Skip to content

AI Fake Citations: Why They Happen and How to Check

AI fake citations look like ordinary Indian case law. Here is why a model invents them, the red flags, and a five-step check before you rely on one.

<AUTHOR NAME> · <REVIEWER NAME, Advocate> · Updated · 10 min read

On this page

AI fake citations are case names and reporter references that a language model writes because they look like law, even when no court decided that case. They survive a quick read because the shape is familiar: parties, a year, a volume, a page. The check is the same one a junior already owes the senior who will sign: find the judgment, match the cite, and read the passage.

This is for the associate turning an overnight research note into a list of dates, and for the advocate who has to decide whether that note can go into a petition. You get the reason these citations appear, the red flags, and a five-step routine you can run without trusting the model. The wider account of grounded research is [LINK: ai-legal-research-india].

Key takeaways

  • A citation can be well formed and still name no case. Search the parties before you argue the proposition.
  • Match the court, the year and the reporter cite to the judgment you actually opened.
  • Read the passage, then find the quoted words in that text. A real case can carry a sentence it does not contain.
  • On 27 February 2026 the Supreme Court said that deciding on non-existent judgments is misconduct, and that a legal consequence follows.
  • Scafffold can retrieve a judgment from Indian Kanoon. It does not certify a draft.

Why do AI fake citations look like real law?#

A language model continues a pattern. A case citation is one it has seen many times: two names, a year, a reporter, a volume, a page. Asked for authority, and not looking at a retrieved judgment, it can finish that pattern with names and numbers that were never published. The result is fluent, and fluency is the hazard.

It is producing the next likely words. It is not opening a report. A citation is short and regular, so it is easy to complete. That is why the failure looks like a proper SCC or AIR reference, and why asking the same model "are you sure?" is a weak test. It can repeat the cite and still not have opened anything.

Two limits sit beside that. A model tied to a database can still misread a real judgment: right case, wrong holding, or a paraphrase set in quotation marks. A miss in one database is a warning, not proof that no such case exists. Indian Kanoon is a large search collection of Indian judgments and statutes. It is not an official law report. How to search it is [LINK: indian-kanoon-search-tips]. What a general chatbot gets wrong on Indian law is [LINK: chatgpt-for-indian-lawyers].

What are the red flags in fake case citations?#

Read the citation before you read the proposition. Several failures are visible in the cite itself.

What you seeWhy it is a warning
A year that cannot be a decision of that court. The order of 27 February 2026 records a reporter-style cite carrying the year 1071.The Supreme Court of India did not exist in 1071. The cite was never a report of its judgment.
Party names that are generic labels rather than the names of litigants.On 17 February 2026, LiveLaw reported Justice B.V. Nagarathna referring to a citation, "Mercy v. Mankind," which did not exist.
A reporter, volume and page that do not belong to the judgment you opened.The case may be real and the cite wrong, or both may be invented. Either way the note is not ready.
A quotation you cannot find by searching the judgment.A real judgment can be dressed with words it does not contain.
A holding with no paragraph and no report to open.You cannot check a proposition you cannot find.
A second answer from the same chatbot, offered as confirmation.That is the same generator, not a source.

The year 1071 is one of four reporter-style cites recorded in Gummadi Usha Rani v. Sure Mallikarjuna Rao, Supreme Court of India, 27 February 2026, SLP (C) No. 7575/2026. Open the order for the list. Do not copy those strings into a note. A familiar reporter is a reason to check, not a reason to stop.

How do you check AI fake citations in five steps?#

Run the steps in order. Stop at the first one that fails. A citation that fails step 1 does not get a clever explanation in the petition.

StepWhat you doStop when
1. Find the caseSearch the party names. Open the judgment, not only a snippet or a headnote.No judgment of that court has those parties.
2. Match the citeCompare court, year, reporter, volume and page with what is printed on the judgment you opened.The cite belongs to a different judgment, or the year cannot be a decision of that court.
3. Read the passageFind the paragraph the note is using. Read the sentences around it, and the holding in the court's own words.The text does not say what the note says.
4. See later treatmentRead cases that cite this judgment, especially a later decision of the same court or of a higher court that the note ignores.A later judgment takes a different line on the point you need.
5. Find the quotationSearch the judgment for the quoted words.Those words are not in the text.

Search on Indian Kanoon and on the report you use in chamber. Open the full text. A headnote is only a lead. Indian Kanoon's list of later citing cases is not a holding that the case was approved or overruled. Open the later case that matters, or say you have not. A year such as 1071 fails step 2 on its face: the Supreme Court of India did not exist then. Quotation marks in a chatbot answer are not evidence the court used those words.

A clean note records the parties, the court, the year, the matched reporter cite, the paragraph, and one sentence on why the case is there. Prompts for the research, each with its own check, are [LINK: ai-prompts-for-lawyers].

