# Court AI disclosure rules: what your legal AI must do **TL;DR:** Reported draft regulations in India point in one clear direction - lawyers may use AI, but they must be able to disclose it, stand behind every output, and ensure a human remains responsible for each decision. That changes the standard for what a legal AI tool needs to do. General-purpose chatbots cannot meet that bar. A tool built around traceable citations, auditable sources, and clean data handling can. --- ## On this page - [What the reported draft direction says](#what-the-reported-draft-direction-says) - [Why courts are taking AI seriously now](#why-courts-are-taking-ai-seriously-now) - [The US precedent that every Indian lawyer should know](#the-us-precedent-that-every-indian-lawyer-should-know) - [What "disclosure-friendly output" actually means](#what-disclosure-friendly-output-actually-means) - [Why general-purpose AI fails the disclosure test](#why-general-purpose-ai-fails-the-disclosure-test) - [Compliant versus risky: a side-by-side comparison](#compliant-versus-risky-a-side-by-side-comparison) - [Traceable citations: the non-negotiable feature](#traceable-citations-the-non-negotiable-feature) - [Auditable sources and the chain of custody for your research](#auditable-sources-and-the-chain-of-custody-for-your-research) - [Data privacy under the DPDP framework](#data-privacy-under-the-dpdp-framework) - [What human-in-the-loop means in daily practice](#what-human-in-the-loop-means-in-daily-practice) - [How Niyam was built for this moment](#how-niyam-was-built-for-this-moment) - [Frequently asked questions](#frequently-asked-questions) --- ## What the reported draft direction says In the first half of 2026, reporting from within the Indian legal and regulatory space has described draft guidelines and policy signals moving in a consistent direction. The general thrust, as reported, is this: lawyers are not prohibited from using AI tools in their work, but when they do, they should be in a position to disclose that use and, more importantly, to vouch for every output as their own professional conclusion. Judicial decisions themselves, the reported direction makes clear, must rest with the judge. A court cannot delegate reasoning or findings to an algorithm. That principle has constitutional grounding: judicial power under the Constitution of India vests in courts and the judges appointed to them, not in software. For practising lawyers, the practical implication is straightforward. If you use an AI tool to research case law, draft a submission, or analyse a contract, you must be able to stand behind every sentence in that document as something you have checked, understood, and taken professional responsibility for. That is not a new duty in substance; lawyers have always had to verify their research. What is new is that AI raises the stakes, because AI can generate plausible-looking text that is factually wrong at a speed and volume that makes manual checking genuinely hard. The direction of travel is also consistent with what legal regulators in other jurisdictions have already articulated. The Bar Council of England and Wales, the American Bar Association, and bar authorities in Singapore and Australia have all moved toward disclosure-and-verify frameworks rather than blanket prohibitions. India appears to be moving along the same arc. Nothing here should be read as a precise statement of enacted law. The regulations described are in draft or reported form and the final text, if and when notified, may differ. What this piece addresses is the functional question: given the direction of travel, what does your legal AI tool need to be able to do? --- ## Why courts are taking AI seriously now The speed at which AI has entered legal practice has surprised even optimistic observers. Two years ago, a lawyer using a large language model for research was an early adopter. Today it is routine. That speed is precisely what has prompted courts and regulators to respond. Three concerns recur across every jurisdiction that has started addressing AI in legal practice. The first is hallucination. AI systems built on large language models can and do produce citations to cases that do not exist, quotations that were never written, and legal propositions that no court has ever endorsed. The problem is not that the AI is being dishonest - it has no capacity for honesty or dishonesty. The problem is structural: the model predicts the most statistically plausible continuation of your prompt, and in legal writing, a plausible continuation often includes a case citation. If you ask about the standard for interim injunctions in India and the model generates text that looks like a Supreme Court ratio, it will include a citation in the style of a Supreme Court citation regardless of whether that particular case exists. The second concern is accountability. A lawyer submits arguments to a court. That submission is a professional act carrying obligations of accuracy, candour, and good faith. If an AI generated part of that submission, who is accountable for its contents? The answer in every disclosure framework being proposed is: the lawyer, always and without exception. AI cannot be a professional. It cannot be a party before a court. It cannot be held in contempt. The