TL;DR: If you use an AI tool to help draft a pleading, an application, or evidence, India’s draft court rules want you to say so. The Supreme Court’s AI Committee released the Draft Regulations for Use of Artificial Intelligence in Courts, 2026 on 3 June 2026, and the disclosure duty sits at the centre of them. As reported in The Leaflet’s clause-by-clause explainer, Regulation 42 of the draft requires a party or counsel who used AI in preparing any document to disclose that at the time of filing, and lets the court ask which AI system was used, how much it did, and what steps you took to verify accuracy. Regulation 43 then puts the whole risk on you: if the AI-generated content turns out to be fabricated or false, you bear the responsibility, and “it was the AI” is not a defence. These are draft provisions open for comment, not yet binding law. But the duty they describe is already being enforced through case law and one High Court’s binding policy. This guide is about that narrow obligation, what you must declare, who signs, and what happens when AI use is hidden, not a general tour of the AI rules.
On this page
- What the disclosure duty actually requires
- Who certifies, and what the declaration must say
- Your verification duty does not transfer to the machine
- What happens when AI use is hidden and the citations are fake
- How the disclosure angle differs from a generic AI rules overview
- Settled versus proposed: what binds you today
- A disclosure checklist before you file
- How Niyam helps you file AI-assisted pleadings safely
- Frequently asked questions
What the disclosure duty actually requires
Start with the thing that is new. Most coverage of the 2026 draft talks about AI in courts as a broad theme. The disclosure duty is sharper than that, and it is the part that touches your daily filing practice.
The Supreme Court’s AI Committee published the draft on 3 June 2026 and invited public comments until 20 June 2026, with submissions to [email protected]. The framework is built on a stated presumption in favour of responsible AI adoption, so it does not ban the tools. It conditions their use on candour. The price of using AI to help with a court document is telling the court you did.
The disclosure obligation, as set out in The Leaflet’s reading of the draft, works in two directions. First, a litigant or legal representative who used AI in the preparation or summarisation of any document must disclose that use to the court at the time of submission. Second, the court is given a matching power to ask for details: which AI system was used, the nature and extent of the assistance it gave, and the steps taken to verify the accuracy of any AI-generated content. Disclosure is not a one-line ritual you tick off. It is the entry point to a possible audit of your process.
One scope point matters before you assume this is a Supreme Court problem. The draft, by its own terms, reaches across the system. As Business Standard reported, the regulations are framed to apply to the Supreme Court, every High Court, subordinate courts, and tribunals and statutory commissions exercising adjudicatory functions. Whether the Supreme Court can validly bind High Courts through such regulations is itself contested, a question examined in LiveLaw’s analysis of the draft’s reach over High Courts. That debate is live. The disclosure principle, though, is the same wherever you practise.
Two facts to hold from this section. The duty triggers on use, not on outcome, so a clean AI-assisted draft still needs disclosure. And the court’s power to probe your verification means the declaration is the start of accountability, not the end of it.
Who certifies, and what the declaration must say
Here is where the draft is precise about one thing and quiet about another, and you need to know which is which.
The draft is clear that the disclosure rests on the person filing. The litigant and the legal representative carry the duty to inform the court. There is no provision shifting it to the registry, the court’s own systems, or the vendor of the tool. If your name is on the vakalatnama, the candour obligation is yours. If you are unsure what that filing authority involves, the basics are set out in what a vakalatnama is and why it matters.
What the draft does not yet spell out, in the version released for comment, is a single mandatory certificate format. Several early summaries described a “declaration or certificate in a prescribed format,” and the direction of travel clearly points to a standard form. But the careful clause-by-clause readings of the draft do not reproduce a fixed template, and you should treat the exact wording as something the final version will settle. The honest position today is that the obligation is firm and the form is not.
That gap is not an excuse to say nothing. A workable disclosure, on the face of the draft’s requirements, should cover four things:
- That AI was used, and at what stage. Drafting, summarising, research, or formatting are different levels of involvement, and the court is entitled to know which.
- Which tool or system. A named tool tells the court whether it was a general consumer chatbot or a legal-specific platform, which bears on reliability.
- The extent of the assistance. A model that generated a full argument is a different risk from one that tidied your grammar.
