TL;DR: AI can do real work on a writ petition - a first cut of the synopsis, a structured set of grounds, a summary of a bulky annexure, a tightening pass on prose. What it cannot be trusted to do is cite case law, apply jurisdiction-specific High Court rules, or make the final legal call. The danger that has put lawyers in front of disgruntled judges, in the United States and now in India, is the hallucinated citation: a confident, well-formatted reference to a judgment that does not exist. In February 2026 the Supreme Court of India said reliance on AI-generated fake judgments can amount to misconduct, not merely an error. This post sets out the anatomy of a writ petition, where AI genuinely helps, where it must never be trusted, and a verification discipline that lets you keep the speed without filing fiction.
On this page
- Why this matters more for a writ than almost anything else
- The anatomy of a writ petition under Article 226 and 32
- Where AI genuinely helps in drafting a writ
- Where AI must never be trusted
- The hallucinated-citation danger: real cases, real consequences
- What the Indian courts are now saying
- A safe human-in-the-loop drafting workflow
- The verification discipline: every cite, against the source
- Prompting tips for petition drafting
- Confidentiality and privilege when using foreign LLMs
- Court-rule compliance: the part AI keeps getting wrong
- Frequently asked questions
- Drafting writs with Indian-law grounding
Why this matters more for a writ than almost anything else
A writ petition is not an ordinary plaint. When you move a High Court under Article 226 or the Supreme Court under Article 32, you are invoking an extraordinary jurisdiction. The court is being asked to step in where an ordinary remedy is inadequate, often urgently, often against the State. The bar for getting in is higher, the scrutiny is sharper, and the consequences of a sloppy filing land faster.
That is exactly why AI is so tempting here, and so dangerous. A writ rewards clean structure, a crisp statement of facts, and grounds that map fact to legal principle without padding. Those are things a good model can help you assemble quickly. But a writ also lives and dies on its authorities. A ground that rests on a misread judgment is worse than no ground at all, and a ground that rests on a judgment that does not exist is a professional emergency.
The argument of this post is not that you should avoid AI. It is that you should use it the way you would use a sharp but unreliable junior: take its first drafts seriously, take its citations on faith never, and own every line that goes out under your signature. Used that way, AI compresses the mechanical work so you spend your hours on the analysis that actually wins.
The anatomy of a writ petition under Article 226 and 32
Before deciding where AI fits, you need a clear map of what a writ petition actually contains. The exact sequence varies by High Court rules, but the standard components are stable across jurisdictions. A typical Article 226 petition is built from the following parts.
| Component | What it does | Drafting weight |
|---|---|---|
| Cause title | Names the court, the parties, and the article invoked | Mechanical but must be exact |
| Memo of parties | Full name, parentage, address, and email of each party | Detail-heavy, error-prone |
| Synopsis | A concise narrative of the dispute and what is sought | Persuasive, sets the tone |
| List of dates | Chronological table of relevant events with dates | Factual, must be accurate |
| Facts | The full factual account leading to the cause of action | The factual spine |
| Grounds | Self-contained legal arguments tying fact to principle | The legal core |
| Prayer | The precise relief sought from the court | Decisive, must be specific |
| Affidavit | Sworn verification that the contents are true | Procedural, non-negotiable |
| Annexures | Supporting documents, orders, correspondence | Bulky, needs indexing |
A few of these deserve a closer look, because they are where AI either helps the most or hurts the most.
The synopsis and list of dates
Most High Court rules require a brief synopsis and a list of dates and events to be annexed to the petition. Under the Delhi High Court rules, for instance, a concise statement of relevant facts in chronological order is expected with every petition, and documents are to be typed in double space. The synopsis is the first thing a judge reads, so it carries persuasive weight out of all proportion to its length. The list of dates is pure fact: a clean chronology that lets the bench grasp the matter in thirty seconds.
The facts and the grounds
The facts set out what happened. The grounds are where the petition earns its keep. As one practical drafting guide puts it, each ground is a self-contained argument connecting a specific fact to a legal principle, explaining precisely why the respondent’s action is illegal, arbitrary, or unconstitutional. A writ under Article 226 challenging an administrative order will usually also have to assert that no equally efficacious alternative remedy is available, since that is a recognised hurdle to the High Court’s writ jurisdiction.
