# A lawyer's duty to verify AI legal output in India # A lawyer's duty to verify AI legal output in India **TL;DR:** AI tools can speed up legal research, but the lawyer who signs the brief owns every citation and argument in it. Indian professional conduct rules have not carved out an exception for software. An AI that produces a plausible but non-existent judgment is not a defence - it is an embarrassment, and potentially a disciplinary matter. This post sets out why the duty exists, where AI tends to fail, what the Supreme Court's draft guidelines expect, and a five-step verification workflow that takes under ten minutes per citation. --- ## On this page - [Why this is not an abstract concern](#why-this-is-not-an-abstract-concern) - [How AI generates text and why hallucinations are inevitable](#how-ai-generates-text-and-why-hallucinations-are-inevitable) - [The professional duties that already apply](#the-professional-duties-that-already-apply) - [What the Supreme Court draft guidelines say](#what-the-supreme-court-draft-guidelines-say) - [The five-step citation verification workflow](#the-five-step-citation-verification-workflow) - [Confidentiality and DPDP when you upload client data](#confidentiality-and-dpdp-when-you-upload-client-data) - [Where a grounded legal AI helps and where it does not](#where-a-grounded-legal-ai-helps-and-where-it-does-not) - [Common verification mistakes and how to avoid them](#common-verification-mistakes-and-how-to-avoid-them) - [Checking whether a judgment is still good law](#checking-whether-a-judgment-is-still-good-law) - [The bench and date problem](#the-bench-and-date-problem) - [Frequently asked questions](#frequently-asked-questions) - [The bottom line](#the-bottom-line) --- ## Why this is not an abstract concern In 2023, lawyers in a US case called *Mata v. Avianca* submitted a brief to the Southern District of New York that cited several precedents. Some of those cases did not exist. The citations were fabricated by ChatGPT, and the lawyers who filed the brief had not checked them. The court sanctioned the attorneys. The incident was embarrassing enough to prompt bar associations across the United States to issue guidance, and it has been cited in courtrooms and law schools on every continent since. The Indian legal market is at a similar inflection point. AI tools are now in use across chambers in Delhi, Mumbai, Chennai, and Bangalore. Some lawyers use general purpose chatbots. Others use purpose-built legal AI. Either way, the fundamental risk is identical: an AI can produce confident, well-formatted output that refers to a judgment that was never delivered, by a bench that never sat, on a date that never happened. No bar authority in India has yet issued specific AI guidance as of mid-2026. The Bar Council of India has not published regulations on the use of AI in legal practice. That absence of specific regulation does not create a safe harbour. The general duties - competence, candour to the court, and confidentiality - have existed for decades, and they apply regardless of which tools a lawyer uses to prepare a submission. If you cite a judgment that does not exist, the fact that an AI gave it to you will not impress the court. ## How AI generates text and why hallucinations are inevitable To understand why verification is non-negotiable, it helps to know what AI tools actually do when they answer a legal question. Large language models work by predicting the next token (roughly, the next word or piece of a word) in a sequence, based on patterns in the text they were trained on. They are not databases. They do not retrieve a stored record of a judgment and quote it back to you. They reconstruct what a plausible answer might look like, drawing on statistical patterns from millions of documents. This is genuinely useful for many tasks. It means these models can summarise arguments, identify relevant legal principles, draft boilerplate, and explain complex statutes in plain language. But it also means they can produce text that sounds like a real citation without any actual judgment behind it. The model has learned that a legal argument often looks like: *proposition X, as held by the court in [case name], [year], [court]*. When it generates that pattern, it can fill in a plausible-sounding case name that does not correspond to anything in the official reports. Worse, the model will often do this confidently. It will give you a complete citation with a year, a bench, and a short description of what the judgment held. The description will sound reasonable. The citation style will look correct. None of that means the case exists. This is not a bug that will be fixed in the next version. It is a structural property of how these models work. Even retrieval-augmented systems - those that pull from a real database before generating an answer - can misquote, misattribute, or mischaracterise what they retrieve. The retrieval helps, but it does not make the output self-certifying. The practical consequence is simple: every citation an AI gives you needs to be checked before it goes into a submission. ## The professional duties that already apply Indian advocates are governed by the Advocates Act 1961 and the Bar Council of India Rules framed under it. Three duties matter most in the AI context. **Competence.