# Supreme Court AI rules for courts: what the draft says # Supreme Court AI rules for courts: what the draft says **TL;DR:** The Supreme Court of India has circulated a draft framework governing how artificial intelligence may - and may not - be used in judicial proceedings. The draft draws a firm line between assistive AI (legal research, translation, transcription, drafting support - all permitted with safeguards) and decisional AI (AI pronouncing judgment, imposing sentence, or substituting for judicial reasoning - all barred). For practitioners, the most consequential obligation is verification: any AI output used in court must be checked and owned by the human lawyer presenting it. --- ## On this page - [What prompted this framework](#what-prompted-this-framework) - [The core distinction: assistive versus decisional AI](#the-core-distinction-assistive-versus-decisional-ai) - [What the draft reportedly permits](#what-the-draft-reportedly-permits) - [What the draft reportedly bars](#what-the-draft-reportedly-bars) - [The verification duty: what it actually requires](#the-verification-duty-what-it-actually-requires) - [Transparency and disclosure obligations](#transparency-and-disclosure-obligations) - [Data confidentiality and the handling of case information](#data-confidentiality-and-the-handling-of-case-information) - [What these rules signal for legal-AI tooling](#what-these-rules-signal-for-legal-ai-tooling) - [The assistive AI that does comply: a field guide](#the-assistive-ai-that-does-comply-a-field-guide) - [What practitioners should do right now](#what-practitioners-should-do-right-now) - [How this fits with Niyam's approach](#how-this-fits-with-niyams-approach) - [Frequently asked questions](#frequently-asked-questions) - [Start with a compliant workflow](#start-with-a-compliant-workflow) --- ## What prompted this framework The Supreme Court's move to formalise AI governance in judicial proceedings did not happen in a vacuum. Across several common-law jurisdictions, courts have been forced to grapple with AI-generated material appearing in filings - sometimes disclosed, often not - and the consequences when that material turned out to be fabricated or misleading. In the United States, courts have already imposed sanctions on lawyers who submitted briefs containing AI-generated citations to cases that did not exist. The material looked authentic. The citations were formatted correctly, the names were plausible, and the supposed holdings were coherent. The only problem was that none of the cases had been decided. Judges who went to verify them found nothing. India has not been insulated from similar incidents. Reports have emerged of AI-generated content appearing in documents before Indian courts, including instances where cited judgments could not be located in standard databases. The Bar Council of India and various High Courts have issued notices and informal guidance, but no systematic, court-level framework had been articulated until the Supreme Court began work on this draft. There is also the broader institutional picture. The judiciary's SUPACE (Supreme Court Portal for Assistance in Courts Efficiency) project demonstrated that the court has been thinking seriously about AI integration for years. That project used AI to analyse case materials and flag relevant precedents to assist judges - not to decide cases, but to surface information faster. The draft AI framework can be read as the governance layer that makes responsible expansion of such tools possible. The timing matters too. Indian courts are processing a docket that routinely exceeds any reasonable institutional capacity. Pendency figures in the millions are the ongoing reality. Anything that can responsibly compress the time a judge or clerk spends locating relevant authority is worth examining. The draft reflects a court that sees AI as a possible partial answer to pendency, while remaining acutely aware of the constitutional stakes if AI begins to replace, rather than support, judicial reasoning. --- ## The core distinction: assistive versus decisional AI The reported framework rests on one foundational distinction. Understand it and everything else follows. **Assistive AI** does work that a paralegal, research assistant, or court clerk might otherwise do: locating judgments, translating documents into the language of the court, transcribing oral proceedings, and helping draft orders or arguments in structured form. The human - lawyer or judge - reviews the output, makes the decisions, and takes professional responsibility for what goes forward. **Decisional AI** is something categorically different. It would mean delegating the exercise of judicial power to an algorithm: letting a machine determine guilt or innocence, impose a sentence, award damages, or craft the legal reasoning that binds the parties and becomes precedent. That is what the draft bars, and the reasoning behind the bar is not merely prudential but constitutional. Article 21 of the Constitution - the right to life and personal liberty - has been interpreted by the Supreme Court over decades to include the right to fair hearing, reasoned adjudication, and access to a court that exercises human judgment. The right to appeal a judicial decision is premised on the existence of a human judicial mind that made a decision which can be reviewed. An AI decision is not subject to the same scrutiny. The chain of constitutional accountability runs from litigant to court to appellate bench to the constitutional texts. Inserting an opaque algorithmic process into that chain breaks the