The prayer and the affidavit
The prayer tells the court exactly what you want: quash this order, issue a writ of mandamus directing that act, stay this notice pending hearing. Vagueness here is fatal. The affidavit is the sworn statement, made before a notary or oath commissioner, confirming that the contents of the petition are true to the petitioner’s knowledge. It is the document that puts your client’s credibility, and yours, on the line.
Where AI genuinely helps in drafting a writ
Now place AI against that map. There is a clear set of tasks where a capable model adds real speed without touching the parts that must be human.
| Task | Why AI helps | What you still own |
|---|---|---|
| First draft of the synopsis | Turns your facts into clean narrative prose fast | The framing and emphasis |
| Structuring the grounds | Organises arguments into discrete, numbered grounds | The legal substance of each |
| Drafting the list of dates | Converts a messy file into a clean chronology | Verifying every date |
| Summarising annexures | Compresses a 60-page order into the operative points | Confirming the summary is faithful |
| Tightening prose | Cuts repetition, fixes flabby sentences | The final voice |
| Spotting gaps | Flags a missing jurisdiction averment or prayer | Deciding what to add |
The pattern is consistent: AI is strong wherever the task is to transform material you already have into a cleaner form. Give it your facts and it will give you a serviceable synopsis. Give it a long impugned order and it will give you a usable summary of what the order actually held. Give it your scattered grounds and it will arrange them into a numbered structure that reads logically.
This is the same strength that makes AI useful in contract drafting: the mechanical, transform-this-into-that work that consumes a junior’s evening but adds little legal value per hour. Offload that, and you free up the hours for the work that does add value.
One more genuinely useful role: structural critique. Ask a model to read your draft and tell you whether the jurisdiction is properly pleaded, whether the prayer matches the grounds, whether the alternative-remedy point has been addressed. It will not always be right, but it surfaces gaps you can then check yourself. That is a safe use, because you verify everything it flags.
A worked example: the synopsis
Take a concrete case. You act for a contractor whose registration has been cancelled by a State department without a hearing, and you want to move the High Court under Article 226. You have the cancellation order, the dates, and the correspondence. You feed the model these facts and ask only for a synopsis.
A weak first draft tends to ramble: it repeats facts, buries the central grievance, and reaches the relief sought only at the end. A good model, given a clean instruction, will produce something tighter: a two-paragraph opening that names the parties, states the impugned action, identifies the breach of natural justice, and arrives at the relief without detours. You then take that draft and do the part the model cannot: decide which fact to lead with, sharpen the characterisation of the department’s conduct, and make sure the framing sets up the grounds that follow.
That division of labour is the whole skill. The model gives you a clean canvas in two minutes instead of twenty. You supply the judgment that turns a competent synopsis into a persuasive one. Nothing in that exchange touches a citation, a rule, or a legal conclusion, so nothing in it can put you in front of a displeased bench. This is the safest, highest-value use of AI in the entire workflow, and it is worth getting good at.
What AI does badly even within its safe zone
It helps to be honest about the limits even where AI is broadly useful. A model summarising a long order will sometimes flatten an important nuance, treating an obiter remark as if it were the operative finding, or missing a conditional in the order that changes its effect. A model structuring grounds may merge two distinct legal arguments into one because they look similar on the surface, when keeping them separate is what wins. And a model drafting a list of dates will faithfully reproduce a wrong date if you fed it a wrong date, because it does not know your file.
None of these are reasons to avoid AI. They are reasons to read its output as a draft by a fast but green junior. The summary is a starting point you check against the order. The grounds are a structure you refine. The chronology is a table you verify date by date. Treated that way, the limits are manageable. Treated as finished work, they become the errors that surface in court.
Where AI must never be trusted
The other half of the map is the part you protect. There is a category of work where AI output must be treated as a hypothesis to be verified, never as an answer to be relied on.
Case citations. This is the single most dangerous output. A language model generates text that looks like a citation by pattern, not by retrieval from a verified database. It will produce a case name, a citation, a year, a holding, and a quoted passage, all formatted perfectly, all potentially invented. We will come back to this, because it is where lawyers are getting sanctioned.
