Does Legal AI Hallucinate Citations? What Indian Lawyers Need to Know (2026)
Yes, general-purpose AI invents Indian case citations, and Indian courts are now penalising lawyers and even judges who rely on them unchecked. Here is what a hallucinated citation looks like, why it happens, and the four checks every AI-suggested citation needs before it goes into a filing.
Legal AI · Trust and Verification
Yes. General-purpose AI tools such as ChatGPT and Gemini can and do invent Indian case citations, and this is not a rare glitch. Indian courts, including tribunals, High Courts, and the Supreme Court, have already flagged filings and orders built on cases that never existed. The fix is not to avoid AI. The fix is to use AI that is grounded in a real judgment database, and to verify every citation before it goes anywhere near a court filing, exactly as you would verify a citation handed to you by a junior who has not yet earned your trust.
- Does AI make up Indian case citations? Yes. General-purpose, ungrounded AI tools can and do invent cases, wrong propositions, and fake paragraph numbers.
- Have Indian courts caught this? Yes, repeatedly, including a 2026 Supreme Court ruling calling reliance on fabricated AI citations misconduct.
- The fix: use AI grounded in a real judgment database, and always run a four-point verification before filing: exists, court and date match, proposition supported, still good law.
- Even grounded AI is not 100%. Verification is never optional for a filing.
01What counts as a hallucinated citation
A hallucinated citation is not always an obviously fake case name. It shows up in three different forms, and the second and third are the ones that catch experienced lawyers off guard.
A case that does not exist at all
The AI generates a party name, a court, a year, and a citation number that look completely plausible, in the correct format, but no such judgment was ever delivered. Nothing exists to find, because there is nothing to find.
A real case cited for the wrong proposition
The case is genuine and the citation is correct, but the AI attaches it to a legal proposition the judgment never actually decided, or even contradicts. This is harder to catch because the citation itself checks out on a quick search.
An invented paragraph or citation number
The underlying case is real, but the AI makes up a paragraph number, a page number, or a parallel citation that does not correspond to anything in the actual judgment. Reported instances in India include a case cited for paragraphs that do not exist in a judgment far shorter than that.
All three are dangerous in the same way. A wrong or invented authority in a filing is worse than no authority at all, because it invites the court and opposing counsel to question everything else you have cited.
02Why ungrounded AI invents case law
A general-purpose language model like a base ChatGPT or Gemini session does not have a live connection to a database of Indian judgments unless it is specifically built or plugged in to have one. When you ask it for a case on a point of law, it is not retrieving a document. It is predicting, word by word, what a citation on that topic would statistically look like, based on patterns in its training data.
That process is very good at producing text that reads like a real citation, correct format, plausible party names, a year that fits the era of the legal principle. It is not designed to check whether that citation corresponds to an actual judgment anywhere. The model has no internal concept of "this case is real" versus "this case is invented". Both come out with the same confidence.
An ungrounded AI model does not look up a judgment. It predicts what a citation should look like, based on patterns it has seen, which is why a fabricated case can look completely real.
This is why simply asking the AI "are you sure this case is real" rarely helps. The model will often reassert the same fabricated citation with equal confidence, because it is still generating text from the same underlying process, not checking a source.
03The Indian context: courts are cracking down
This is not a hypothetical risk. Indian courts and tribunals have, over the past two years, repeatedly caught filings and even orders built on fabricated AI citations, and the judiciary’s tolerance for it has run out.
Tribunals and High Courts have already been caught out
The Income Tax Appellate Tribunal in Bengaluru recalled an order in a dispute worth roughly Rs 669 crore after it emerged that four case citations relied on did not exist. The Bombay High Court quashed a tax assessment of roughly Rs 28 crore that rested on three non-existent AI-generated precedents, on the ground that it breached natural justice. In Karnataka, a High Court ordered an inquiry after a trial court judge relied on two non-existent Supreme Court decisions while deciding an application. A litigant in Mumbai was made to pay a Rs 50,000 cost after "dumping" AI-generated content that cited a judgment which did not exist.
The Supreme Court has now drawn a hard line
In July 2026, the Supreme Court, in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., set aside orders that had relied in part on fabricated case law and held that citing unverified AI-generated precedents is misconduct on the part of an advocate, and a serious lapse where a judge does it. The Court said an order built even partly on hallucinated material is "no decision in the eyes of the law", and directed the Bar Council of India to frame disciplinary norms for the profession.
The pattern across these cases is consistent. Someone, a litigant, a lawyer, or in one case a tribunal itself, used an ungrounded AI tool to find or draft citations, and skipped the step of checking whether the case was real. The cost was not just an adverse order. It was a professional lapse, a cost award, or a judicial inquiry.
04Grounded AI vs ungrounded AI: the difference that matters
The difference between an AI tool that is safe to use for citations and one that is not comes down to whether it is grounded.