accountability link in a way the Constitution cannot accommodate. This is not a technical objection - it is a structural one. Even if AI were demonstrably accurate in 99% of cases, the 1% of errors would not be correctable through the mechanisms the Constitution provides. You cannot cross-examine an algorithm. You cannot ask it to explain its reasoning in terms that can be subjected to legal argument. You cannot appeal to a higher court on the ground that the algorithm applied the wrong legal test, because the algorithm does not apply legal tests - it predicts outputs. The draft is reported to take this position clearly: AI in courts is a tool for the judicial mind, never a replacement for it. --- ## What the draft reportedly permits Based on reporting and commentary available as of mid-2026, the draft framework is understood to allow the following categories of AI use, each with attendant conditions. ### Legal research assistance AI may be used to search for, summarise, and surface relevant judgments and statutory provisions. The operative word is "surface": the tool locates candidates; the lawyer or judge evaluates them. Tools that search across a corpus of Indian judgments and return citations with passages can be used in preparation and in court, provided the user can verify the output. This is precisely what retrieval-grounded tools do - and why the distinction between a general-purpose chatbot and a purpose-built legal AI matters enormously in this context. For more on what this means for tool selection, see [how to vet legal AI citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy) and the broader discussion in [AI legal research India](/blog/ai-legal-research-india). ### Translation India's multilingual reality means that parties, witnesses, and documents routinely arrive in languages other than the language of the court. AI-assisted translation of documentary evidence and pleadings is permitted under the draft, subject to human review where the translation is being relied upon in proceedings. The framework reportedly distinguishes between AI translation used for internal comprehension (lower bar) and AI translation placed before the court as evidence of what a document says (higher bar requiring verification). ### Transcription Court proceedings generate enormous volumes of text that currently require manual transcription, a resource-intensive process prone to error and delay. AI transcription of oral proceedings is reportedly within the permitted category, again subject to human review before the transcript is finalised and relied upon. ### Drafting assistance This is the category most likely to affect practising lawyers directly. AI may be used to assist in drafting orders, judgments, written arguments, and pleadings - but the draft is reported to be emphatic that assistance means assistance. The final product must represent the genuine work of the human drafter. An AI-drafted document that has been read, assessed, corrected, and signed off is compliant. An AI-drafted document submitted without meaningful human engagement is not. The line between these two is not always comfortable to draw, but the framework reportedly provides a useful test: if the drafter cannot explain and defend every substantive choice in the document, the level of AI reliance has exceeded what is permitted. --- ## What the draft reportedly bars The prohibitions are, if anything, more clearly articulated than the permissions. ### AI as decision-maker No AI system may issue a judicial decision. This covers final orders, interim orders, procedural rulings, or any other act that constitutes the exercise of judicial power. The draft is reported to make no distinction based on the severity or complexity of the matter: a traffic fine and a capital sentence are equally within the bar. ### AI sentencing recommendations without disclosure and human override The use of AI-generated risk scores or sentencing recommendations - a practice that has generated substantial controversy in the United States through tools like COMPAS - is either barred outright or subject to conditions that make undisclosed use impermissible. The concern is that algorithmic sentencing recommendations can encode historical biases in ways that are not visible to the parties and cannot be effectively challenged. ### AI substituting for judicial reasoning This is the most conceptually important prohibition. It covers the scenario where a judge uses AI to generate the reasons for a decision and adopts them wholesale without genuine independent reasoning. The output might be grammatically correct, legally formatted, and cite real cases. But if the judge did not independently reason to the conclusion, the judgment is not a judicial act in the constitutional sense. Indian courts have long held - and the Supreme Court has repeatedly affirmed - that a speaking order, one that gives reasons, is not just a formal requirement but a constitutional guarantee. Reasons allow appellate review. Reasons show the parties that their arguments were heard. Reasons create precedent that can be distinguished or followed. AI-generated reasons that the judge did not independently reach hollow out this guarantee. ### Undisclosed use Perhaps the most practically significant prohibition: any use of AI in the preparation of material placed before the court must be disclosed. The draft reportedly requires disclosure both in pleadings and by judicial officers in administrative communications about their decision-making process. The consequences of undisclosed use are treated as equivalent to the consequences of