Jurisdiction-specific High Court rules. Each High Court has its own rules on format, court fees, the number of copies, affidavit requirements, and what must be annexed. A model trained on a global corpus has no reliable grasp of the Delhi High Court (Original Side) Rules, the Bombay rules, or the local practice of the registry you are filing in. It will guess, and it will sound confident.
Statutory provisions and their current text. A model may cite a repealed provision, an old section number, or a numbering that has shifted. With the criminal codes having moved to the Bharatiya Nyaya Sanhita and the Bharatiya Nagarik Suraksha Sanhita, this risk is acute right now: a model may give you a CrPC section that no longer exists in the same form.
The final legal judgment. Whether to file, what relief to claim, which ground to lead with, how to characterise the respondent’s conduct: these are acts of professional judgment that carry your liability. AI can inform them. It cannot make them.
The rule of thumb is simple. If a wrong output would mislead the court or harm the client, that output must be verified by a human against an authoritative source before it goes anywhere near the petition.
There is a subtler version of this risk worth naming. A model will sometimes give you a correct legal principle but attach it to the wrong vehicle: the right proposition, the wrong case. Or it will state a rule that was correct under the old CrPC but has shifted under the BNSS. Because the substance reads as familiar and true, your eye slides past the error. This is why the verification has to be mechanical rather than impressionistic. You are not checking whether the output sounds right. You are checking, item by item, whether each cite, section, and rule is real, current, and correctly attributed. A thing can sound entirely right and still be wrong in a way that a registry or an opponent will catch.
The hallucinated-citation danger: real cases, real consequences
The hallucinated citation is not a theoretical risk. It has a documented track record, and the consequences are severe.
The case that put it on the map is Mata v. Avianca, decided in the Southern District of New York in 2023. Two attorneys submitted a brief built on judicial decisions that ChatGPT had simply invented, complete with names, citations, procedural histories, and quoted passages. When the lawyer asked ChatGPT to confirm the cases were real, the chatbot said they were. They were not. Judge P. Kevin Castel imposed sanctions of USD 5,000 against the lawyers and their firm, and required them to send corrective letters to every judge falsely named as the author of a fabricated opinion. As the Wikipedia summary records, the episode became the reference point for what blind reliance on AI looks like in practice.
What made it instructive was not the size of the fine. It was the mechanism. The lawyer did not set out to deceive. He used a tool he did not understand, asked it to verify its own output, and trusted the answer. The model produced text that was indistinguishable in form from a real citation. That is the trap: a hallucinated citation does not look wrong. It looks exactly right.
| Failure mode | What the AI produces | Why it is hard to catch |
|---|---|---|
| Invented case | A plausible name, citation, and year | Reads like any real cite |
| Misquoted holding | A real case with a wrong proposition attached | The case exists, so it passes a glance |
| Overruled authority | A real case that is no longer good law | Looks valid until you check its history |
| Wrong court or year | A real principle attributed to the wrong bench | Subtle, easy to miss under deadline |
The consequences are not confined to a fine. Submitting fabricated authority can expose an advocate to professional misconduct proceedings, contempt, and the reputational damage of having a judge record in an order that your citations were fake. That is a stain that follows a practitioner.
What the Indian courts are now saying
For a while, Indian lawyers could treat Mata v. Avianca as an American cautionary tale. That is no longer true.
In December 2024, the Bengaluru bench of the Income Tax Appellate Tribunal passed an order in the Buckeye Trust matter that cited Supreme Court and Madras High Court rulings which did not exist. As Analytics India Magazine reported, the tribunal recalled the order within a week. The problem had crossed over from speculation to documented fact in an Indian forum.
It escalated sharply in early 2026. The Supreme Court took suo motu cognisance of an Andhra Pradesh trial court order in a property dispute that had relied on four AI-generated, non-existent judgments. A bench of Justices P. S. Narasimha and Alok Aradhe held that a decision based on such fake judgments would not be an error in decision-making but would amount to misconduct, attracting legal consequences. The Court appointed senior advocate Shyam Divan as amicus and issued notices to the Attorney General, the Solicitor General, and the Bar Council of India.
Around the same time, a separate bench led by Chief Justice Surya Kant flagged the practice of lawyers filing AI-drafted petitions without checking them, after encountering a non-existent judgment styled “Mercy versus Mankind”. As India TV reported, the bench observed that some lawyers had started using AI to draft petitions and called it “absolutely uncalled for”. The Court later described the wider problem of fabricated AI judgments as a “rampant menace” affecting courts across jurisdictions, and asked the Bar Council of India to constitute an expert committee on AI use in court proceedings.