** An advocate must have the legal knowledge and skill necessary to handle a matter. This does not mean knowing everything - it means knowing enough to serve the client properly, and knowing when to seek assistance or verify information. Using an AI tool without understanding its limitations, and without checking its output, is arguably an exercise of incompetent practice. The tool does not possess legal knowledge in any meaningful sense; it produces plausible text. A lawyer who relies on that text without verification has effectively delegated judgment to a statistical model. **Candour to the court.** The Bar Council Rules prohibit an advocate from knowingly misleading a court. Submitting a non-existent judgment plainly misleads the court. The more difficult question is whether submitting an AI-generated citation that the lawyer has not verified, but believes to be correct, constitutes a knowing misrepresentation. Courts are unlikely to be sympathetic to the argument that a lawyer did not know because they trusted their software. Counsel is expected to know what is in the brief they sign. **Confidentiality.** Client information is privileged. An advocate must not disclose confidences or use them adversely. When a lawyer uploads client documents to an AI platform, they are sharing that information with a third-party system. Whether that constitutes a breach depends on how the platform handles data - whether it trains on user input, where it stores it, who can access it. This is addressed separately below in the section on DPDP, but the duty exists independently of any statute. None of these duties are new. What is new is that lawyers now have access to tools that can produce authoritative-looking legal output very quickly, creating a temptation to skip the verification steps that have always been part of competent practice. ## What the Supreme Court draft guidelines say As of mid-2026, the Supreme Court of India has circulated draft guidelines on the use of AI in courts and legal practice. These are draft guidelines - not a final order, not a statutory instrument - and they may be revised or replaced before any formal adoption. Practitioners should treat them as indicative of the direction in which institutional thinking is moving, not as settled rules. With that caveat, the draft guidelines are clear on one point: AI tools are assistive. They are not a substitute for human legal judgment, and responsibility for submissions made to a court remains with the advocate who makes them. The guidelines contemplate that courts may take a dim view of submissions that rely on AI output that has not been independently verified. The draft also gestures at the distinction between grounded AI systems - those that draw on real, verified legal databases - and general-purpose language models. The implication is that the former carry lower (though not zero) risk than the latter. This is a sensible distinction, but it does not change the verification duty. Even a grounded system can misquote or mischaracterise what it retrieves. If you want to follow the spirit of where Indian court practice appears to be heading, the position is: use AI to help you work, verify everything material before it goes to a court, and be able to demonstrate that you did so if asked. For more on how these draft guidelines affect the tools you use, see our post on [what the Supreme Court AI rules mean for legal tools](/blog/supreme-court-ai-rules-what-tools-need) and [Supreme Court AI rules in India explained](/blog/supreme-court-ai-rules-india). ## The five-step citation verification workflow This is the practical core of the post. For every citation you intend to use in a submission, run through these five steps. On a well-organised research workflow, this takes five to ten minutes per citation - less if the citation is from a grounded source that already links to the original text. **Step 1 - Check that the judgment exists.** Pull up the citation in an official or authoritative source. For Supreme Court judgments, this means the official Supreme Court website (main.sci.gov.in), SCC Online, or Manupatra. For High Court judgments, the relevant High Court website. Look for the exact case name and year. If you cannot find the case, do not use it. This sounds obvious, but it is the step that was skipped in *Mata v. Avianca*. The lawyers trusted the output. They did not run the citation. **Step 2 - Read the actual judgment.** Find the paragraph or passage the AI claims supports the proposition you want to advance. Read it. Check that the judgment actually says what the AI summary says it says. AI tools - even good ones - can mischaracterise what a judgment holds. A case cited for proposition X may actually stand for a narrower version of X, or may have significant caveats that the summary omits. Courts read judgments. If your submission rests on a mischaracterisation, opposing counsel will find it. **Step 3 - Confirm that the judgment says what you claimed.** This is slightly different from Step 2. After reading the passage, ask: if I quote this judgment in my submission for proposition X, is that an accurate representation of what the court held? Sometimes a judgment is real and the passage is real, but the AI has characterised it in a way that overstates its holding. This matters. Courts do not look kindly on submissions that quote out of context. **Step 4 - Run a good-law check.