- The verification you did. This is the part the court can ask about under the draft, so it is the part worth documenting before you are asked.
The reason the “who” is so blunt is that the rest of the framework refuses to let accountability drift. The draft’s accountability provision treats every decision involving AI as advisory and attributes it to the human officer who made it, and it says accountability cannot be avoided by pointing to a “black box” or a hallucination. The person stays in the frame. Disclosure is simply the first place that principle bites.
Your verification duty does not transfer to the machine
This is the heart of the disclosure regime, and it is older than the 2026 draft. The duty to check your own citations did not arrive with AI. AI just made it easier to forget.
The draft states the rule plainly. Where the AI-generated content in any document, pleading, or evidence turns out to be fabricated, false, misleading, or inaccurate, the responsibility is borne by the person who submitted it, and that person cannot rely on the AI character of the output as a defence. Read that twice. The rule is not “disclose and you are safe.” It is “disclose, and you still own every word.” Disclosure buys candour, not immunity.
The duty has independent roots in professional conduct and in how courts have always treated citations. A lawyer who signs a pleading is representing that the law in it is real. That has not changed. What the verification duty adds in the AI age is the discipline of treating model output as a draft to be checked against a primary source, never as a source in itself. The practical mechanics of that check are covered in the lawyer’s duty to verify AI output and in how to vet a legal AI tool for citation accuracy.
Judges have been saying the same thing in plainer language for over a year. Justice B.R. Gavai, then a Supreme Court judge and later Chief Justice of India, warned in March 2025 that “relying on AI for legal research comes with significant risks, as there have been instances where platforms like ChatGPT have generated fake case citations and fabricated legal facts,” as reported by LiveLaw. The warning is not anti-technology. It is a reminder that a tool that sounds authoritative is the most dangerous kind to leave unchecked.
The verification duty is why the choice of tool matters as much as the choice to disclose it. A general chatbot trained to produce fluent text will invent a citation that looks perfect rather than admit it does not know. A legal research system that answers only from real judgments, with each proposition linked to a case you can open, is a different risk profile entirely, a distinction drawn out in native legal AI for India versus generic GPT. Verification is far cheaper when the tool was built to be verified.
What happens when AI use is hidden and the citations are fake
The draft regulations describe a duty. The case law shows the cost of breaching it. India already has a run of orders and filings undone by AI-generated authorities that nobody checked, and these are decided matters, not hypotheticals.
The most cited example is a tribunal, not a lawyer. In Buckeye Trust v. PCIT, ITA No. 1051/Bang/2024, the Bengaluru bench of the Income Tax Appellate Tribunal passed an order on 30 December 2024 in a dispute over a trust settled with assets of about ₹669 crore. The order relied on Supreme Court and High Court authorities that did not exist, including a fabricated “S. Gurunarayana v. S. Narasimhulu (2004) 7 SCC 472.” Within roughly a week the tribunal recalled the order under Section 254(2) of the Income Tax Act. The original order remains on the ITAT’s own record, a useful reminder that a phantom citation can survive into a signed order if no human opens the case.
The High Courts have been blunter. On 6 October 2025 the Bombay High Court quashed a faceless income tax assessment that had pushed a company’s assessed income to ₹27.91 crore on the strength of non-existent, AI-generated case laws. The division bench held that authorities exercising quasi-judicial power cannot blindly rely on AI results and must cross-verify every reference before using it, as set out in Taxguru’s report on the judgment. The court did not just set the order aside. It built in a safeguard, directing that any judgments relied on in the fresh proceeding be put to the assessee with at least seven days’ notice. A breach of natural justice and a fabricated authority, in other words, travel together.
The Delhi High Court caught the problem on the lawyer’s side. A single-judge bench of Justice Girish Kathpalia was hearing a homebuyers’ petition when senior counsel flagged that it leaned on ChatGPT-generated material. The petition had quoted paragraphs 73 and 74 of Raj Narain v. Indira Nehru Gandhi, a judgment that, as ThePrint reported, contains only 27 paragraphs in reality. The court called the use of AI in that manner “impermissible” and a serious concern, and allowed the plea to be withdrawn. The phantom paragraph is the tell. AI does not just invent cases. It invents passages inside real ones.