human lawyer remains entirely responsible. The third concern is data. When a lawyer feeds a client's matter into an AI tool, questions arise about where that data goes, how it is stored, who can access it, and whether it is used to train future models. In an environment where India's Digital Personal Data Protection Act (DPDP) has established a statutory framework for personal data handling, this is not a theoretical concern. All three concerns point to the same conclusion: the relevant question is not whether to use AI but which AI, and under what conditions. --- ## The US precedent that every Indian lawyer should know Before discussing what Indian rules require, it is worth pausing on the case that put the legal profession on notice globally. In *Mata v. Avianca*, decided in the Southern District of New York in 2023, a lawyer filed a brief that cited multiple cases as authority for various propositions. The cases did not exist. They had been generated by ChatGPT, which the lawyer had used to research the submission. When the court asked for copies of the cited judgments, the lawyer could not produce them because they had never been decided. The court imposed sanctions, making observations about professional responsibility that have been widely cited in subsequent AI-and-law discussions. The *Mata* case is not Indian law. But it is the clearest illustration available of what happens when a lawyer uses an ungrounded AI system for legal research and submits its output without independent verification. Every bar authority and court that has subsequently addressed AI disclosure has cited or echoed this episode. It is the cautionary baseline from which every policy framework has proceeded. The lesson for Indian lawyers is not "do not use AI." The lesson is: if you use AI that cannot show you where its answer came from, you are in the same position as the lawyer in *Mata*. You cannot verify what you cannot trace. Read more about hallucination risk in the Indian context in our piece on [AI legal research in India and how to avoid citation errors](/blog/ai-legal-research-india). --- ## What "disclosure-friendly output" actually means The concept of disclosure-friendly output is worth unpacking because it is more specific than it sounds. A disclosure-friendly output is one where, if a court or opposing counsel asked you to demonstrate the basis for every claim in your submission, you could do so. That means three things. First, every legal proposition must trace back to a specific, identifiable source - a judgment, a statute, a notification. Not "according to various courts" or "case law establishes that" but a citation you can look up, read, and confirm. Second, the source must be verifiable independently of the AI tool. If the only way to check a citation is to ask the same AI again, the citation is not verifiable - it is circular. You need to be able to pull the judgment from an official source and confirm the proposition matches the actual holding. Third, the output must not overstate what the source says. An AI that takes a two-judge bench observation made in passing and presents it as a binding ratio of the Supreme Court has produced a disclosure-hostile output even if the underlying case exists. The characterisation matters as much as the citation. These three requirements - traceability, independent verifiability, and accurate characterisation - are not features that can be retrofitted onto a general-purpose language model. They require the system to be built around a corpus of real judgments and to surface the source alongside every answer. --- ## Why general-purpose AI fails the disclosure test ChatGPT, Claude, and Gemini are powerful tools. They are also, by design, unsuited for use as primary legal research engines in a disclosure-required environment. Understanding why clarifies what you need from a purpose-built alternative. A general-purpose LLM is trained on a broad corpus of internet text and generates responses by predicting likely continuations. It has no persistent, queryable database of Indian judgments that it searches when you ask a legal question. It draws on patterns learned during training. That means: - It cannot tell you which database it searched, because it did not search one. - It cannot give you a neutral citation (like (2024) 3 SCC 411) with confidence, because it is generating the number by pattern rather than retrieving it from a record. - It cannot confirm a case is still good law, because checking current citation status requires real-time access to a citator, not a language model. - It cannot distinguish what a judgment actually held from what commentators have said about it. None of this means these tools are useless for lawyers. They are excellent for drafting, summarising known positions, explaining unfamiliar concepts, and structuring arguments. The problem arises specifically when they are used as research tools and their outputs are cited to courts without independent verification. In a disclosure environment, the question shifts from "is this answer plausible?" to "can I prove this answer?" Plausibility and provability are different standards. General-purpose AI clears the first. It cannot clear the second. For a detailed comparison of generic versus purpose-built tools, see our [comparison page](/compare) and our piece on [native legal AI versus generic GPT for Indian practice](/blog/native-legal-ai-india-vs-generic-gpt). --- ## Compliant versus risky: a side-by-side comparison The following table illustrates the difference between output that is disclosure-ready and output that creates professional risk. | Dimension | Disclosure-ready output | Risky output | |---|---|---| | Citation format | Full neutral citation with year, volume, court, page - e.g. (2023) 5 SCC 210 | Partial or invented citation - e.g. "Supreme Court, 2022 (unreported)" | | Source link | Direct link to the judgment in the AI tool's corpus | No link, or a link that resolves to a different case | | Citation status | Citator check run and confirmed as good law | No citation check; case may have been overruled | | Proposition accuracy | Quoted text matches the actual judgment paragraph | Summary is a paraphrase that shifts the meaning | | Data handling | Client matter is not stored or used for model training; DPDP-compliant | Unclear data retention; no privacy assurance | | Auditability | Full trail from query to source judgment is available | No audit trail; cannot reconstruct how answer was reached | | Overstatement risk | Tool flags when a case is from a non-binding bench or jurisdiction | Tool presents High Court observation as binding ratio | | Human sign-off | Tool outputs are presented as drafts requiring lawyer review | Tool presents outputs as final answers | The distinction is not about whether the answer is correct in a general sense. It is about whether you can prove it is correct in a way a court would find satisfactory. --- ## Traceable citations: the non-negotiable feature If you take one functional requirement from the emerging disclosure framework, it is this: every legal proposition your AI produces must be traceable to a specific judgment in a specific database. Traceability has two parts. The first is retrieval - when the AI answers your question, it must be pulling from actual judgments, not generating text by statistical inference. Retrieval-augmented generation, where the model searches a real corpus before generating a response, is the technical mechanism that makes this possible. The second part is surfacing. Retrieval is not enough if the system does not show you what it retrieved. A tool that searches a judgment database internally but only shows you the answer (not the source) does not give you the traceability you need. You need to see the citation, click through to the judgment, and read the relevant passage yourself. Niyam searches across 72,000+ Supreme Court and High Court judgments from India and surfaces the source alongside every answer. The citation, the court, the bench, the date, and the relevant passage are all visible. You are not being asked to trust the AI's characterisation. You are being given the primary material to read. This is the minimum standard a disclosure-friendly tool must meet. Anything less and you are in the position of the lawyer in *Mata* - asked to produce the source and unable to do so. For more on evaluating citation accuracy in legal AI tools, see our guide on [how to vet legal AI citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy) and our [good law checking explainer](/blog/good-law-checking). --- ## Auditable sources and the chain of custody for your research Traceability covers individual citations. Auditability covers your research process as a whole. In a legal context, auditability means you can reconstruct what you searched, what the tool returned, what you relied on, and what you chose not to rely on. This matters in at least two scenarios. The first is professional accountability. If a court questions how you arrived at a legal position, you should be able to walk through your research process. "I searched for precedent on X, the tool returned three cases, I verified each of them, and I relied on the first and third because they were binding on this bench" is an auditable account. "I asked an AI and it gave me this answer" is not. The second is internal review. Law firms have quality control processes. Senior lawyers review submissions before they go out. If the junior who did the AI-assisted research cannot explain the chain from query to citation, the review process is compromised. A well-designed legal AI tool maintains a session history, stores your queries and the results alongside the source judgments, and allows you to export or share that research trail. This is not a luxury feature. In a disclosure environment, it is a basic professional requirement. Explore how Niyam structures research sessions on our [research solutions page](/solutions/research). --- ## Data privacy under the DPDP framework The Digital Personal Data Protection Act and its implementing rules create obligations for anyone who processes personal data in India. For lawyers, this creates a set of questions about every AI tool you use in practice. When you describe a client's situation to an AI system - the facts of their dispute, the names of parties, commercial terms of a transaction - you are processing personal data. The question is: what does the tool do with it? General-purpose AI tools commonly use conversations for model improvement. Your input may be reviewed by