The Case Law Did Not Exist. The Tax Officer Did Not Check.

⏱ 6 min read|Published 5 October 2026|By Dibyendu Choudhury

On 20 August 2026 the Gujarat High Court quashed a GST order passed against a firm called Faiz Enterprise, and the reason is the kind of sentence that would have sounded like satire two years ago. The State Tax Officer who passed the order had relied on case law that did not exist. The judgments cited were generated by artificial intelligence, and the officer had not verified them. According to MediaNama's report of the matter, the officer admitted this and apologised. The court ordered a fresh notice and directed compliance with the department's own guidelines on the use of AI, issued on 18 August 2026.

Read that last date again. The guidance existed two days before the court had to enforce it. That single detail changes what kind of story this is.

What actually went wrong

Large language models are very good at producing text that looks authoritative. Ask one for a case that supports a position and it can produce a title, a court, a year and a tidy paragraph of reasoning that has the exact shape of a real citation. Sometimes the case is real. Sometimes it is a confident fabrication. The model does not know the difference in the way a lawyer or a tax officer must. Verifying is the human's job, and verifying is dull: you look the citation up in a trusted database, you read the paragraph, you check that the court actually said what the draft claims.

In this matter, by the reported account, that dull step was skipped. An order that affected a small business's tax position rested, in part, on authorities that were never real. Nobody intended to deceive. That is almost the point. The failure was not malice or even incompetence in the usual sense. It was an unchecked shortcut inside a process that carries legal consequences for the person on the receiving end.

Why the taxpayer's side of the table matters

Most commentary on stories like this focuses on the officer, or on the technology. I want to look at the other end. A small firm received a demand built on fabricated law. To undo it, the firm had to take the matter to a High Court. For a large corporate, that is a line item. For a proprietorship or a ten-person manufacturer, High Court litigation means a lawyer's fees, months of hearings, and the quiet anxiety of an unresolved liability sitting on the books. The cost of the officer's shortcut was borne almost entirely by someone else.

That asymmetry is the heart of why the quality of an order matters so much in a tax system. A speaking order, one that gives reasons and cites authority, is not a courtesy. It is the thing that lets the taxpayer understand what is being alleged and respond to it. If the reasons are hollow, the taxpayer is arguing with a ghost. Here, the ghost was literally a set of invented judgments.

It is a governance story, not an anti-AI story

I would be doing my readers a disservice if I used this case to tell small business owners to fear AI. I use these tools myself, and the great majority of MSMEs I speak to stand to gain from them in drafting, translation, bookkeeping support and customer service. The lesson is more precise: a tool that can be wrong must be used inside a process that checks it, especially when the output can change someone's legal position.

The department appears to have understood this, because it issued guidelines on AI use on 18 August. The court's direction to comply with those guidelines tells us the rule was already written. What was missing was the discipline to follow it in the moment, and perhaps the supervision to notice when it was not followed. That distinction, between a missing rule and a missing habit, is where most real-world failures live. Writing a circular is easy. Making a thousand officers verify a citation at four in the afternoon on a Friday is the hard part.

What an MSME owner can do

I am not a lawyer, and nothing here is legal advice. But as someone who works around small businesses, I can offer a practical checklist of habits that cost little and protect a lot.

First, read every order or notice you receive as a document to be checked, not merely obeyed. If an order cites a judgment, note the court, the year and the case title. Second, ask your chartered accountant or advocate to confirm that the cited authority exists and says what the order claims. A citation that cannot be found in a standard legal database deserves a pointed question, and it can be raised in your reply. Third, mind the timelines. Under the GST law an appeal against an order generally has to be filed within a limited period from the date it is communicated, commonly three months, with narrow room for condonation, so a suspicious order should reach a professional quickly rather than sit in a drawer. Fourth, keep a clean paper trail of your own returns, invoices and replies, because a fresh notice, which is what the court ordered here, will arrive and must be answered properly.

And the same discipline applies to you

It is tempting to read this story and feel comfortable that the error was on the other side of the table. Do not. The same temptation sits in every small office that now uses an AI assistant to draft a reply to a notice, a contract clause, a customer policy or a compliance note. If you paste a generated paragraph into a reply to a tax authority, and it includes a case that does not exist, the embarrassment is yours. The standard of care is identical for the person who sends the document and the person who receives it: nothing generated goes out unverified.

A simple rule works well. Any fact, figure, section number or citation that comes out of an AI tool must be traced to a primary source before it leaves your hands. If you cannot trace it, delete it. It costs ten minutes and removes an entire category of risk.

The enforcement gap

For policy makers, the moral is institutional. Guidelines are necessary but not sufficient. If a department issues an AI-use policy, the sensible next steps are to build verification into the workflow itself, for example by requiring that any authority cited in an order be accompanied by a reference to where it can be checked, to train officers on what hallucination looks like, and to make the reviewing authority responsible for spot checks. The High Court has effectively said the same thing in a different register: follow your own rules.

There is a hopeful reading of this episode. The matter surfaced, the officer owned the mistake, the court corrected it, and the department's guidance was already on the books. That is a system learning in public. The risk is that it learns slowly while the number of orders drafted with AI help grows quickly.

The sentence worth remembering

The case law did not exist, and the tax officer did not check. Everything else follows from the second half of that sentence. Technology will keep producing plausible falsehoods for as long as it is built the way it is today. The countermeasure is not fear and it is not a ban. It is the old, unglamorous virtue of checking the source before you act on it, practised consistently by officials, advisers and business owners alike.

Published 5 October 2026 · dibyenduchoudhury.com