The failure is structural, not careless. Understanding why it happens tells you exactly which check catches it.
A general AI model produces text that resembles a citation rather than looking one up, so it can return a case name, a year and a citation that do not exist. The fix is structural, not stylistic: an answer must be tied to a real judgment on file, pin-cited to the paragraph, with its good-law status shown.
Three ordinary facts, none of them mysterious, combine to produce a citation that reads perfectly and refers to nothing.
A general AI assistant is asked a question and produces the text that most plausibly follows it. A case citation has a very regular shape — two party names, a year, a reporter, a volume, a page. Producing something of that shape is easy. Producing something of that shape that also corresponds to a judgment actually delivered by an Indian court is a completely different operation, and a general assistant is not performing it. When it has no real judgment to draw on, it does not fall silent. It produces the shape.
General-purpose assistants have read far more American and English case law than Indian case law, because far more of it circulates in the material they were built from. The consequence is not that Indian answers are worse in tone. It is that the assistant has less real Indian material to draw on and therefore more room to compose. Ask about a well-known American doctrine and you may get something serviceable. Ask about a High Court's line on a specific proviso and there may be nothing there to draw on, so it will compose.
An advocate spots a shaky argument by tone. That instinct does not work here. Legal prose is written in a settled register — declarative, unhedged, citation-carrying — and a fabricated passage is written in exactly the same register as a correct one. There is no wobble to detect. This is the part that catches experienced practitioners, because every other unreliable source they have ever encountered gave itself away in the writing.
None of these is proof. Each is a reason to open the judgment before the brief goes any further.
Party names that sound entirely ordinary attached to a year that does not fit — a judgment on a provision inserted in 2013 dated 2006, or a Supreme Court decision on a court constituted later. The names pass. The arithmetic does not.
A volume and page in a series that did not run that volume in that year, or a neutral citation in a format the court does not use. Reporter conventions are rigid, which makes an invented citation testable against the format alone.
A holding that answers your question exactly, with no qualification, no proviso and no remand. Real judgments are decided on their facts and carry conditions. A proposition that fits your matter perfectly is the one to check first, not last.
The most common misattribution, and the most consequential, because it converts persuasive authority into binding authority in the reader's mind. It also survives a casual check, because the judgment exists — only the court is wrong.
The citation resolves, the judgment opens, and the paragraph number cited does not contain the proposition. The judgment was real; the pin cite was composed. Verifying the case name alone does not catch this.
A proposition stated under a provision that no longer governs, because the criminal law numbering changed on 1 July 2024. Correct authority, wrong code. The IPC to BNS mapping.
Filings containing citations to judgments that do not exist have reached courts in several jurisdictions since 2023, and the professional consequences where it has happened have been serious. The pattern is consistent wherever it has occurred: a general assistant was used to prepare or supplement a brief, the citations were not opened before filing, and the fabrication was discovered by the court or by opposing counsel rather than by the advocate who filed it. The advocate is answerable in every one of those accounts, because the signature on the filing is the advocate's.
In India, judicial and bar-level guidance on the use of AI in legal work is developing rather than settled. Courts have been raising the concern, and practitioners should expect the position — on disclosure, on verification obligations and on consequences — to become firmer over time.
This page deliberately does not name individual advocates, matters or benches. Two reasons. Naming people adds nothing to the practical lesson, which is entirely about verification. And an account written today would be out of date within months, which is why this page carries a review date rather than a claim to be current. It was last reviewed in July 2026.
What is worth taking from the record is narrow and durable: the failure has never been discovered by reading the brief. It has always been discovered by opening the judgment. That is the entire lesson, and it is the reason the checklist below is written the way it is.
Run in this sequence, the first check eliminates the fabricated citations and the rest are quick. Run out of order, all six are slow.
Locate the full text of the judgment itself. If the citation cannot be resolved to a judgment on record, stop — there is nothing to verify and nothing to file. This single check disposes of every wholly fabricated citation, which is why it goes first.
Check that the judgment was delivered by the court named, and note the bench strength. A High Court judgment presented as a Supreme Court judgment changes what the authority does in your favour, and a Division Bench judgment presented as a Full Bench does the same.