humans at the AI company, stored, and potentially used to train future model versions. This is typically disclosed in terms of service that most users do not read. Under the DPDP framework, sharing personal data with a third-party processor without appropriate consent and safeguards creates legal exposure. A legal AI tool built for professional use should operate on a different basis. It should process your query to return results but not retain client-related data for training purposes. Data handling practices should be transparent and consistent with the obligations a lawyer carries toward a client. Niyam's approach is to treat your queries as professional work product. We do not use your research sessions to train our models. Our data handling is designed around the DPDP framework and the professional confidentiality obligations that govern legal practice. For a fuller treatment of DPDP and what it means for legal AI use, see our [DPDP rules 2025 overview](/blog/dpdp-rules-2025) and the comparison of [AI tools for lawyers in India](/blog/best-ai-tools-for-lawyers-india). --- ## What human-in-the-loop means in daily practice The phrase "human-in-the-loop" comes from AI safety literature and has a specific technical meaning - a human checks AI output before it has real-world effect. In the context of legal practice and the direction of regulatory travel, it translates into concrete habits. It means you read every case the AI cites before you cite it yourself. You do not forward the AI's output to a client without reading it. You do not file a submission that includes AI-generated text you have not personally reviewed paragraph by paragraph. This sounds obvious but in practice, time pressure creates shortcuts. A junior associate under deadline hands off AI output to a partner who is also under pressure. The partner signs off without reading the citations. One of them does not exist, or has been overruled, or says the opposite of what the summary claimed. None of this would happen if every person in the chain treated AI output the way they would treat a trainee's first draft: as a starting point that requires verification, not a finished product that can be submitted as-is. Human-in-the-loop is not an inconvenience imposed by regulators. It is a description of responsible practice that lawyers have always been expected to follow. AI does not change that duty. It raises the consequences of ignoring it. Tools that make human review easier - by surfacing the source alongside the answer, by flagging overruled cases, by presenting outputs clearly as drafts - support responsible practice. Tools that present AI output as authoritative answers work against it. Our [citator tool](/solutions/citator) is specifically designed to support the verification step - checking that a case is still good law before you cite it. --- ## How Niyam was built for this moment Niyam is a legal AI built specifically for Indian practice, on Indian law, with the professional obligations of Indian lawyers in mind. The architecture choices we made from the start happen to align closely with what a disclosure framework requires. Every answer Niyam gives traces back to a real judgment in our corpus of 72,000+ Indian Supreme Court and High Court decisions. We do not generate citations - we retrieve them. When you ask Niyam a question, the answer comes with the source judgment, the relevant passage, and the citation in the format Indian courts expect. Our citator checks whether the cases Niyam surfaces are still good law. An answer built on an overruled case is not just unhelpful - it is professionally dangerous. Citator checking is not an add-on. It is part of the research loop. We do not use your research sessions to train our models. Your client matters stay within your session. Our outputs are presented as research support, not as legal conclusions. The interface is designed to prompt verification, not to short-circuit it. Every response is a starting point for your professional judgment, not a substitute for it. We believe that when Indian courts and regulators finalise AI disclosure frameworks, the standard they set will look very much like what Niyam already does. Not because we predicted the specific text of any regulation but because the underlying values - accountability, accuracy, privacy, human judgment - are the same values that principled legal AI design starts from. Read how our approach compares to general-purpose tools in our [AI legal research guide for India](/blog/ai-legal-research-india), our [ChatGPT for lawyers India analysis](/blog/chatgpt-for-lawyers-india), and the full [tools comparison](/blog/best-ai-tools-for-lawyers-india). **Ready to use a legal AI that is built for disclosure?