These are not isolated. The broader pattern, including a property dispute where the Andhra Pradesh High Court flagged four non-existent Supreme Court judgments as AI-generated, is mapped in LiveLaw’s survey of phantom precedents in Indian courts. The thread running through all of them is the same: the AI was never disclosed, the citations were never checked, and the filing collapsed once someone opened the cases. A deeper treatment of the failure mode itself sits in AI-hallucinated citations in India.
How the disclosure angle differs from a generic AI rules overview
It is worth being precise about what this duty is and is not, because the disclosure obligation is easy to fold into the larger “AI in courts” conversation and lose. For the wider framework, the structure of the regulations, the governance body, and what tools qualify, see the Supreme Court’s AI rules for India and what those rules require of the tools you use. This section stays on the narrow question of declaration.
A general AI rules overview answers “is AI allowed in court, and how is it governed.” The disclosure duty answers a different and more personal question: “what must I, the person filing, say and do when I have used a tool.” The first is institutional. The second is a conduct rule that attaches to your signature on a specific document.
The cleanest way to see the difference is to compare two filings that are identical except for the AI step.
| Element | AI-assisted draft, disclosed and verified | AI-assisted draft, hidden and unchecked |
|---|---|---|
| AI use stated to the court | ✓ Declared at the time of filing | ✗ Concealed |
| Tool and extent of assistance | ✓ Named and described | ✗ Unknown to the court |
| Every citation opened and confirmed | ✓ Checked against the primary source | ✗ Pasted from model output |
| Position if a citation is fake | ✓ Candour on record, error correctable | ✗ No AI defence, full responsibility |
| Risk to the client and the matter | ✓ Contained | ✗ Order recalled, plea withdrawn, costs |
The table makes the practical point. Disclosure does not make a bad citation good. It changes whether you were candid about the process that produced it, and candour is what courts reward when something goes wrong. The verification column is what actually protects the client. Disclosure plus verification is the safe combination. Either one alone is not.
There is a second difference worth naming. A generic overview treats AI use as a yes or no. The disclosure regime treats it as a spectrum, which is why the draft asks for the “nature and extent” of assistance. Using a tool to summarise a long judgment is not the same as using it to write your grounds, and the disclosure should reflect that. If summarisation is your use, the reliability questions specific to it are covered in AI judgment summarisation.
Settled versus proposed: what binds you today
This is the section to read twice, because mixing up what is law with what is draft is the fastest way to get the duty wrong. The 2026 Supreme Court regulations are a draft. They were released for public comment and the comment window has closed, but the version you have read about is not yet a binding rule of court. Treat its specific clause numbers and any “prescribed certificate” as provisional until the final text issues.
What is already binding is narrower and more concrete. One High Court has put a formal AI policy in force. On 19 July 2025 the Kerala High Court issued its Policy Regarding Use of Artificial Intelligence Tools in District Judiciary, the first such binding document from any High Court in the country. It applies to the district judiciary, the staff assisting them, and interns and law clerks. It says AI may be used only as an assistive tool, never as a substitute for decision-making or legal reasoning. It restricts general-purpose generative tools on confidentiality grounds and steers users to approved tools. And it has teeth: a violation can lead to disciplinary action under the Kerala Civil Services (Classification, Control and Appeal) Rules, 1960. That policy binds judicial officers, not advocates, but it is the clearest signal yet of how an Indian court treats unverified AI in its own house.
The other thing that already binds you is the existing law of professional conduct and the courts’ inherent power over their own process. You do not need the 2026 regulations to be penalised for filing a fabricated citation. The decided cases above, Buckeye Trust, the Bombay High Court assessment, the Delhi High Court plea, all happened under existing law, before any AI-specific rule was in force. The duty to verify is not waiting for the draft to be notified. It is here now.
| Source | Status | What it binds | Force today |
|---|---|---|---|
| Draft 2026 Supreme Court AI regulations | Proposed, out for comment | Courts, parties, counsel across the system | ✗ Not yet binding |
| Kerala High Court district judiciary AI policy | In force from July 2025 | Judicial officers and staff in Kerala | ✓ Binding now |
| Existing professional conduct and court powers | Settled law | Every advocate and litigant | ✓ Binding now |
| Decided hallucination cases | Decided | Persuasive across courts | ✓ Already enforced |
The takeaway is simple. Do not wait for the regulation to be notified before you change how you file. The conduct it describes, disclose and verify, is already enforceable through the routes that undid the cases above. For the wider context of how AI is being received by the judiciary, AI in Indian courts sets the scene.
