Bombay set aside a customs order with phantom precedents. Verify authority before consequence

THE SIGNAL IN ONE SENTENCE

A legal citation is a small piece of text with an unusually large lever attached. Put the wrong one in a book review and you have an embarrassing correction. Put it in a customs order, a benefits decision, a disciplinary finding or a bail application and it can move money, property, status or liberty. That is why a four-page order from the Bombay High Court deserves attention far beyond the lawyers who read it first. Quanta Logistics challenged an Order-in-Original dated August 11, 2026 from the Commissioner of Customs (Import-I). The company told the High Court that the customs order had relied on two supposed Supreme Court judgments that, to the best of its knowledge, did not exist in any law report or database. It said the citations appeared to be an AI hallucination and that the order had been made without an independent application of mind. Those are the petitioner's allegations. The distinction matters. The High Court did not conduct a merits trial and declare that a named official deliberately fabricated cases. Counsel for the respondents told the court that the challenged order could be quashed and the matter returned for a fresh decision. By consent of both parties, Justices M. S. Karnik and Sandesh D. Patil set aside the August 11 order on October 30. The court directed Quanta Logistics to appear before the customs authority on November 5 and required the fresh proceedings to conclude within eight weeks from that date. The order was uploaded on November 5. Local legal reporting brought it to wider public attention on November 7. That procedural shape is important. It keeps the article honest and sharpens the actual signal. The system did not need a philosophical seminar about whether a machine can practice law. It needed reliable authority for the authorities it cited. According to the order, Quanta Logistics had asked the customs authority on September 1 for copies of the two judgments and did not receive them. That is a wonderfully plain test. If a consequential decision invokes a precedent, the decision maker should be able to produce the precedent. AI makes this old obligation more urgent, not more exotic. Language models can produce case names, courts, dates and holdings in the polished shape of a citation. That shape is persuasive because it resembles the compressed signals lawyers use every day. An invented citation can look tidier than a real one. It has no coffee stains, no split authority, no awkward procedural posture and no paragraph that ruins the neat summary. The danger is not merely that a model can make something up. Humans have always made citation errors. The danger is speed plus authority. A system can generate a plausible chain of support faster than a reviewer can feel suspicious, then place it inside an official document whose formatting tells the reader that the checking has already happened. So the verification cannot be an informal final glance. Every authority used to justify a consequence should carry a small, reviewable receipt. The receipt begins with the citation exactly as it appeared in the draft. It then records the normalized case name, court, case or neutral citation number, decision date, source document, pinpoint page or paragraph and retrieval date. Next comes the proposition. What sentence is this case supposed to support? Copy the relevant passage with enough surrounding context to preserve its meaning. Record whether the passage states a binding rule, applies a rule to different facts, quotes another court or merely summarizes an argument. Then record the reviewer. A name or accountable role should confirm that the cited document was opened and matched to the proposition. If the search failed, record that too. An empty result should not be silently replaced by the model's confidence. Finally, connect the citation to consequence. Does it support a background sentence, or is it part of the reasoning that changes a person's duty, payment, penalty, access or legal position? The greater the consequence, the stronger the source and review gate should be. This is not a request to ban useful tools. Search systems, citation parsers and language models can help find possible authorities, compare versions, translate dense passages and flag missing fields. They can turn an afternoon of clerical work into a useful first pass. But discovery is not verification. A candidate becomes authority only after the source document is found, opened, matched and placed in context. The source hierarchy matters. A court copy or official repository should sit above a model answer, a search snippet, a headnote, a blog post or a database summary. Secondary sources can point a reviewer in the right direction. They should not be promoted into primary law by repetition. The same discipline belongs outside legal practice. A clinician using a study to support a treatment note should preserve the paper, population, endpoint and limitation. A bank using a policy citation to deny access should preserve the operative rule and version. A school using a regulation to discipline a student should preserve the actual text in force on the relevant date. A public agency using a technical standard should preserve the edition, clause and interpretation. In each case, the authority check asks the same questions: does the source exist, is it authentic, does it say what the draft claims, was it in force, does it govern this situation and who verified it before the consequence occurred? Teams can build this into ordinary workflow. At drafting time, force every citation into a structured field instead of accepting free-floating prose. Resolve the case or document identifier against an approved source list. Refuse to treat a title alone as proof. Require a stable source link or archived document and a pinpoint reference. At review time, separate citation existence from proposition support. A real case can still be cited for a rule it never announced. A valid statute can be obsolete. A genuine paper can be irrelevant to the population in front of you. Verification needs two checks, not one green badge. At release time, block consequential output when any authority remains unresolved. The system should show which claim is waiting, which sources were searched, which reviewer owns the check and what safe fallback is available. Sometimes the correct fallback is to remove the claim. Sometimes it is to ask for more time. Sometimes it is to decide on another valid ground. After release, keep the receipt with the decision. If a citation is challenged, a reviewer should be able to reconstruct the source, version, passage and approval without searching through private chats or trying to remember which tool produced a sentence. Automation can help police automation. A citation checker can test format, duplicate identifiers, dead links, missing pinpoints and conflicts between case name and number. A second retrieval system can compare the quoted passage with the source. But the final gate should remain accountable to a human who understands the consequence and can explain the reasoning. That human review should not be ceremonial. Do not ask someone to approve 70 citations in one undifferentiated scroll. Rank them by consequence and uncertainty. Put unresolved or model-suggested authorities first. Show the exact proposition beside the source passage. Make rejection easy. Record the reason. If the reviewer cannot access the source, the interface should say unverified, not probably fine. There is a useful cultural rule hiding here too: do not punish the person who stops the line. Citation failures survive when speed, hierarchy or embarrassment makes it easier to wave a suspicious reference through than to hold a decision. A clerk, analyst, junior lawyer or case worker who cannot find an authority should be able to pause publication without needing to prove that a hallucination occurred. The burden belongs on the citation to become verifiable. The Bombay matter also shows why labels must remain precise. It is tempting to summarize the episode as a court finding that AI invented two judgments. The order itself says the petitioner made that case, the respondents consented to quashing and the court remanded the matter. Those are strong facts. Inflating them would repeat the exact problem the story warns about: letting a convenient sentence outrun its authority. The better habit is almost boring. Open the source. Match the identifier. Read the passage. Record the version. Name the reviewer. Stop if the support is not there. AI can help find a door. Before a decision walks through it, someone still has to check that the door exists.