** Start your research with the ₹100 trial - 200 credits, cancel anytime. [Start for ₹100](https://app.niyam.ai/register) or write to us at hello@niyam.ai with questions. --- ## Frequently asked questions ### Are Indian courts currently enforcing AI disclosure rules for lawyers? As of mid-2026, reported draft guidelines and policy signals suggest the direction is toward requiring disclosure. Enacted, notified rules with specific compliance obligations have not been confirmed in final form. Lawyers should follow the direction of travel, apply the underlying principles now, and monitor for formal notification. ### What does "AI disclosure" mean practically for a lawyer in court? At minimum, it means being able to tell the court that AI was used in preparing a submission and to confirm that every citation and legal proposition in that submission has been independently verified by the lawyer. It does not mean attaching a printout of your AI session. It means being able to stand behind every word as your own professional judgment. ### Does using AI for legal research breach professional conduct rules? Not inherently. The consensus across jurisdictions - and the reported direction in India - is that AI use is permissible provided the lawyer verifies outputs, takes responsibility for accuracy, and does not allow AI to substitute for professional judgment. The breach arises when unverified AI output is submitted to a court or relied on without checking. ### What happened in the Mata v. Avianca case? In this 2023 case from the US Southern District of New York, a lawyer filed a brief citing multiple cases generated by ChatGPT that did not exist. When the court demanded copies of the judgments, they could not be produced. The court imposed sanctions and made pointed observations about professional responsibility in the AI context. The case is now the standard reference point in global discussions of AI risk in legal practice. ### Can a general-purpose tool like ChatGPT or Claude be used safely for legal research in India? These tools can support drafting, summarising, and structuring. They are not safe to use as primary legal research engines when citations will be submitted to a court, because they generate citations by pattern rather than retrieving them from a real database. You cannot independently verify a citation that was generated rather than retrieved. In a disclosure environment, that makes unverified chatbot output professionally risky. ### What is retrieval-augmented generation and why does it matter for legal AI? Retrieval-augmented generation (RAG) is a technical approach where the AI first searches a real database - in Niyam's case, a corpus of Indian judgments - and then generates a response grounded in what it actually found. This is structurally different from a pure language model generating text by prediction. RAG produces traceable, verifiable answers rather than statistically plausible but potentially invented ones. ### What is a citator and why do I need one when using AI for research? A citator tracks the subsequent history of a judgment - whether it has been followed, distinguished, overruled, or disapproved by later courts. A case that was good law when decided may have been overruled since. An AI that retrieves a judgment without checking its current status may give you a case that no longer represents the law. A citator closes that gap. Niyam includes citation status checking as part of the research workflow. ### How does the DPDP Act affect my use of AI tools in legal practice? The Digital Personal Data Protection Act creates obligations around how personal data is processed. When you describe a client's facts to an AI tool, you may be processing personal data. If the tool retains that data, shares it with third parties, or uses it for model training, there are DPDP implications. Lawyers should use tools that clearly limit data retention and processing to what is necessary to return search results, and that do not use professional queries for model training. ### Is client confidentiality at risk when using AI for legal research? It can be, depending on the tool. General-purpose AI services commonly retain and review inputs. If you describe the details of a client dispute - names, facts, commercial terms - to such a tool, that information has left your control. Purpose-built legal AI tools designed for professional use should have explicit policies limiting retention and prohibiting use of inputs for training. Check the privacy policy before using any tool with client information. ### What does "human-in-the-loop" mean in the context of legal AI? It means a qualified human - in legal practice, the lawyer - reviews and takes responsibility for every AI output before it has real-world effect. Specifically: you read the cited judgments, verify the propositions, and make your own professional judgment about what to rely on. You do not file AI output without reading it. The AI assists; the lawyer decides. ### Can AI make judicial decisions in Indian courts? No. Judicial decisions must be made by appointed judges exercising the judicial power vested in them by the Constitution. This is not a matter of policy preference - it is a constitutional requirement. AI can assist with research, drafting, and case management, but the decision itself must be a human judicial act. ### What features should I look for in a legally compliant AI research tool? Look for: (1) retrieval from a named, verifiable corpus of real judgments; (2) citations surfaced alongside answers with links to source material; (3) citator functionality confirming cases are still good law; (4) clear data handling policies limiting retention and prohibiting training use of your queries; (5) outputs presented as research support requiring lawyer verification, not as final legal conclusions. ### Why