AI Citations - Anatomy of a Grounded Answer
27 July 2026
Everyone in legal tech now sells "grounded" AI. Grounded. Verified. Source-backed. It's reassuring the way "artisanal" is reassuring on a jar of jam, you nod, you feel better, you have no idea what happened in the factory. So let's open the factory. One question, run through two very different machines.
Here's the question, the kind a junior types at 11 p.m.: "What's the limitation period to file under Section 34 of the Arbitration Act, and can a court condone delay beyond it?"
The confident stranger. A generic model doesn't look anything up, it reads that question twice. It predicts the most statistically plausible next word, forever. It's the most well-read person at the party who has never once said "I'm not sure." Ask it the easy version and it nails it: three months plus thirty days, cites Union of India v. Popular Construction Co. — a real case that really says that. Which is exactly why it's dangerous; it's right often enough that you stop checking. Now ask a niche follow-up and watch it wander: "See Meridian Infrastructure Pvt. Ltd. v. State of Maharashtra, (2021) 6 SCC 214, where the Bombay High Court held…" Beautiful. Plausible. Correctly formatted. But complete fiction. The case name is invented, the docket number is invented, the ratio is invented. All dressed in the exact typography of something true. A January 2025 NLSIU study found frontier models "frequently generated hallucinations" on specialised legal research.¹ That's in calm conditions, not a junior copying the good-looking paragraph into a reply due at 10:30 a.m.
The boring, trustworthy librarian. A grounded system flips the order: it retrieves first, reasons second. Before writing a word, it pulls the actual documents, the text of Section 34, the judgments interpreting it, the amendment history, and then reasons only over those specific documents, not the whole foggy ocean of the internet. The generic model is a brilliant friend recounting a case from memory over dinner. The grounded one is a librarian who won't answer until she's set the actual open volume on the table. The first is faster and more charming. The second is the one you cite in court. So every load-bearing claim comes with a little door: click the provision, land on Section 34(3); click the case, land on the real judgment. Every output needs to be traceable to a primary source², an architectural promise with a cost: the system can only tell you what it can show you.
But "traceable" is a verb, not a sticker. Traceability isn't something the software has so you can relax. The machine puts the door there; you open it. A citation you never clicked isn't verified, it's just a citation with better manners. And here's the part nobody puts on the billboard: grounding dramatically reduces hallucination, it doesn't abolish it, and it certainly doesn't abolish you. A grounded system can still surface a real case that's been distinguished into irrelevance, or a provision since amended. Retrieval fixes "does this case exist." It does not fix "should I rely on this case." That second question is law, and law is still your job. What grounding kills is the worst failure; the pure fabrication, the Meridian Infrastructure that never existed. It turns "trust me" into "here, check," and changes your job from detective to editor. Smaller. Saner. Not no job.
So, two answers on your screen and a hearing in the morning: look for the door. Not the confidence, not the word "verified" those are free, the fabricators hand them out too. Look for whether you can click the citation and land on the actual judgment. If you can, open it, read the ratio, then decide. If you can't, you're quoting a very well-read stranger to a judge. That check takes four seconds. The alternative has taken lawyers a lot longer, usually in a sentence that begins "counsel, this case does not appear to exist."