It is worth being precise about what this body of observations does and does not establish. As of mid-2026, the Bar Council of India has not yet published formal, India-specific AI ethics guidelines for advocates; an expert committee process is underway. What is already clear from the bench is the direction of travel: knowingly placing fabricated authority before a court is being treated as misconduct, and the duty to verify rests squarely on the advocate who signs the filing. The duties already exist under the Advocates Act 1961 and the Bar Council rules: not to mislead the court, to act competently, to maintain confidentiality. AI does not dilute those duties. It tests them.
A safe human-in-the-loop drafting workflow
So how do you actually use AI on a writ without ending up in the next news report? The answer is a workflow with the human at the centre and the AI on a tight leash. Here is a sequence that works.
Step 1: Brief the model on facts you control. Start by feeding the model the facts, the impugned order, and the relief sought, drawn from your own file. Do not ask it to research the law yet. Ask it only to produce a first draft of the synopsis and a structured set of grounds based on the material you have given it.
Step 2: Let AI draft the structural parts. Have it draft the synopsis, organise the grounds into a numbered structure, and produce a draft list of dates from the chronology you supplied. This is the transform-this-material work where it is strong.
Step 3: Take ownership of the legal substance. Now you, the advocate, supply the authorities. Identify the cases and provisions yourself, from a reliable source, and slot them in. Never let the model “fill in” the case law. If you ask it to suggest authorities, treat every suggestion as a lead to be checked, not a citation to be used.
Step 4: Verify every cite against the source. Pull up each judgment you are relying on from a trusted database and read the actual text. Confirm the case exists, the citation is correct, and the holding genuinely supports the proposition you are using it for. This is non-negotiable and it is covered in detail below.
Step 5: Run a good-law check. A real case is not enough. Confirm it has not been overruled, reversed, or distinguished into irrelevance. A judgment that was good law five years ago may not be today. See our note on good-law checking for how to do this properly.
Step 6: Compliance and final review. Check the petition against the specific High Court rules for format, annexures, affidavit, and court fees. Then read the whole thing yourself, end to end, as the advocate who is about to sign it.
Step 7: Sign. Your signature is your verification. By the time you sign, every fact, every date, and every citation should have been checked by a human against an authoritative source.
| Stage | Owner | The discipline |
|---|---|---|
| Synopsis and grounds draft | AI | Drafts from your facts only |
| Authorities | Lawyer | You find them, you verify them |
| Citation check | Lawyer | Every cite against the source |
| Good-law check | Lawyer | Confirm it is still binding |
| Rule compliance | Lawyer | Local High Court rules |
| Final read and sign | Lawyer | Own every line |
The structure of this loop is the whole point. AI does the drafting. The lawyer does the law. The two never blur.
The verification discipline: every cite, against the source
The citation check is where most AI accidents are prevented or caused, so it deserves its own discipline. The rule is absolute: read the source, not the summary.
For each authority in your draft, do the following.
- Confirm the case exists. Open a reliable database and find the judgment by name and citation. If you cannot find it, it does not go in. A citation you cannot locate is a citation you cannot use.
- Confirm the citation is accurate. Check the reporter, volume, page, and year against the actual report. AI frequently produces a real case with a mangled citation, or a real citation pointing to a different case.
- Read the holding yourself. Do not trust the model’s one-line summary of what the case held. Read the relevant paragraphs. A model will confidently attach a proposition to a case that the case never decided.
- Check the proposition actually fits. A case can exist, be cited correctly, and still not support your ground. Make sure the ratio applies to your facts, not just the headnote.
- Run the good-law check. Confirm the judgment has not been overruled or doubted by a larger bench.
This is also why grounded tools matter. There is a meaningful difference between a general chatbot that generates a citation from pattern, and a system that retrieves from a verified corpus of real judgments and shows you the source. The first invites the Mata problem. The second lets you click through to the actual judgment and read it. When you are choosing what to use for the law part of the workflow, that distinction is the one that counts. Our piece on finding similar judgments covers how grounded retrieval changes the research half of this work.