does Niyam search 72,000+ judgments specifically? That figure represents the depth of the Supreme Court corpus from the court's digital records - a comprehensive body of decisions that allows Niyam to answer questions about constitutional law, commercial law, criminal procedure, and every other domain the Supreme Court has addressed. Breadth of the corpus matters because a narrower database produces more gaps, and gaps produce either wrong answers or admitted ignorance. Both are better than hallucination but only breadth gives you confidence. ### How is Niyam different from a keyword search on Indian Kanoon? Indian Kanoon is a keyword search engine across a large judgment database. Niyam uses semantic search - you can ask a question in plain language and the system understands the legal concept you are asking about, not just the words you used. The retrieval is then combined with a language model that synthesises the relevant passages and surfaces the most applicable authorities, rather than returning a list of raw results that you have to read and rank yourself. ### What should I do before submitting an AI-assisted submission to court? Verify every citation independently: confirm the case exists in an authoritative database, read the relevant passage yourself, confirm the proposition accurately represents what the court held, and run a citator check to confirm the case is still good law and has not been overruled. Sign off on the submission as your own professional work, not as AI output. Be prepared to explain your research process if asked. ### Will disclosure rules make AI tools more expensive to run for legal teams? Compliance-oriented features (retrieval over generation, citator integration, audit trails, privacy-preserving data handling) require more infrastructure than a raw language model. But the cost of using a non-compliant tool - sanctions, professional conduct proceedings, reputational damage - vastly exceeds any price difference between tool tiers. The relevant comparison is not compliant tool versus cheap tool. It is compliant tool versus professional liability. ### Do solo practitioners and small firms need to worry about AI disclosure, or is this only for large firms? Every lawyer who submits a document to a court carries the same professional obligations regardless of firm size. A sole practitioner whose AI-assisted submission relies on a fabricated citation faces the same sanctions exposure as a partner at a large firm. Firm size is irrelevant to professional conduct obligations. ### How do I explain AI use to a client who is unfamiliar with the technology? Tell them that AI tools are used to search a large database of court decisions for relevant precedent, in the same way a junior researcher would search a law library. Every result is reviewed and verified by the lawyer before being relied on. The AI assists with finding material; the legal judgment about what to use and how to argue it is yours. ### Are the DPDP rules already in force for legal AI tools? The Digital Personal Data Protection Act received Presidential assent in 2023. Implementation rules and specific sectoral guidance continue to be developed. The underlying obligation to handle personal data lawfully and with appropriate safeguards is in force. Lawyers and legal tech providers should be operating on DPDP-consistent principles now rather than waiting for sector-specific guidance. ### Where can I read more about evaluating AI tools for Indian legal practice? Start with our guides on [AI legal research in India](/blog/ai-legal-research-india), [how to vet legal AI citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy), [good law checking](/blog/good-law-checking), [ChatGPT for lawyers in India](/blog/chatgpt-for-lawyers-india), and [best AI tools for lawyers in India](/blog/best-ai-tools-for-lawyers-india). Our [comparison page](/compare) sets out how purpose-built legal AI differs from general-purpose tools on the dimensions that matter for professional use. --- ## The decision you need to make now The direction of travel is clear. Indian courts and regulators are moving toward a framework where AI use in legal practice is permissible but must be disclosed, verified, and traceable to real sources. The lawyers who are positioned well for that environment are not the ones who avoid AI. They are the ones using AI that was built to meet exactly these requirements. Choosing a tool now, before formal rules are notified, is not premature. It is rational preparation. A tool that makes your research traceable, keeps your client data private, and presents outputs as drafts requiring your professional judgment is not just compliance-ready. It is also better research practice, regardless of what any regulation ultimately requires. Niyam is built for Indian legal practice on Indian law, with a corpus of 72,000+ judgments, a citator that checks whether cases are still good law, and data handling designed around your professional obligations. The ₹100 trial gives you 200 credits to run real research queries and see for yourself how sourced, traceable answers compare to what you get from a general-purpose chatbot. [Start for ₹100](https://app.niyam.ai/register) Questions about whether Niyam fits your practice? Write to hello@niyam.ai.