A disclosure checklist before you file
Theory is easy to nod at and hard to apply at 11pm before a filing deadline. Here is the duty reduced to steps you can actually run.
Before you draft, decide your tool. If a document is going to a court, use a system built to answer from real Indian judgments with citations you can open, not a consumer chatbot that optimises for fluent text. The reliability gap is the whole game, and it is the difference traced in AI legal research for India.
While you draft, log what the tool did. Note the stage, the tool, and the extent of assistance as you go. Reconstructing it later, when the court asks under the draft’s verification power, is harder and less accurate. Contemporaneous notes are the best evidence that you took the duty seriously.
Before you sign, open every authority. Each case, every quoted paragraph, the citation, the holding, the proposition you are relying on it for. If a passage is quoted, confirm the paragraph exists and says what the draft claims, because the phantom paragraph inside a real case is the failure that caught the Delhi petition. Then confirm the case is still good law, since an overruled judgment is as useless as an invented one. The method is in good law checking, and the citation mechanics are in how to cite Indian judgments correctly.
At filing, state the AI use. Cover the four points: that AI was used, at what stage, which tool, and the verification you did. Keep it factual and specific. A vague “AI may have been used” is worse than a precise account, because it reads as evasion.
If you are using AI further upstream, in drafting the petition itself, the same discipline applies to the drafting tool, and the workflow is set out in using AI to draft a writ petition and AI contract drafting. The rule does not change with the document type. Disclose the use, own the output, check the law.
How Niyam helps you file AI-assisted pleadings safely
The disclosure duty is really two duties wearing one coat: be candid that you used AI, and be sure the law you filed is real. The first is on you. The second is where the tool you pick decides whether the duty is easy or impossible to meet.
Niyam is built so that the verification half is the default, not an afterthought. Ask a question in plain English, and Niyam answers from real Indian judgments, with every proposition tied to a case you can open and read. That is the opposite of the failure mode in the cases above, where a fluent-sounding tool produced citations that pointed at nothing. When the source is a real judgment you can click through to, checking it before you sign takes seconds, and your disclosure can honestly say the authorities were confirmed against the primary source.
It also makes the candour half cleaner. Because Niyam shows you the cases it relied on, you can describe the extent of the assistance accurately, name the tool, and record that you verified each authority, exactly the four things a disclosure should cover. Before you rely on a precedent, you can check whether it is still good law so you do not file an overruled case, and you can see how to vet any legal AI tool for citation accuracy to hold every tool, including this one, to the same standard. If you want to compare approaches before you commit, the comparison page lays out the difference.
The point is not that AI makes filing risky. It is that undisclosed, unverified AI does. A tool that is built to be checked, used by a lawyer who discloses and verifies, is exactly the responsible adoption the draft regulations say they want.
Start for ₹100
Try Niyam on your next pleading. For ₹100 you get credits to run real research grounded in Indian judgments, with every answer cited to a case you can open before you file. Create your account and start for ₹100.
Frequently asked questions
Do I have to disclose AI use in every court document?
Under the draft 2026 regulations, the duty triggers whenever you use AI in preparing or summarising a document, pleading, or evidence that you submit to the court. It is tied to use, not to whether the output was good. So a clean, fully verified AI-assisted draft still attracts the disclosure obligation. The duty is currently in a draft open framework, but the safest practice is to disclose now, because the existing law on citation accuracy already applies.
Is the “Regulation 19” label for AI disclosure correct?
Be careful with regulation numbers. The clause-by-clause readings of the published draft place the disclosure duty in Regulation 42 and the responsibility provision in Regulation 43, with accountability addressed separately. Because this is a draft released for comment, the numbering and exact wording may change in the final version. Rely on the substance of the duty, disclose AI use and own the output, rather than a specific clause number that has not been finalised.