Ungrounded AI is a general chatbot generating an answer from its training data, from memory, with no link back to an actual judgment it can show you. It cannot point to the document the citation came from, because in most cases there is no retrieval step at all.
Grounded AI is built differently. It searches a real corpus of judgments first, retrieves the actual document that matches your query, and then returns a citation that traces back to that specific judgment, one you or the court can open and check. The AI’s job shifts from "compose a plausible-looking citation" to "find and surface a real one".
This distinction is measurable, not just a claim either side can make. See the citation accuracy benchmark for what grounded citation accuracy actually looks like when it is tested, and what makes a citation court-ready for the specific elements a citation needs before it can go into a filing.
05The four-point verification every lawyer should run
Whether a citation comes from a grounded tool, an ungrounded chatbot, or a junior colleague, the same four checks apply before it goes into a filing. Run all four against a primary source, not against the AI that suggested the case.
| Check | What you are confirming | Where to verify |
|---|---|---|
| 1. The case exists | A judgment with this party name, court, and citation actually exists, not just a plausible-looking format. | Indian Kanoon, SCC Online, or the court’s own judgment portal |
| 2. Court and date match | The court, bench, and date cited match the source document exactly. | The judgment itself, cross-checked on Indian Kanoon or SCC Online |
| 3. Proposition is supported | The paragraph you are relying on actually says what you are citing it for, in context. | Read the actual paragraph in the judgment, not an AI summary of it |
| 4. Still good law | The judgment has not since been overruled, reversed, or distinguished on the point you are relying on. | SCC Online citator or Indian Kanoon’s "cited by" references |
None of these checks take long individually. Skipping all four, on the other hand, is what turned an AI suggestion into a professional lapse in every one of the Indian cases above. Verification is a five-minute habit that protects a filing, a client, and a licence.
06Where Claw fits
Claw is an all-in-one legaltech platform for Indian advocates, law firms, and corporate legal teams, combining AI-based case search, an AI legal assistant (Legal GPT), case management, and compliance automation across all Indian courts and tribunals. For case search specifically, Claw is a grounded tool: it retrieves from a base of 30 crore-plus real judgements across the Supreme Court and 25 High Courts, and returns citations that trace back to an actual judgment rather than to the model’s memory, formatted to be court-ready.
That said, grounded is not the same as infallible. No AI tool, Claw included, removes the need for a lawyer to check a citation before it is filed. What grounding changes is the starting point: you are verifying a real, retrievable judgment against the four checks above, not chasing down whether a plausible-sounding case exists at all. See the citation accuracy benchmark for the measured accuracy figures by tier, and the best AI tools for law students and junior advocates for how grounded tools fit into everyday research and drafting.
07Sources and further reading
Primary sources for verifying a citation, and further reading on this issue:
- Indian Kanoon (free case law search): indiankanoon.org
- SCC Online (citation and citator database): scconline.com
- Supreme Court of India (official judgments): sci.gov.in
- Coverage of the Indian AI-hallucination cases and the Supreme Court ruling: iPleaders and MediaNama
- Claw: clawlaw.in
This page does not report every case in this area, and figures and rulings should be reconfirmed against the primary judgment before being relied on in a filing.
08Frequently asked questions
Does ChatGPT make up Indian case law?
Yes, it can. A general-purpose ChatGPT session, without a legal database connected, generates citations by predicting plausible text, not by retrieving real judgments, so it can produce case names, citations, and even paragraph numbers that do not exist. Indian courts have already recorded several instances of this happening in real filings and even in a tribunal order.
How do I know if an AI citation is real?
Check it against a primary source before you rely on it, never against the AI that gave it to you. Search the case name and citation on Indian Kanoon or SCC Online, confirm the court and date match, read the actual paragraph to see it supports your point, and confirm the case has not since been overruled.
Is legal AI safe to use in India?
It is safe to use as a research and drafting aid if you treat every AI-suggested citation as unverified until you check it yourself. Grounded tools that retrieve from a real judgment database are inherently safer than a general chatbot generating from memory, but neither removes the lawyer’s duty to verify before filing.
Have Indian courts penalised fake AI citations?
Yes. Tribunals and High Courts have recalled and quashed orders built on non-existent citations, imposed costs on litigants who filed AI-generated fake case law, and ordered an inquiry into a judge who relied on non-existent Supreme Court decisions. In July 2026, the Supreme Court called reliance on unverified AI-generated precedents misconduct and directed the Bar Council of India to frame disciplinary norms.
Which is safer, a chatbot or a grounded legal AI?
A grounded legal AI is safer for citation work, because it retrieves from an actual judgment corpus and returns a citation you can trace back to a real document, rather than generating one from memory. It is still not a substitute for verification. Even a grounded tool is not 100 percent accurate, so the four-point check remains necessary before anything goes into a filing.