misrepresentation. --- ## The verification duty: what it actually requires A disclosure obligation without a verification obligation would be largely empty. The draft reportedly treats the two as linked. A lawyer who discloses that AI was used in research but makes no effort to verify the AI's output has not met the standard. What does verification require in practice? The framework, based on available reporting, seems to adopt something close to the standard a careful senior lawyer would apply to work done by a junior: you do not submit it until you are confident enough to put your name on it and defend it under questioning. For legal research specifically, this maps to a set of concrete steps that practitioners already know, even if they do not always follow them under time pressure: First, every citation that AI surfaces must be confirmed to exist. The case name, year, court, and reporter must correspond to a real decision. This means checking the citation in a reliable database, not asking the AI to confirm its own output. Second, the proposition the citation is said to stand for must match what the judgment actually says. An AI tool might correctly identify a case as relevant but characterise its holding in a way that is subtly or materially wrong. You need to read the relevant passages of the judgment, not just the AI's summary of them. Third, the judgment must still be good law. A Supreme Court decision from 2008 that was overruled in 2019 is not authority for anything, and citing it as if it were could mislead the court. Built-in citator functionality, or a separate citator check, is the only reliable way to catch this. Fourth, the judgment must be binding or persuasive in the forum where you are appearing. A High Court judgment from one state is not binding on a coordinate bench in another state, though it may be persuasive. A Tribunal decision is not binding on a High Court. These distinctions matter. The [lawyer's duty to verify AI output](/blog/lawyer-duty-verify-ai-output) is not created by the draft rules - it already exists in professional conduct norms and flows from the advocate's duties to the court. What the draft does is operationalise and make explicit something that ethical rules already require. For a practical checklist, see [AI in Indian courts](/blog/ai-in-indian-courts). --- ## Transparency and disclosure obligations The disclosure provisions in the draft are worth examining separately because they have structural implications for how law firms and individual practitioners will need to work. The reported approach distinguishes between two levels of disclosure: **Disclosure in pleadings:** Where AI was used in the research, drafting, or analysis underlying a pleading, affidavit, or written submission, this must be stated. The form of the disclosure is not yet finalised in the draft, but it is likely to include a statement identifying the nature of the AI use (research, drafting assistance, translation) and affirming that the output was verified by the signatory advocate. **Disclosure by judicial officers:** Where AI was used by a judge or court officer in preparing an order, draft judgment, or cause list, this is separately required to be disclosed in the document or in accompanying administrative records. The disclosure requirement for judges is reported to be calibrated to the level of AI use - light use of AI-assisted transcription in chambers administration may not require the same disclosure as AI assistance in drafting reasons. The underlying rationale for disclosure is appellate reviewability. If AI contributed materially to a judicial product, the appellate court is entitled to know this in assessing whether the required judicial reasoning was actually performed. If AI contributed to an argument placed before a court, the opposing party is entitled to know this because it affects how the argument should be assessed and challenged. From a practical standpoint, disclosure requirements of this kind will push law firms toward formalising their AI use policies. Ad hoc use of various tools by individual lawyers, with no firm-level record-keeping, becomes legally problematic once disclosure is mandatory. Firms will need to know what tools were used, for which matters, and by whom - partly for compliance, and partly because they will need to be able to make the disclosure accurately. --- ## Data confidentiality and the handling of case information This is the section of the draft that has received comparatively less coverage in legal commentary, but it may be the most consequential for day-to-day practice. When a lawyer uses an AI tool to research or draft material for a client matter, the prompts and documents fed into that tool contain confidential client information. Instructions about the facts of the case. Details of the client's position. Potentially privileged communication about litigation strategy. The moment that information leaves the lawyer's device and is processed by a third-party AI system, questions of professional duty arise. The draft is reported to address this directly, though the exact provisions remain subject to revision. The reported position is: AI tools used in judicial or litigation contexts must not transmit confidential case information to external servers in ways that allow that information to be used for training, shared with third parties, or accessed by the tool provider's personnel. Where a tool does transmit data externally, it must do so under terms that provide the same confidentiality protections as a lawyer's own duty of confidentiality to the client. This is a