** A judgment that was good law in 2018 may have been overruled, distinguished, or substantially modified since. Check whether the judgment you are citing is still good law. Has it been appealed? Has the Supreme Court subsequently taken a different view? Has the court that delivered it subsequently distinguished it on facts similar to your own? For more on how to do this properly, see our post on [good law checking in Indian legal research](/blog/good-law-checking) and [how to vet legal AI citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy). **Step 5 - Verify bench composition and date.** This is the step most lawyers skip, and it can matter. A judgment from a three-judge bench of the Supreme Court carries more weight in certain contexts than one from a two-judge bench. A constitution bench judgment is binding in a different way than a division bench judgment. Check that the bench size and composition the AI gives you is correct. Check the date. These details affect how you should frame the citation in your submission, and getting them wrong is the kind of error that makes a court look at you sideways. ## Confidentiality and DPDP when you upload client data A practical issue that does not get enough attention: when you upload client documents to an AI platform to get help with research or drafting, you are sharing those documents with a third party. What happens to them depends entirely on the platform. Some AI tools - particularly general-purpose consumer tools - train on user inputs. Your client's contract, your private advice, the sensitive facts of a dispute: all of that can become training data. Most tools that are positioned for professional use have provisions preventing training on user data, but "we don't train on your data" and "your data is secure" are not the same thing. The Digital Personal Data Protection Act, 2023 (DPDP) adds a statutory dimension to this. DPDP requires data fiduciaries to implement reasonable security safeguards, and creates obligations around how personal data can be used. When your client's documents contain personal data - which they almost always do - you need to be confident that the platform you are using handles that data consistently with your client's reasonable expectations and with DPDP obligations. This does not mean you cannot use AI tools with client data. It means you should check the data handling terms before you do, and you should know what you are agreeing to. For matters involving particularly sensitive information, a tool that processes data locally or within a clearly defined secure environment is worth the extra cost. Niyam is built to be aware of DPDP obligations. If you want to understand how we handle data, the [responsible AI page](/responsible-ai) sets this out directly. ## Where a grounded legal AI helps and where it does not There is a meaningful difference between a general-purpose language model and a legal AI system that retrieves from a real, verified corpus before generating an answer. A general-purpose chatbot has no legal database. When you ask it about Indian case law, it is drawing on whatever it absorbed during training - a mixture of legal texts, commentary, Wikipedia, blog posts, and anything else that was in the training set. It has no reliable mechanism for knowing whether a judgment exists, and it cannot check its own output against a source of record. A grounded legal AI retrieves actual documents from a structured database and uses those documents to generate its answer. Niyam, for example, works from a corpus of 72,000+ Indian judgments. When it cites a case, the citation is drawn from that corpus. This is a substantially different and lower-risk model than generation from memory. But it still does not remove the verification duty. The system can still mischaracterise what it retrieves. The lawyer still needs to read the judgment. The practical difference is that with a grounded tool, Steps 1 and 2 of the verification workflow are significantly faster. The tool provides a link to the actual judgment. You can read the passage directly. You are verifying characterisation, not existence. For a detailed comparison of native legal AI versus general purpose tools, see [our post on native legal AI for India versus generic GPT](/blog/native-legal-ai-india-vs-generic-gpt). Niyam also has a built-in citator that lets you check whether a judgment remains good law - which is Step 4 in the workflow above. This makes the verification process faster. It does not make it unnecessary. See [how the Niyam citator works](/solutions/citator). For an honest comparison of how Niyam differs from general chatbots, see the [comparison page](/compare/chatgpt-general-ai). ## Common verification mistakes and how to avoid them A few patterns come up repeatedly when lawyers talk about close calls with AI-generated citations. **Trusting the summary, not reading the judgment.** The AI gives a confident two-sentence summary of what a judgment holds. The lawyer checks that the case exists, sees the correct citation, and uses the summary in the submission. The problem: the summary was accurate for one part of the judgment but missed a significant qualification in the next paragraph. Read the judgment. Not just the headnote, not just the AI summary - the actual passage. **Checking existence but not characterisation.** This is the subtler version of the same error. The judgment is real. The passage exists. But the AI has framed the holding more broadly than the court actually framed it. Courts notice. Opposing counsel notices. **Not distinguishing between obiter and ratio.