Eleven at night, reply due at half past ten in the morning. The junior asks her question and the AI gives her a Delhi High Court judgment.
She checks it, because she reads our blog now and has developed trust issues. Real case. Citation opens. Ratio says what the summary said it said. She files and goes to bed rather pleased with herself.
The matter is in Bombay. So the judgment binds nobody in the room. It's a suggestion. Nicely dressed, correctly formatted, entirely optional. And there's a Bombay judgment the other way that never came up, which the other side is now reading out with an enjoyment he isn't bothering to hide.
Nothing in the AI's answer was untrue. That's what makes it awkward. The problem wasn't a fake case. It was authority.
Who binds whom
Article 141 is the easy bit. The Supreme Court speaks, everybody follows.
After that it gets fussy. A High Court binds the courts under it, in its own territory, and nowhere else. Madras in Calcutta is a guest, not the landlord. A single judge who disagrees with an earlier single judge doesn't simply depart from it, she refers. Division Bench over single judge, Full Bench over Division Bench. Tribunals sit inside all of this at angles that never quite form a ladder. And state law runs beside central law, so the same facts in two states can land differently.
Nobody thinks any of this is hard. It's first-year material. Which is exactly the trouble, it's so obvious that nobody writes it down inside a judgment, so there's nothing in the corpus for a machine to learn it from.
Why the machine can't see it
Ask a general model for case law and what you're really getting is similarity. It finds the text nearest your question, which is useful, and does turn up things a Boolean search won't.
But relevant and binding are not the same property, and only one of them is in the document. No judgment carries a line saying which courts have to follow it. That depends on where you're appearing, and the model has no idea where you're appearing.
So it hands you Madras for a Karnataka matter with exactly the confidence it brings to a right answer. Generic AI can't "distinguish binding vs persuasive precedent," and misreads "legal context."¹ In India those are the same complaint.
It isn't lying to you. It's laying everything out at one weight, Supreme Court and tribunal and long-dead authority all in the same font, which is far harder to spot than an invented case because every piece of it checks out.
Three ways it bites
Wrong forum, as above.
Then the split you weren't told about. High Courts disagree all the time. You get whichever side reads closest to your query and never hear the other exists, so you don't lose the argument so much as walk in unaware there was one.
And the dead one. Binding in principle, distinguished into nothing in practice, or overtaken by an amendment younger than the judgment. It opens. It reads well. It's a corpse with your name on it. "Does this case exist" catches none of these. All three exist.
Which is why it has to come first
We keep saying the destination is adversarial AI, something that attacks your draft before the other side does. That only works if it can rank what it finds. An adverse judgment on its own isn't news; what you want is this one binds your bench, this one you can bury in a footnote, this one's before a larger bench and nobody's leading with it. Take the hierarchy out and you're left with a long list of things to worry about and no order to worry in.
Getting that right isn't a feature you add later. It means tagging every judgment with its court, bench, date, territory and treatment, keeping amendment histories and state variations, and filtering on forums before you rank on relevance. Corpus-level work. No prompt fixes it.
The bit that's still yours
None of this decides your case. Whether the binding judgment actually fits your facts is judgment. Whether a Kerala ruling, persuasive only, is written well enough that this bench takes it anyway, also judgment. Persuasive authority isn't junk — used properly it's often the most interesting thing in a submission.
We were never trying to make the machine choose. Only to stop it printing a suggestion and an order in the same typeface.