stringent standard. Many general-purpose AI tools - including some of the best-known chatbot products - do not meet it out of the box. Default settings on several platforms allow user inputs to be used for model improvement. Some platforms grant their personnel access to user conversations for safety monitoring. Even where a lawyer uses these tools with good intentions, the act of inputting client information into a non-compliant system may constitute a breach of professional duty independent of any disclosure obligation. The implication is clear: legal AI tools used in the Indian judicial context will need to offer verifiable data handling commitments. Vague assurances about "enterprise security" will not suffice. The standard the draft is reported to contemplate is closer to what the Digital Personal Data Protection Act, 2023 requires for sensitive personal data - clear purpose limitation, defined retention, no secondary use without consent - applied to the lawyer-client relationship. For tools built with Indian legal practice in mind, data localisation and compliant data handling are not differentiators; they are table stakes. --- ## What these rules signal for legal-AI tooling The draft framework, read carefully, functions as a quality filter on the legal-AI market as much as a regulation on court conduct. Tools that would not survive scrutiny under the draft's requirements are, broadly, the tools that should not be used in legal practice anyway - and the draft makes this explicit rather than leaving it to individual professional judgment. Consider the implied requirements for a compliant tool: | Requirement | General-purpose chatbot | Purpose-built legal AI | |---|---|---| | Retrieval from verified Indian judgment corpus | ✗ | ✓ | | Every answer cited to a real, checkable judgment | ✗ | ✓ | | Built-in citator to check good-law status | ✗ | ✓ | | Data handling that does not train models on client inputs | Varies - often ✗ | ✓ (by design) | | Outputs auditable for disclosure compliance | ✗ | ✓ | | Transparency about what the corpus covers | ✗ | ✓ | | Designed for Indian court hierarchy and citation formats | ✗ | ✓ | This is not a theoretical distinction. A lawyer who runs a client matter query through a general-purpose chatbot is, under the reported draft, potentially in breach of the disclosure requirement (if AI was used and not disclosed), the verification requirement (if the chatbot's output was not independently confirmed), and the data confidentiality requirement (if the chatbot's default settings permit secondary use of inputs). Three independent compliance failures from one tool choice. The rules effectively require lawyers to know what their tools do - not just what they produce. A tool that produces fluent, plausible-looking legal text while hallucinating citations and uploading client instructions to a training corpus is not a legal AI tool. It is a liability with a chat interface. For a fuller treatment of what specific technical features a rules-compliant tool needs, see [what tools need to comply with Supreme Court AI rules](/blog/supreme-court-ai-rules-what-tools-need). --- ## The assistive AI that does comply: a field guide Given the framework's contours, what does compliant AI-assisted practice actually look like? Working through a few scenarios is useful. **Legal research for a bail matter.** A lawyer uses an AI research tool to find Supreme Court and relevant High Court judgments on the bail-rule-jail-exception principle. The tool returns citations drawn from a corpus of 72,000+ Indian judgments, each result showing the exact passage relied upon and the full citation. The lawyer opens two or three of the returned judgments in the primary source, reads the relevant passages, and confirms the holdings match what the tool described. The tool's citator shows all returned cases as current good law. The lawyer then drafts the bail application incorporating those authorities. Disclosure: "AI-assisted legal research was used in preparing this application. All cited authorities have been independently verified by the undersigned advocate." This is compliant. **Translation of a regional-language document.** A client produces a registered sale deed in Telugu. The lawyer uses an AI translation tool to get an initial English rendering for internal comprehension. Before placing the translation before a court or relying on it in a pleading, the lawyer engages a certified translator to verify the AI's output on the substantive terms. The AI translation is used as a starting draft that the certified translator reviews and corrects, reducing the translator's time while maintaining the accuracy standard. Compliant. **Drafting assistance for a writ petition.** A lawyer uses an AI drafting tool to generate an initial skeleton for a writ petition challenging an administrative order under Article 226. The tool produces a structured draft identifying the grounds of challenge, relevant constitutional provisions, and placeholder argument sections. The lawyer rewrites every substantive argument section in her own words, drawing on the research she has done, and adds the specific facts of the matter. The AI's skeleton is a scaffold she has entirely rebuilt, not a document she has signed off on wholesale. Compliant. **What is not compliant.