How To · AI Citations
How to Verify an AI Generated Legal Citation in India
An AI tool gives you a case name, a citation and a crisp holding, and it reads like something a senior would hand you. The problem is that an AI model can produce a citation that is entirely invented, with a plausible party name, a real-looking reporter string and a confident proposition, and nothing about how it looks will tell you it is fake. The only way to know is to verify it. This guide takes a slightly different angle from a general check: it shows you the red flags that a citation was AI-generated rather than retrieved, and then how to verify each component of it against free primary sources. For the quick two-minute version, see the companion guide; this one is about recognising and dismantling a fabricated citation. Last verified October 2026.
- Why it happens: an ungrounded AI model assembles a plausible citation pattern, so a fake looks exactly like a real one.
- Red flags: no source shown, a suspiciously perfect holding, an off-looking citation string, a court or date that does not fit, or a citation that shifts when you ask again.
- Verify by component: party names, citation string, court and date, the proposition, and whether it is still good law, each against a primary source.
- If it will not verify: search for a real authority for the point, and drop the AI citation; never dress a fake up for filing.
- Your duty: the responsibility for a cited authority is the lawyer's, not the tool's.
01Why AI invents citations in the first place
A general-purpose AI model predicts the next most likely piece of text. When you ask it for authority on a point, it produces something that looks like a citation, because citations in its training data follow a pattern: a party name, a versus, a year, a reporter abbreviation, a number. It can assemble that pattern convincingly without any real case behind it. This is not a bug you can prompt away; it is how an ungrounded model works.
That is why a fabricated citation is so dangerous. It does not look wrong. It follows every convention of a real one, which is exactly why you cannot judge it by appearance. For the fuller mechanism, see why legal AI hallucinates citations.
02Red flags that a citation is AI-generated, not retrieved
Some signs raise the odds that a citation was generated rather than pulled from a real database. None is proof on its own, but any of them means verify before you rely:
- The tool cannot show a source. If it will not give you a link or a database reference to open, treat the citation as unverified.
- The holding is suspiciously perfect. A case that supports your exact point in your exact words, with no nuance, is the kind of thing a model produces to please you.
- The citation string looks slightly off. A reporter abbreviation that is not quite right, an implausible volume or page number, or a neutral citation format that does not match the court.
- The court and year do not fit. A proposition attributed to a court that would not decide that kind of matter, or a date that predates the law it supposedly interprets.
- The same case keeps shifting. Ask twice and the citation number or year changes, which a real retrieved case would not.
A fabricated citation is built to pass a glance. The red flags are not in how it reads, they are in what you cannot confirm when you go looking.
03Verify it component by component
A citation is made of parts, and an AI model can get some right and others wrong, blending two real cases or attaching a real holding to a wrong citation. So verify each component against a primary source, not the citation as a whole:
| Component | What to confirm | Where to check |
|---|---|---|
| Party names | A real, matching case exists with these parties in this court | Indian Kanoon, the court's own website, eCourts |
| Reporter / neutral citation | The citation string is real and points to the same case | Indian Kanoon, a reporter or research platform |
| Court and date | The bench and the judgment date are consistent with the citation | The primary copy of the judgment |
| Proposition | The judgment actually states the point the AI attached to it | The full judgment text, read in context |
| Good law | It has not been overruled, reversed or stayed since | Citing references on a research platform |
The order matters. If the party names do not resolve to a real case, stop; there is nothing to verify further. If they resolve but the citation string does not match, the AI may have blended cases. Only once the case is confirmed real and correctly identified is it worth reading for the proposition and checking it is still good law. For the streamlined version of this flow, see the two-minute four-step check.
04What to do when a citation will not verify
Sometimes a citation simply does not check out. Handle it in order:
- Ask the tool for its source. A grounded tool can show the judgment it retrieved. If it cannot, that is informative in itself.
- Search the proposition, not the citation. If the legal point is sound, there is usually a real authority for it. Find that real case and cite it instead of the AI's version.
- Do not reverse-engineer a fake into a filing. Never tidy up an unverifiable citation to make it look filed-ready. An invented citation that reaches a court is a professional problem, not a formatting one.
- Drop it. If no real authority supports the point, the point is not supported, however confidently the AI stated it.
05Your professional duty does not transfer to the tool
The responsibility for what goes into a filing is the lawyer's, not the tool's. Courts have made clear that citing a non-existent authority is the person's failure, regardless of how the citation was produced. An AI tool is a research aid; it does not carry the duty of candour to the court, and it cannot be blamed when an invented case is cited.
Practically, that means the verification step is not optional polish. It is the part of the work that an AI tool does not do for you, and it is the part that protects you. Treat every AI-generated citation as unverified until you have confirmed it against a primary source yourself.
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.
The whole reason this check exists is that an ungrounded tool gives you a proposition with no checkable source, so the verifying falls on you. A grounded legal tool is designed to remove most of that work: instead of free-writing a case name, it searches an actual judgment database and returns the specific citation it found, with the source shown, so you can open the same judgment in one step. Claw's case search and Legal GPT work this way, grounded in 30 crore+ judgments with Supreme Court decisions from 1950 and High Court decisions from 1990, returning verified, court-ready citations in under 5 seconds, and Claw does not use your documents to train its AI models. That removes the first components of the check above; reading the judgment for the exact proposition and confirming it is still good law remain good practice on any citation from any tool. You can create a FREE Claw account to try it. For the measured figures, see the citation accuracy benchmark, and for the tools compared, the best legal GPT for Indian judgments.
07Sources and further reading
- Indian Kanoon (free case law search): indiankanoon.org
- Supreme Court of India (official judgments): sci.gov.in
- eCourts (district and High Court case status): ecourts.gov.in
- SCC Online: scconline.com
- Claw: clawlaw.in
Last verified October 2026. Court and platform features change over time; always confirm against the primary source at the time you check. This is general information, not legal advice.
08Frequently asked questions
How can I tell if an AI made up a case citation?
You cannot tell from how it reads, because an AI model assembles citations that follow every convention of a real one. You tell by verifying: confirm the party names resolve to a real case, the citation string points to the same case, the court and date fit, the judgment states the proposition, and it is still good law, each against a primary source like Indian Kanoon or the court's own website.
What are the warning signs a citation was AI-generated?
The tool cannot show a source, the holding supports your point too perfectly, the citation string looks slightly off, the court or year does not fit the proposition, or the citation changes when you ask the tool again. Any one of these means verify before you rely on it.
What should I do if an AI citation does not check out?
Ask the tool to show its source, then search for the legal point itself rather than the citation, because if the point is sound there is usually a real authority for it. Cite that real case instead. If no real authority supports the point, drop it. Never edit an unverifiable citation to make it look filing-ready.
Is it my fault if I cite a case the AI invented?
Yes. The responsibility for a cited authority is the lawyer's, not the tool's, and courts have treated citing a non-existent case as the person's failure regardless of how it was produced. Verification is the part of the work the tool does not do for you, and it is what protects you.
How is this different from the four-step citation check?
The four-step check is a quick, general two-minute pass over any citation. This guide focuses on AI-generated citations specifically: the red flags that a citation was invented rather than retrieved, and a component-by-component dismantling of it. They are complementary, and this page links to the quick version.
Does a grounded legal tool remove the need to verify?
It removes most of it. A grounded tool searches a real judgment database and shows the source, so the case exists and the citation matches by design. Reading the judgment for the exact proposition and confirming it is still good law remain good practice on any citation, from any tool.