Read the paragraph that is said to support the proposition. Confirm that it says what is claimed and that the conditions attached to the holding have travelled with it. A real judgment with an invented paragraph reference is the failure mode that survives a name check.
Establish whether the judgment has been overruled, distinguished, doubted or held per incuriam on the point relied on. An authority that is real, correctly attributed and no longer good law is a worse position to be in than a citation that never existed. How the check is done.
Confirm that every statutory provision cited belongs to the code governing your matter. Since 1 July 2024 the same section number can mean two different things depending on whether the CrPC or the BNSS applies, and section 482 is the sharpest example. The three numbering collisions.
Record the paragraph or page against each proposition as you verify it. The value is on the day of the hearing: a bench that questions a proposition gets the paragraph immediately, and nobody repeats the verification at midnight because the note said only "SC, 2019".
Lawgger answers only from judgments it holds. The collection is 2.07 crore-plus Indian judgments in full text, covering the Supreme Court and all 25 High Courts. A proposition that cannot be attached to one of those judgments has nowhere to go, so it is not produced. That is the structural difference, and it is the only kind of difference that matters here.
Three consequences follow, and they are worth stating separately because they are separate protections:
The honest limit. This does not remove the advocate from the loop, and Lawgger does not claim it does. A pin cite makes verification take seconds instead of an evening; it does not perform the verification. The advocate who signs the filing is the person who confirms that the paragraph says what the brief says it says. Any tool that tells you otherwise is selling you the exact risk this page is about.
Related reading: how research answers are built and cited, how accuracy and data handling work, and ChatGPT for lawyers in India — where a general assistant helps and where it must not be used.
A general chat assistant can return a case name, a year and a citation that do not correspond to any judgment. It produces text that resembles a citation rather than looking one up in a collection of judgments. The output is fluent and correctly formatted, which is precisely why the error is hard to notice. Where a general assistant helps and where it must not be used.
Three reasons combine. A general model completes patterns rather than performing a lookup. Indian judgments are thinly represented in general training material compared with American and English case law, leaving more room to compose. And the register of legal writing is confident by convention, so a fabricated citation reads exactly like a correct one.
It depends entirely on whether the answer is tied to judgments on file. A tool that composes an answer from memory is not reliable for citations. A tool that builds an answer only from judgments it holds, pin-cites each proposition to the page, and shows the good-law status of each authority can be relied on as a research aid — with the advocate verifying before filing.
Open the judgment itself rather than a summary. Confirm the court and the bench. Find the paragraph the proposition is drawn from and read it. Check whether the judgment is still good law on that point. Check the section numbering against the code that governs your matter. Record the pin cite so the check need not be repeated.
Filings containing citations to non-existent judgments have reached courts in several jurisdictions since 2023, with serious professional consequences. In India, judicial and bar-level guidance on the use of AI in legal work is developing rather than settled. This page does not name individual matters or advocates, and is reviewed as the position changes — last reviewed July 2026.
Lawgger builds answers only from judgments on file across the Supreme Court and all 25 High Courts, and pin-cites every proposition to the page of the judgment it comes from. A citation it returns resolves to a judgment you can open. It remains a research aid: the advocate signing the filing verifies the authority before it is used.
The advocate who signs and files a document is answerable for its contents, whoever or whatever drafted it. Courts have treated citations to non-existent judgments seriously. This page does not offer legal advice on professional consequences. The practical position is that verification is not delegable, and a pin cite is what makes verification quick enough to actually do.
No tool should be described as incapable of error, and Lawgger does not describe itself that way. What can be changed is the structure. If every proposition must be attached to a judgment on file and pin-cited to a page, a fabricated citation has nowhere to attach itself. That converts an invisible failure into a verifiable one, which is a different thing from eliminating error.
That every authority resolves to a real judgment, that the court and bench are correctly stated, that the paragraph relied on says what the brief claims, that the authority is still good law on that point, and that statutory provisions match the code governing the matter. Authority Check in Lawgger returns a Table of Authorities with a verdict against each case.
Answers built only from 2.07 crore-plus Indian judgments on file, pin-cited to the page, each authority carrying its good-law verdict. A 14-day free trial of the full product, no card required.