** A lawyer asks a general-purpose AI tool: "Write me a bail application for a case where my client was arrested under BNSS section X on allegations of Y. Include Supreme Court judgments supporting bail." The tool produces a document citing several cases. The lawyer reads it quickly, sees it looks right, and files it. Two of the cited cases do not exist. The lawyer had no idea because he did not check. This is not compliant on any dimension: the tool is not retrieval-grounded, the output was not verified, client facts were fed into a non-compliant platform, and no disclosure was made. --- ## What practitioners should do right now The draft is circulating, not yet in force. But the verification and disclosure obligations it articulates are not new requirements - they restate and make explicit what competent, ethical practice already demands. There is no sensible reason to wait for the draft to become final before adjusting your workflow. **Audit your current AI use.** Which tools are you using? For which tasks? What happens to the information you input? If you do not know the answers to these questions, you are not in a position to make a disclosure that is both accurate and compliant. **Establish a verification protocol.** The five-step verification process - confirm the case exists, read the judgment, confirm it is good law, match the proposition to the holding, confirm jurisdiction - should become habitual. A tool that makes this easier (by returning the source passage alongside the citation, and by running a citator check automatically) compresses the verification time without removing the step. **Think about disclosure language.** Even before mandatory disclosure rules are in force, getting into the habit of noting AI use in your file notes is good practice. When mandatory disclosure arrives, you will have the records to make accurate statements. More practically, drafting a standard disclosure clause now - one you can review and adapt for each matter - means you are not improvising under pressure when the rules kick in. **Review data handling on every tool you use.** Check the terms of service. Check the privacy policy. Where the tool's default is to use inputs for training, check whether there is an opt-out and whether that opt-out is effective. Where you cannot determine what happens to client data, treat that as a red flag. Client confidentiality is not a secondary concern to be addressed after the tool has already proved useful. **Keep reading the primary sources.** AI-assisted research surfaces candidates for your reading list. It is not a substitute for reading the judgments. A practitioner who has not read the cases they cite is professionally exposed, AI rules or not. --- ## How this fits with Niyam's approach Niyam is a legal AI built for Indian practice. The connection between the draft framework and how Niyam works is direct, not incidental. The research module searches across 72,000+ Indian judgments and returns answers cited to the specific passages in the actual judgments - not to the AI's training data memory, but to retrievable documents you can open and verify. This is the retrieval-grounded approach the draft framework implicitly requires. When you get a research result from Niyam, you get a citation you can check, because it was retrieved from a real source document, not generated from statistical pattern-matching. The built-in citator lets you confirm whether a judgment is still good law before you rely on it. This is the citator step in the verification protocol - made faster but not skipped. Drafting, notices, and translation are all assistive functions. Niyam generates material for the lawyer to review, revise, and own. The product is your document, informed by AI assistance you have verified. Data handling is built around DPDP-aware principles. Client matter information does not train the model. The design assumption is that information entered in the context of a legal matter is privileged and must be treated accordingly. For a side-by-side view of how Niyam's approach compares to general AI tools across the criteria the draft raises, see the [compare page](/compare). For a more detailed breakdown of what the draft framework implies for tool architecture, see [what tools need to comply with Supreme Court AI rules](/blog/supreme-court-ai-rules-what-tools-need). For lawyers evaluating AI tools specifically for citation accuracy, [how to vet legal AI citation accuracy](/blog/how-to-vet-legal-ai-citation-accuracy) walks through the criteria. The draft framework is not a barrier to AI adoption in legal practice. It is a statement of what responsible AI adoption looks like. Tools designed from the ground up for Indian legal practice, with retrieval-grounded research, built-in verification, and compliant data handling, fit within these rules naturally. The compliance overhead falls primarily on practitioners who have been using general-purpose tools in ways that were always professionally questionable. --- ## Frequently asked questions ### Is the Supreme Court AI framework already in force? No. As of June 2026, the framework is a draft that has been circulated and reported on. It has not been formally adopted as a binding circular or direction. The provisions described in this post reflect what has been publicly reported about the draft's contents. Until the final version is published and formally issued, its exact terms remain subject to change. ### What is the difference between assistive AI and decisional AI under the draft? Assistive AI helps a human do their job faster and more accurately - research, translation, transcription, drafting support. The human reviews the output and takes professional responsibility for it. Decisional AI replaces the human's judgment - a machine determining the outcome of a case, imposing a sentence, or substituting for judicial reasoning. The draft permits the former