A practical habit: keep a verification log for each petition. One line per authority, with a tick once you have located it, read it, and confirmed it is good law. When you sign, the log is your evidence that you did the work. If a citation is ever challenged, you have a record.
Prompting tips for petition drafting
How you prompt shapes how safe and useful the output is. A few habits make a large difference.
Give it your material, not a blank slate. The more grounded the prompt, the safer the output. Paste the facts, the impugned order, the relief sought. A model drafting from your file is transforming; a model drafting from nothing is inventing.
Forbid invented citations explicitly. Tell the model, in the prompt, not to cite any case unless you have supplied it. A useful instruction: “Draft the grounds using the legal principles I describe, but do not insert any case citations. I will add authorities myself.” This removes the single biggest failure mode at the source.
Ask for structure, not conclusions. “Organise these points into discrete grounds, each tying a fact to a legal principle” is a good prompt. “Tell me whether this petition will succeed” is not, because the answer is a guess dressed as advice.
Use it to critique, not to certify. “Read this draft and list any gaps in the jurisdiction pleading or the prayer” is a safe, useful prompt. You then verify each gap it flags.
Iterate on the synopsis. The synopsis is persuasive writing, and AI is genuinely good at a second and third pass. Ask it to tighten, to lead with the strongest point, to cut repetition. This is pure drafting craft, and it is safe.
Make the model state its uncertainty. A small instruction pays off: ask the model to mark anything it is unsure about rather than smoothing over it. Phrases like “flag any point where you are guessing” or “do not state as fact anything I have not given you” push the output toward honesty. A model that flags a gap is far more useful than one that quietly fills it with a plausible invention. You cannot rely on this to catch every hallucination, and it is no substitute for the verification step, but it shifts the default from confident fabrication toward visible doubt, which is the safer failure mode to work with.
Keep prompts narrow and sequential. One task per prompt produces cleaner, more checkable output than a single sprawling instruction that asks for the synopsis, the grounds, the chronology, and the authorities at once. Drafting in small, ordered steps also mirrors the workflow above: structure first, your authorities second, verification third. Each step is easier to audit when it is the only thing the model just did.
| Prompt goal | Safe phrasing | What to avoid |
|---|---|---|
| Draft grounds | ”Structure these points into numbered grounds" | "Find cases that support this” |
| Summarise an order | ”Summarise the operative findings of this order" | "Tell me if this order is wrong in law” |
| Critique structure | ”List gaps in the jurisdiction averment" | "Confirm this petition is complete” |
| Improve prose | ”Tighten this synopsis, lead with the stay point" | "Add authority to strengthen this” |
Confidentiality and privilege when using foreign LLMs
There is a quieter risk that sits underneath the citation problem: what happens to the data you paste in.
An advocate owes a duty of confidentiality to the client, and legal communications attract privilege. When you paste a client’s facts, a draft petition, or the contents of an impugned order into a consumer AI tool, you are sending privileged material to a third-party server, often outside India. As commentators have noted, many general-purpose LLMs may use user inputs to train and refine the system, which is plainly problematic where the input is confidential or privileged.
This creates a real tension. The default settings of a free consumer chatbot are not built for privileged legal work. Before you put client material into any AI tool, you need to be clear on a few things.
| Question | Why it matters |
|---|---|
| Is my input used for training? | If yes, privileged material may leak into the model |
| Where is the data stored? | Foreign storage raises cross-border and jurisdiction concerns |
| Is there a data-retention setting? | Retained prompts are retained risk |
| Does the client know? | Where AI did meaningful work, best practice is to tell the client |
| Is the tool built for legal use? | Enterprise legal tools usually offer no-training and retention controls |
The practical guidance: do not paste sensitive client material into a free consumer chatbot. Use a tool with enterprise terms that contractually exclude your inputs from training and give you control over retention. Where AI has done meaningful work on a matter, the emerging best practice, drawn from how the Bar Council of England and Wales frames it, is to disclose that to the client. The duty of confidentiality under the Advocates Act 1961 does not pause because a tool is convenient.
Court-rule compliance: the part AI keeps getting wrong
The last gap is the most mundane and the one AI is least equipped to close: local procedure.