Who has to sign or certify the AI declaration?
The duty rests on the person filing, meaning the litigant and the legal representative on record. The draft does not shift it to the court registry or the tool’s vendor. There is no separate official who “certifies” your declaration. You make it, and you stand behind it, which is consistent with the framework’s refusal to let accountability move to a “black box.”
Is there a prescribed format for the declaration?
The version of the draft released for comment does not reproduce a single fixed certificate template, even though some early summaries described one. Treat the format as unsettled and the obligation as firm. A workable disclosure states that AI was used, at what stage, which tool, the extent of the assistance, and the verification you carried out.
Does disclosing AI use protect me if a citation turns out to be fake?
No. Disclosure is about candour, not immunity. The draft says that if AI-generated content is found to be fabricated, false, or misleading, the person who submitted it bears the responsibility and cannot rely on the AI character of the output as a defence. You still have to verify every authority. Disclosure plus verification protects you. Disclosure alone does not.
Has any Indian court actually penalised AI-generated fake citations?
Yes, in decided matters and under existing law. The Bengaluru bench of the Income Tax Appellate Tribunal recalled its order in Buckeye Trust v. PCIT after it cited non-existent authorities. The Bombay High Court quashed a ₹27.91 crore faceless tax assessment on 6 October 2025 for relying on AI-generated case laws. The Delhi High Court treated a ChatGPT-assisted petition with invented paragraphs as impermissible and allowed it to be withdrawn.
Are the 2026 Supreme Court AI regulations binding now?
Not yet. They are draft regulations released by the Supreme Court’s AI Committee on 3 June 2026 for public comment. The final, notified version may differ. What binds you today is the existing law of professional conduct and the courts’ power over their own process, plus, for judicial officers in Kerala, that High Court’s binding AI policy.
Do these rules apply to High Courts and district courts or only the Supreme Court?
The draft is framed to apply across the Supreme Court, the High Courts, subordinate courts, and tribunals and statutory commissions exercising adjudicatory functions. Whether the Supreme Court can validly bind the High Courts through such regulations is a contested question that commentators have raised. The disclosure and verification principle, though, is being applied by courts at every level already.
What is the Kerala High Court AI policy, and does it bind advocates?
It is the Policy Regarding Use of Artificial Intelligence Tools in District Judiciary, issued by the Kerala High Court in July 2025, the first binding AI policy from an Indian High Court. It restricts AI use by judicial officers and staff, bars AI as a substitute for legal reasoning, and provides for disciplinary action on breach. It binds the judiciary and court staff, not advocates, but it signals how seriously courts treat unverified AI.
What exactly should my AI disclosure say?
Keep it specific. State that AI was used, the stage at which it was used, the name of the tool or system, the nature and extent of the assistance it provided, and the verification steps you took. The draft empowers the court to ask for precisely these details, so providing them upfront is both compliant and credible.
Does using AI only for summarising a judgment still need disclosure?
On the face of the draft, yes, because the duty covers preparation and summarisation of documents. Summarisation is a lower level of assistance than drafting an argument, and your disclosure should say so. It carries its own accuracy risks, since a summary can drop a crucial qualification, which is why the source judgment should still be read.
How is this different from the general Supreme Court AI rules?
The general rules govern how AI is allowed and supervised across the court system as an institution. The disclosure duty is a personal conduct rule attached to your specific filing: what you must declare and verify when you, as the person on record, have used a tool. One is institutional governance; the other is a duty that follows your signature.
Can I just avoid all this by not using AI?
You can, and for some filings that is a reasonable choice. But the verification duty still applies to anything your juniors or research staff produced with AI, and to authorities handed to you by others. The Bombay High Court’s seven-day-notice safeguard exists precisely because fabricated authorities can enter a matter from the other side too. Knowing the duty protects you whether or not you use AI yourself.
What does responsible AI use in pleadings look like in practice?
Pick a tool built to answer from real judgments with openable citations. Log what it did as you work. Open and confirm every authority and quoted paragraph against the primary source, and check it is still good law. Disclose the use plainly at filing. That combination meets both halves of the duty, candour and accuracy, and it is the responsible adoption the draft regulations say they are trying to encourage.