and bars the latter. ### Does the draft apply only to judges, or does it cover lawyers as well? Based on available reporting, the framework addresses both. Lawyers are covered by the disclosure and verification obligations when they place AI-assisted material before courts. Judicial officers are covered by separate provisions governing their own use of AI tools in administrative and judicial work. ### What is the disclosure obligation for lawyers under the draft? Where AI has been used in the research, analysis, or drafting of a pleading, written submission, or affidavit, the lawyer is reportedly required to disclose this in the document. The disclosure is expected to state the nature of the AI use and affirm that the output was independently verified by the signatory advocate. ### Can a lawyer use ChatGPT or Gemini for legal research? The draft does not name specific tools. However, the requirements it imposes - retrieval from a verified corpus, citable output, data confidentiality, and verifiability - are not met by general-purpose chatbots as they are typically used. The compliance risk of using a general-purpose chatbot for Indian court work sits with the lawyer. For a detailed comparison, see [Claude vs ChatGPT vs Gemini for Indian legal work](/blog/claude-vs-chatgpt-vs-gemini-legal-india) and the [compare page](/compare). ### What happens if a lawyer does not verify an AI-generated citation? Submitting an unverified citation to a court is professionally dangerous irrespective of AI rules. Indian courts have the power to impose costs, issue show-cause notices, and refer matters to the Bar Council for professional misconduct. The draft's verification obligation makes explicit what courts already expect. A lawyer who submits a fabricated or wrong citation cannot avoid responsibility by pointing to the AI tool that generated it. ### Does the draft affect the use of AI in arbitration or tribunal proceedings? The Supreme Court's framework directly governs court proceedings. Its application to arbitral tribunals and statutory tribunals is not clear from current reporting. Arbitral tribunals set their own procedural rules, and statutory tribunals are governed by their own enabling legislation. That said, the professional conduct obligations of advocates - including verification duties - apply in all forums. ### What is the draft's position on AI-generated evidence? This is a nuanced area. AI-generated analysis of data (for example, pattern analysis offered as expert evidence) sits in a different category from AI-assisted research or drafting. The draft reportedly does not resolve all questions about AI-generated evidence; that area is likely to require separate treatment under the Evidence Act framework and the Bharatiya Sakshya Adhiniyam, 2023. ### Can AI be used to generate a first draft of a judgment? Under the reported framework, AI can assist in the drafting of a judgment in the same sense it can assist in drafting any document: it can produce structured text, suggest organisation, and surface relevant authorities. What it cannot do is produce the reasoning. The judge must independently reach the conclusions, and the draft must go through enough human engagement that the final product genuinely reflects the judge's reasoning rather than the AI's pattern output. ### What does "retrieval-grounded" mean and why does it matter? A retrieval-grounded AI tool searches a corpus of actual documents and bases its answers on retrieved passages. A general-purpose language model generates text based on statistical patterns in its training data. The difference is that a retrieval-grounded tool can show you exactly what source it drew from, and you can verify that source. A general-purpose model cannot do this because its output is not linked to any specific document - it is generated text that resembles what a legal answer might look like. Under the draft's verification requirements, only retrieval-grounded output gives you something verifiable to check. ### Will the draft create problems for lawyers who have been informally using AI tools? It depends on what "informally using" means. If a lawyer has been using AI tools to accelerate research and then verifying the output before relying on it, the draft's formal requirements largely codify good practice they are already following. If a lawyer has been submitting AI output without verification, or feeding client information into tools with permissive data terms, the draft brings existing professional conduct risks into sharper focus. The rules do not create new risks so much as they formalise consequences for risks that already existed. ### How should a law firm structure its AI use policy to prepare for the draft? A compliant policy should identify which tools are approved for use on client matters, specify the data handling requirements those tools must meet, mandate a verification protocol for all AI-generated research and drafting, and require disclosure logging so that accurate disclosure statements can be made. Firms that already have structured processes for associate work product review will find the adaptation less disruptive than firms where AI use has been entirely ad hoc. ### What is the significance of the SUPACE project in this context? SUPACE - the Supreme Court Portal for Assistance in Courts Efficiency - was an earlier AI initiative in which the court itself used AI to help process case materials and surface relevant precedents. SUPACE was firmly on the assistive side: it helped judges and court staff find information, it did not make decisions. The draft AI framework