Every High Court runs on its own rules. The Delhi High Court rules call for petitions to be typed in double space in a specified font, for a memo of parties listing the email of every party, and for a synopsis and list of dates with every petition. Other High Courts have their own formatting, court-fee, affidavit, and annexure requirements. A registry will return a petition over a formatting defect just as readily as over a substantive one, and a returned petition costs you days.
A general AI model has no reliable knowledge of the registry you are filing in. It may produce a generically correct petition that fails the specific local rule. So this stage stays fully human.
- Check the format rules of the specific High Court: spacing, font, margins, paper.
- Check court-fee requirements and the number of copies.
- Confirm the affidavit is in the required form and properly sworn.
- Index and paginate the annexures as the rules require.
- Confirm the memo of parties carries every detail the local rule demands.
If you want a fuller treatment of the filing mechanics themselves, see our guide on how to file a writ petition, and our explainer on the five writs for choosing the right relief. The point for our purposes is narrow: compliance is not a drafting task you can delegate to a model. It is a procedural task you own.
Frequently asked questions
Can I use AI to draft a writ petition in India?
Yes, for the right tasks. AI can draft a first version of the synopsis, structure your grounds, build a list of dates, and summarise annexures, all from material you supply. What you cannot do is rely on it for case citations, local court rules, or the final legal judgment. The Supreme Court has signalled that filing AI-drafted petitions without verification is unacceptable, so the human verification step is mandatory, not optional.
What is the biggest risk of using AI for legal drafting?
The hallucinated citation. A model can produce a case name, citation, year, and quoted holding that look completely real but refer to a judgment that does not exist. This is exactly what sanctioned the lawyers in Mata v. Avianca in the United States, and it has now surfaced in Indian forums. Every citation must be verified against the actual source before it goes into a petition.
Has any Indian court penalised lawyers for AI-fabricated citations?
Indian courts have flagged the problem seriously. In early 2026 the Supreme Court took cognisance of an Andhra Pradesh trial court order that relied on four non-existent AI-generated judgments and observed that reliance on such fake judgments can amount to misconduct. A separate bench criticised lawyers filing AI-drafted petitions without checking them. The Court has asked the Bar Council of India to form an expert committee on AI use in court proceedings.
Does the Bar Council of India have AI guidelines for lawyers?
As of mid-2026, the Bar Council of India has not published formal, India-specific AI ethics guidelines for advocates. An expert committee process is underway following the Supreme Court’s notice. The existing duties under the Advocates Act 1961 and the Bar Council rules already apply: do not mislead the court, act competently, and maintain confidentiality. AI does not change those duties.
Is it safe to paste client documents into ChatGPT for drafting?
Be very careful. Many consumer AI tools may use your inputs to train their systems, which is a serious problem when the input is confidential or privileged client material. Do not paste sensitive material into a free consumer chatbot. Use a tool with enterprise terms that exclude your inputs from training and give you control over data retention.
How do I verify a case the AI suggested?
Open a reliable legal database and locate the judgment by name and citation. Confirm the citation is accurate, read the actual paragraphs to confirm the holding supports your point, and run a good-law check to confirm it has not been overruled. If you cannot locate the case, it does not go into the petition. Treat every AI-suggested authority as a lead to be checked, never a citation to be used.
What should I never delegate to AI when drafting a writ?
Four things: case citations, jurisdiction-specific High Court rules, the current text of statutory provisions, and the final legal judgment about what to file and how. These are the areas where a wrong output either misleads the court or harms the client, and they carry your professional liability. Keep them fully human.
Drafting writs with Indian-law grounding
The thread running through all of this is grounding. The danger of AI in petition drafting is not the drafting; it is the invention. A model that generates citations from pattern will eventually hand you one that does not exist, and under deadline you may not catch it.
Niyam is built around the opposite premise. Its drafting works from a verified corpus of real Indian judgments, so when it points you to an authority, you can click through and read the actual judgment rather than trusting a generated summary. The structural work, the synopsis, the grounds, the chronology, gets the AI speed; the legal substance stays anchored to sources you can verify. That is the workflow this whole post describes, built in rather than bolted on.
You still own every line. But you own it with a tool that is designed to keep you on the right side of the line the Supreme Court has now drawn.
Start for ₹100 and draft your next writ petition on grounded Indian law. Create your account and put the verified-first workflow to work.