reading the decision on appeal, did not catch the invented authorities, not because the review was inadequate, but because nothing in the face of a plausible citation marks it as fabricated. The detection mechanism has to sit in the retrieval layer, not in the review layer. The courts that have moved decisively on this problem share a common insight: verifiability is a structural requirement, not a downstream check on an otherwise sound process.
Much of the legal AI conversation has been about prompting: knowing what to ask and how to frame it. That is useful knowledge. A better-constructed query returns better output. But it does not change the fundamental architecture of a general-purpose language model, which predicts plausible tokens. When those tokens happen to describe a case that does not exist, nothing in the model signals the error. The fluency is identical to the fluency of a real citation.
A system built on closed primary law sources behaves differently. The retrieval step is constrained to what exists in the corpus. If a case is not in the corpus, it cannot appear in the output. Citations trace to the source document because the source document is what the system queried. The output reflects what the law says, drawn from verified primary materials and attributed to them, rather than a plausible rendering of what it might say.
Habeas searches over 300,000 Australian cases and pieces of legislation from a closed dataset of verified Australian legal sources. Results are grounded in that corpus and traceable to it. The system cannot cite what the corpus does not contain. That architectural constraint is, under the Essel Infraprojects logic, the relevant one. Barristers running authority gathering on the platform work from sources traceable to the source document; foundational research that used to take a full morning can be completed in minutes, with every authority verifiable before it reaches a submission. The speed matters, but the grounding is what the Indian Supreme Court's ruling puts a name to.
The Essel Infraprojects matter poses a question that GPN-AI compliance alone cannot answer: is every authority in this matter traceable to a real source, regardless of how it arrived and who found it?
For counsel, the answer starts with how research is done before a submission is drafted. For the bench, the ruling is a direct prompt. For firms advising on AI governance, it extends the perimeter of the question considerably, because the risk is not confined to filed documents. It lives in any workflow where a general-purpose model returns legal authorities without constraint.
Zero tolerance for fake precedents is a workable standard only where the retrieval layer makes fabrication structurally improbable. That is a design question as much as a discipline question, and it is the design question the Essel Infraprojects matter puts squarely on the table. The sanction cases that preceded it were easy to read as problems belonging to other people. This one is harder to read that way.
See how Habeas approaches the grounding problem at habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