can be understood as the governance infrastructure that allows SUPACE-style tools to expand and for equivalent tools to be used by the broader legal ecosystem, with the assistive/decisional boundary as the governing principle. ### Does the draft address AI bias? From available reporting, the draft does address concerns about algorithmic bias, particularly in the context of sentencing and risk assessment tools. The concern is that AI systems trained on historical data can perpetuate historical biases - in ways that are opaque to the parties and cannot be effectively challenged. The draft's response appears to be prohibition or strict disclosure requirements rather than a technical standard for bias auditing, though detailed provisions on this may be included in the final version. ### What is the position on AI translation of regional language documents for court use? The draft reportedly permits AI translation for internal and preparatory use, with a higher standard applying when AI translation is being placed before the court as evidence of what a document says. In those cases, human verification of the translation - ideally by a qualified translator - appears to be required. Automatic submission of AI translations as certified translations is not within the permitted scope as currently reported. ### How does the draft interact with the DPDP Act, 2023? The Digital Personal Data Protection Act, 2023 sets standards for the processing of personal data. In a litigation context, case files routinely contain personal data of clients, witnesses, and third parties. The draft's confidentiality provisions for AI use in judicial contexts run parallel to DPDP obligations: both require that data be used only for the specified purpose, that it not be shared with third parties without authorisation, and that it be retained only as long as necessary. A tool that is compliant with DPDP in its data handling is likely to meet the draft's confidentiality requirements as well - though the specific obligations are independently grounded. ### What should individual lawyers do while the draft is still being finalised? The safest approach is to adopt the practices the draft would require now: use retrieval-grounded tools with verifiable citations, verify every citation before relying on it, keep records of what AI was used for each matter, and check the data handling terms of every tool you use on client matters. These are all good practices independently of the draft's final form, and they leave you prepared when mandatory requirements come into effect. ### Are there any Indian courts that have already issued binding AI rules? As of mid-2026, binding rules specific to AI use in judicial proceedings are limited. Individual High Courts have issued administrative notices and informal guidance in response to specific incidents, but there is no comprehensive, formally-issued framework from any court, including the Supreme Court. The draft under discussion would be the first such comprehensive framework if it is adopted in its current form. ### Can AI be used to prepare arguments for oral hearing? Preparing arguments is a drafting task, and AI assistance in drafting is within the permitted category. A lawyer can use AI to structure arguments, identify the key points of each ground, and surface supporting authority. The oral argument itself is delivered by the human lawyer. The constraint is that the lawyer must have genuine command of the arguments they present - both because the court will probe them and because the professional conduct obligation requires it. ### Where can I find the text of the draft framework? The draft framework had not been formally published as an accessible public document as of the date of this post. Reporting on its contents has appeared in Indian legal media and practitioner commentary. Once formally issued, it would ordinarily be published through the Supreme Court's official channels. Practitioners should monitor those channels and notifications from the Bar Council of India for the final version. --- ## Start with a compliant workflow The draft framework does not make AI in legal practice harder. It makes the right kind of AI use more clearly defined. Retrieval-grounded research over a verified corpus, every answer cited to a real judgment, built-in citator, and DPDP-aware data handling: these are not features added to Niyam in anticipation of a regulatory checklist. They reflect what responsible legal AI design requires, and they happen to align closely with what the draft framework expects. Niyam's [Research module](/solutions/research) covers the full research workflow - finding authority, verifying it is good law, and presenting it in a form you can check before you rely on it. The [Citator](/solutions/citator) handles good-law confirmation. [Drafting](/solutions/draft), [Notices](/solutions/notices), and [Translation](/solutions/translation) work as assistive tools that produce output for your review, not output that bypasses it. The [Matters](/solutions/matters) module keeps everything organised by matter so your disclosure records are accurate and accessible. For lawyers who want to understand how Niyam's approach compares to other tools across the specific criteria the draft raises, the [compare page](/compare) lays this out directly. For the tooling implications of the draft - what specific technical features distinguish compliant from non-compliant tools - see [what tools need to comply with Supreme Court AI rules](/blog/supreme-court-ai-rules-what-tools-need). 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.