** An AI will often cite a judgment for a proposition that appears in obiter dicta - observations the court made in passing, which do not form part of the binding ratio decidendi. These can be useful in argument, but they carry a different weight, and representing them as the court's holding is inaccurate. Check what part of the judgment you are relying on. **Using judgments without checking subsequent treatment.** A High Court judgment from 2019 may have been overruled by the Supreme Court in 2022. If you cite the 2019 judgment without noting the 2022 development, you are not doing your client any favours. The other side will find it. **Uploading the wrong data to the wrong tool.** This is not a citation error but it is a professional risk. Know what you are uploading and where. See also the DPDP section above and our post on [AI hallucinated citations in India](/blog/ai-hallucinated-citations-india). ## Checking whether a judgment is still good law Good-law checking deserves its own section because it is the step that is most easily skipped and most often consequential. A judgment can be good law in several senses and not good law in others. A Supreme Court judgment may be the binding authority on a general principle but subsequently distinguished on facts that are similar to yours. A High Court judgment may remain good law in that jurisdiction but have been decided differently by another High Court, making the position unsettled. A Full Bench judgment may have been referred to a larger bench, leaving the law in a state of suspense. None of this is visible from a citation alone. You need to check the subsequent treatment of the judgment. Practically, this means: find the judgment in a proper legal database and check what has cited it and how. Has it been followed? Distinguished? Overruled? Doubted? The answer changes how you use it in your submission. This used to require significant time with a subscription database. A citator built into a legal AI platform can do this faster - see [Niyam's citator](/solutions/citator). But the output of a citator is still something you should read and assess, not just accept. ## The bench and date problem Two details that AI tools frequently get wrong are bench composition and the precise date of a judgment. Bench composition matters because it affects precedential weight. In the Supreme Court of India, a judgment of a larger bench prevails over a judgment of a smaller bench on the same point. If you cite a two-judge bench judgment for a proposition and the other side has a three-judge bench judgment taking a different view, you have a problem. An AI tool asked to state the bench size may estimate from context rather than retrieve the actual information. Check. Date matters for several reasons. The chronology of judgments on a point matters when you are tracing how the law developed. The date determines which version of a statute was in force when the judgment was delivered. And, sometimes, date affects jurisdiction: a judgment delivered before a court's territorial jurisdiction changed, or before a statutory amendment, may need to be read in that context. Neither of these checks takes long when you have the original text in front of you. The problem is that lawyers who rely entirely on AI summaries often do not pull up the original text at all. ## Frequently asked questions ### Is it professionally wrong to use AI for legal research in India? No. Using AI tools for legal research is not in itself a breach of any professional duty. The Bar Council of India has not prohibited their use, and as of mid-2026, there is no specific guidance restricting which tools an advocate may use for research or drafting. The duty that applies is the duty of competence - you must produce work that is accurate and serves your client properly. How you get there is not regulated. What matters is whether the product is reliable. ### Does the Bar Council of India have specific AI regulations? Not as of mid-2026. The Bar Council has not issued specific regulations or guidelines addressing the use of AI in legal practice. The existing rules under the Advocates Act 1961 and the Bar Council of India Rules - governing competence, candour, and confidentiality - apply by their terms to the conduct of an advocate regardless of the tools used. It would be surprising if specific AI guidance were not issued at some point, given the pace of adoption, but nothing exists yet. ### What happened in the Mata v. Avianca case and could something similar happen in India? In 2023, two lawyers filed a brief in the United States District Court for the Southern District of New York that cited several cases which did not exist. They had used ChatGPT to research the matter and submitted the AI's output without verifying the citations. When the court asked them to produce the cases, they could not. The court sanctioned them. There is nothing in the Indian procedural or disciplinary framework that would prevent an analogous outcome. Courts have inherent powers to address misconduct before them, and the Bar Council can take disciplinary action for breaches of professional conduct rules. A lawyer who submits a non-existent citation in an Indian court faces real risk. ### If I use a legal AI that retrieves from a real database, do I still need to verify? Yes. A retrieval-grounded