Can You Trust an AI-Generated Case Citation? A Verification Framework for Indian Lawyers

The Supreme Court's decision in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668) - see our companion article - shows what happens when an AI-hallucinated citation reaches an adjudicatory order unverified: the order itself gets set aside, not just the citation. That outcome makes an ad hoc 'it looked right to me' approach to citation-checking genuinely risky. What follows is a concrete, ten-point framework for the verification step - not a substitute for it.

1. Does the case actually exist? Search for it directly on a primary source - don't accept the AI tool's own description as proof.

2. Is the court correct? An AI tool can attribute a real holding to the wrong court or the wrong bench level.

3. Is the case title correct? Party names get transposed or invented more often than the underlying holding does.

4. Is the citation identifier correct? A neutral citation, SCC/AIR reference, or case number that doesn't resolve to the claimed case is a hard stop.

5. Is the date correct? A plausible-sounding date attached to a real case is still a fabrication if the case wasn't actually decided then.

6. Does the cited paragraph exist? Open the actual judgment and locate the specific paragraph being relied upon - don't assume it's there because the AI tool quoted it.

7. Does that paragraph actually support the proposition? A real paragraph, quoted accurately, can still be cited for a proposition it doesn't support.

8. Is the judgment/order distinction correct? A tribunal order, an interim order, and a final judgment carry different weight - conflating them is a separate error from the citation being fake outright.

9. Is the case still good law? A real, accurately quoted holding that has since been overruled or distinguished is a different kind of failure than fabrication, but still a failure.

10. Is a primary source actually accessible for it? If nothing but a secondary summary can be found, that itself is a signal to slow down before relying on the citation.

Running through these ten questions does not make verification automatic or effortless - it's a discipline, not a tool, and it still requires someone to actually open a primary source and read it. It also won't catch every failure mode on its own: a subtle mischaracterization of a real holding can pass all ten checks if the reviewer doesn't read closely enough. The framework structures the review; it doesn't replace the judgment of the person doing it.

It's also worth being honest that this list is not exhaustive - it reflects the failure patterns documented in reported cases so far (see our piece on the real Indian cases where AI-hallucinated citations reached courts and tribunals), not a guarantee against failure modes that haven't surfaced yet.

This doesn't need to be a formal, written sign-off process to be useful - for most firms, it's enough that someone specific is responsible for running through it before a citation reaches a filing, and that this responsibility is assigned before a tool is adopted rather than after a citation gets challenged. See our companion piece on responsible AI use in legal research and drafting for the broader firm-level practice this framework fits into.