What did the Supreme Court say on 27 February 2026?#

On 27 February 2026, Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe passed an order in Gummadi Usha Rani v. Sure Mallikarjuna Rao, SLP (C) No. 7575/2026, arising from an order of 21 January 2026 of the High Court of Andhra Pradesh in CRP No. 2487/2025.

The record of proceedings says the petitioners were defendants in an injunction suit. On 19 August 2025 the trial court dismissed their objections to an advocate commissioner's report while relying on cited decisions. They told the High Court those decisions were non-existent. The High Court treated the judgments as generated by artificial intelligence, recorded a caution, and dismissed the revision on merits. The Supreme Court said the concern was the process of decision, and directed the trial court not to proceed on the commissioner's report while the petition was pending.

In paragraph 7 the Court said:

a decision based on such non-existent and fake alleged judgments is not an error in the decision making. It would be a misconduct and legal consequence shall follow.

The order takes cognizance of a trial court deploying "AI generated non-existing, fake or synthetic alleged judgments." It issued notice returnable on 10 March 2026, and it issued notice to the Attorney General, the Solicitor General and the Bar Council of India. It appointed Mr. Shyam Divan, senior counsel, to assist the Court. The order does not itself fix the penalty. It says a legal consequence shall follow, and that the Court will examine the issue further.

Two later orders in the same petition, both on Indian Kanoon, are procedural. On 5 May 2026 the Court directed that the matter be listed on 26 May 2026. On 26 May 2026 it impleaded, in the words of the order, the Ministry of Information and Technology and the Ministry of Law and Justice, issued notice to them, and listed the petition on 1 September 2026. A search of Indian Kanoon on 8 October 2026 for this SLP number returned those three orders and no later one. [VERIFY: whether any order was passed on or after 1 September 2026 in SLP (C) No. 7575/2026, and whether it changes the declaration in the order of 27 February 2026.]

LiveLaw reported remarks on 17 February 2026 before Chief Justice Surya Kant, Justice Joymalya Bagchi and Justice B.V. Nagarathna. The report quotes the Chief Justice calling AI drafting by some lawyers "absolutely uncalled for," and it reports Justice Nagarathna referring to "Mercy v. Mankind" as a citation that did not exist, and to quotations pinned on real judgments. That is a news report of oral remarks, not an order. The written order is the one of 27 February 2026.

This article explains that record and a checking routine. It is not advice on any pending matter, including that petition.

What usually goes wrong in the check?#

The headnote stands in for the judgment. If the note's sentence appears only in the headnote, step 3 is unfinished. The holding is in the text.

The names are close, and the case is the wrong one. Match the other party, the court and the year before you adopt the holding.

The case is real and the use is not. The judgment exists, the reporter cite matches, and the note still describes a different rule. A database search will not catch this. Step 3 will.

The quote is a polished paraphrase. If you cannot find the words, take the quotation marks off. Quote the court, or drop the sentence.

The same model is asked to audit itself. A second chat that calls the cite correct has not opened a report unless it shows you the paragraph. Then open that paragraph yourself.

The cite is left in because the hearing is in the morning. Leave the case out. A short list of real cases is a better filing than a long list that includes one you have not opened.

Doing this with Scafffold#

Put the note or the draft in the project and ask which authorities it uses. For Indian judgments and statutes, Scafffold searches Indian Kanoon from the conversation and can open the full text there. It is instructed to name cases that search returns, and to give the Indian Kanoon link. That is not a guarantee. A retrieved judgment can be real and still fail the passage check or the quotation check.

On a document you uploaded, the answer cites the page and a short passage from that page. Open the page. If the quoted words are not on it, drop the point.

Scafffold does not scan a finished draft and certify the citations. You still run the five steps. You can choose Claude, Gemini or GPT for that conversation. The choice does not change the check. You can open Scafffold and keep the papers and the judgments in one matter.

Sources#

FAQ

What makes a citation an AI fake citation?
A model can produce a case name, a year and a reporter cite that no court decided. The parts look ordinary. The citation is fake when no judgment matches the parties and the reporter reference.
Does a proper SCC-style cite mean the case exists?
The shape is not proof. The Supreme Court order of 27 February 2026 records reporter-style cites, one of them carrying the year 1071, and treats those judgments as non-existent. Open the text before you rely on the cite.
What are the five checks?
Search the parties, match the court and the reporter cite, read the passage the note relies on, see how later cases have used it, and find the quoted words in the judgment. Stop at the first check that fails.
Does Scafffold certify citations in a draft?
Scafffold can search Indian Kanoon and open the judgment in the chat. It does not certify a draft. You still match the proposition, and any quotation, to the text of the judgment.
Which order should I read on fake judgments?
Read Gummadi Usha Rani v. Sure Mallikarjuna Rao, Supreme Court of India, 27 February 2026, SLP (C) No. 7575/2026, on Indian Kanoon. Later orders in that petition, through 26 May 2026, are procedural. Check for anything filed after that date.

This article is general information about Indian law and legal practice, current as of the "Updated" date above. It is not legal advice. Scafffold is a software company, not a law firm.