Supreme Court on AI Fake Citations (2026): What Every Indian Advocate Must Change

On 2 July 2026, in Pooja Ramesh Singh v. J&K Bank, the Supreme Court set aside orders that rested on invented judgments. The message was not “ban AI”. It was: citing unverified machine output is on the advocate, not the chatbot.

If you are still pasting ChatGPT case lists into a compilation, this is the week to stop.

What actually changed

• Invented authorities in a filing are now a documented Supreme Court problem, not a Twitter scare.
• The Court’s own draft rules still allow research, citation-checking, drafting help, summaries and translation — with human oversight.
• Disclosure is becoming the professional default, not a confession.

The Citation Truth Test™ — 7 checks, 90 seconds

• Exists — parties match on Indian Kanoon.
• Court and year match the document you opened.
• Citation / neutral number match.
• The quoted sentence is actually in the judgment.
• Paragraph number is inside the judgment’s length. (¶73 of a 27-paragraph judgment = fake.)
• Still good law — open “cited by”. Look for overruled / reversed / per incuriam.
• The section it applies is still in force — BNS / BNSS / BSA, not repealed IPC numbers.

All seven must pass. One fail = it does not go in the draft. Average advocates score 4/10 on a planted gauntlet. That is the skill gap AI Vakeel is built to close.

Chamber policy you can paste this week

Three lines on a card next to the intern desk: (1) no identified client data in consumer AI, (2) no citation leaves the draft until someone opened it, (3) if it is filed, it is yours. Add a one-line disclosure in the covering note when AI assisted the first cut.

AI Vakeel Day 0 and Day 2 are exactly this: red lines, Truth Test, Fake-Citation Gauntlet, plus a compliance kit. One afternoon. Then it is a habit.