system is substantially safer than a general-purpose chatbot, because the citations it produces are drawn from real documents. But the system can still mischaracterise what it retrieves. The AI reads a passage and summarises its holding - that summary can be imprecise or incomplete. The only way to know whether the summary is accurate is to read the original passage yourself. Retrieval narrows the verification task; it does not eliminate it. ### How long should citation verification take? For a single citation in a well-organised workflow, five to ten minutes. For a complex brief relying on twenty citations, allow a dedicated verification pass of a couple of hours. This is not materially different from what a diligent advocate would have spent pulling print reports, and the speed of the research phase more than compensates. The issue is not that verification takes too long; it is that the speed of AI output creates a temptation to skip it. ### What counts as a "good-law check" in Indian practice? A good-law check means confirming that the judgment you are citing has not been overruled, that it has not been limited to facts significantly different from yours, and that the principle you are relying on has not been substantially modified by subsequent decisions. In practice, this means checking the subsequent judicial treatment of the case - who has cited it, and how. A citator tool can surface this information; you then need to read the relevant subsequent decisions yourself to understand their effect. ### Can I use a general-purpose AI like ChatGPT or Gemini for Indian legal research? You can, but the risks are higher than with a purpose-built legal AI. General-purpose models have no Indian legal database. They are generating responses based on whatever they absorbed in training, which includes varying amounts of Indian legal material but with no guarantees of accuracy, completeness, or currency. Citations from a general-purpose model should be treated as starting points for research, not as verified authorities. You will need to check every citation from first principles. For a more detailed comparison, see [ChatGPT versus native legal AI for India](/compare/chatgpt-general-ai). ### What specific duties under the Advocates Act apply to AI use? The Act itself does not mention AI. The relevant duties come from the Bar Council of India Rules framed under Section 49 of the Act. Rule 11 prohibits an advocate from misleading the court. Rule 22 requires that an advocate not be guilty of conduct unbecoming of an advocate. Separately, the duty of competence is not stated in a single rule but runs through the framework - an advocate is expected to have the knowledge and skill to handle the matter they have taken on. None of these rules turn on whether you used a software tool, and none of them would be satisfied by an argument that the tool was at fault. ### What should I do if I realise I cited a non-existent case? Act immediately. Inform the court at the earliest opportunity. If the matter is before a court, you have a duty of candour that requires you to correct the record. Concealing the error is significantly worse than the error itself. Draft a formal correction, explain what happened, and let the court decide how to proceed. Courts are not uniformly unsympathetic to honest mistakes corrected promptly; they are unsympathetic to discovered concealment. ### How does DPDP affect which AI tools I can use for client matters? The Digital Personal Data Protection Act, 2023 requires that personal data be handled securely and used only for the purposes for which it was collected. When you upload client documents containing personal data to an AI platform, you are sharing that data with a third-party data fiduciary. You should check whether the platform's terms of service permit the use of that data for training, what security measures are in place, and whether the platform stores data in India or transfers it abroad. For highly sensitive matters, a tool that processes data within a defined secure environment reduces risk. The DPDP obligations are on you as the person responsible for the data, not on the software vendor. ### Is there a difference between using AI for research and using it for drafting? Practically, yes. Research outputs carry the specific risk of fabricated citations, which is what this post is mainly about. Drafting outputs carry different risks: the AI may produce arguments that are legally inaccurate, miss key elements of a statutory provision, use the wrong standard of proof, or draft clauses that are unenforceable in an Indian jurisdiction. Both types of output require review and verification, but the nature of what you are verifying is different. For drafting, the question is not "does this citation exist" but "is this legally correct, complete, and appropriate for my client's situation." ### Can AI help me verify citations, or does it introduce circular risk? Do not use the same AI tool that produced a citation to verify it. The tool will tend to confirm what it generated. Use an independent source - the official court website, a separate legal database, or a purpose-built citator. The only reliable verification of a citation is checking it against the original judgment in a trustworthy source. If a legal AI platform provides a citator function backed by a verified corpus, that is appropriate to use for good-law checks, but even then, read the underlying cases yourself for anything that matters. ### What if a client asks me whether I used AI to prepare their matter? Answer honestly. A client who has engaged you for legal advice is entitled to know how you prepared that advice, particularly if the method of preparation could affect the reliability of the output. If you used AI tools and verified the output through the process described in this post, you can describe that with confidence. If you used AI tools without verification, that is a harder conversation to have - and the right response at that point is to go back and do the verification before the matter progresses further. ### What should I look for in a legal AI tool to reduce verification risk? Four things matter most. First, the tool should draw from a real, verified legal database - not generate citations from training memory. Second, it should cite its sources and link you to the original text, so you can verify in one step rather than running a separate search. Third, it should have a built-in good-law check function (a citator), so you can check the subsequent treatment of a case without leaving the platform. Fourth, its data handling should be consistent with your confidentiality obligations - check whether it trains on user input and how it stores data. See [how Niyam approaches each of these](/solutions/research). ### How do I explain AI use to a court if asked? If a court asks about your research methodology, you should be able to explain what tools you used, what they did, and what independent verification you performed. The answer "I used an AI tool and then verified each citation against the original judgment in [source]" is a complete and defensible answer. The answer "I relied on the AI's output" is not. Courts are not hostile to AI use per se; they are hostile to submissions that lack the ordinary verification a competent advocate would perform. ### Does the verification duty change for different types of proceedings? The core duty is the same across proceedings - you must not mislead a court and you must be competent. The stakes and practical consequences vary. An error in a Supreme Court submission before a constitution bench is likely to be treated differently from an error in a routine district court application. But the professional duty is identical. Apply the same verification standard regardless of where the matter sits. ### How often do legal AI tools produce wrong citations? This is genuinely hard to quantify, and any specific number would be unreliable without a controlled study. General-purpose chatbots produce wrong citations often enough that treating every citation as unverified is the only safe approach. Retrieval-grounded legal AI tools produce wrong citations much less frequently, because they draw from real documents - but mischaracterisation of correct documents is still possible. Our own assessment at Niyam is that you should verify every citation you intend to use in a submission regardless of source, because the cost of a wrong citation is high relative to the time cost of verification. ### What is the most common type of AI citation error in Indian legal research? From what practitioners report, the most common error with general-purpose AI is a citation that looks real - correct court, plausible year, reasonable case name - but does not correspond to any actual judgment. The second most common is a real case being characterised for a proposition it does not actually support. Both errors are caught by the same process: find the judgment in an authoritative source and read the relevant passage yourself. ### Is there a way to use AI tools responsibly while still getting the time savings they offer? Yes, and this is the practical point of the post. The time savings from AI come mainly at the research phase - identifying relevant cases, understanding the shape of the law, generating a first-cut draft. The verification step is much faster than the original research would have been without AI, because you already know what to look for. A grounded legal AI that links you directly to the source text makes the verification step faster still. The workflow is not "trust the AI" or "do not use AI" - it is "use AI to narrow the field, then verify what matters." --- ## The bottom line None of this is a reason to avoid AI tools. Used well, they make legal research faster and help surface cases and arguments that might otherwise take days to find. The five-step verification workflow described above is not onerous - it is just the due diligence that competent practice has always required, now applied to a new category of output. The position is straightforward: the tool assists, the lawyer decides, and the submission belongs to the lawyer who signs it. An AI generating a plausible-sounding but non-existent citation is a tool failure. Submitting that citation to a court without checking it is a professional one. If you want a legal AI built on a real Indian judgment corpus - one that cites sources, links to original text, and has a citator built in so you can run Steps 1 and 4 of the above workflow within the same tool - that is what Niyam is. It does not remove your verification duty, and we have never claimed it does. It makes verification faster. When you are ready to try it: [Start for ₹100](https://app.niyam.ai/register) - 200 credits to start, cancel anytime. Questions: hello@niyam.ai.