ARTIFICIAL INTELLIGENCE AND LEGAL RESPONSIBILITY: In light of the Supreme Court's Draft AI Regulations

INTRODUCTION: RECURRING PATTERNS

Every major technological shift in the practice of law has encountered a degree of scepticism. I was not around to witness the transition from typewriters to computers, nor the gradual move from shelves of physical law reports and Bare Acts to digital databases and search engines. Yet, accounts from those who practised through those eras reveal a familiar pattern: many traditionalists initially resisted change, questioning both its necessity and reliability. However, over time, the efficiencies and advantages offered by new technologies became impossible to ignore, and what was once viewed with suspicion eventually became an indispensable part of legal practice.

A similar transition played out within my family’s chambers.

My great-grandfather, C.S. Narayana who practised law in the 1940’s, subscribed to All India Reporter (AIR), a tradition that still continues with my father, G.R. Ravishankar being the subscriber to AIR now. As a result, our office accumulated decades' worth of AIR volumes and digests. Whenever an old precedent had to be located, the exercise often involved hours of searching through shelves of books and then scanning through the few thousand pages.

During one such discussion, Mr Ashutosh Chitale, Director of AIR and a friend of my grandfather, A.R. Shivaprasad, remarked to my grandfather that every precedent reported in the AIR volumes subscribed to by C.S. Narayana over seventy years ago to the latest Supreme Court and High Court decisions was now available at the click of a button through a computer application.

Of course, my grandfather was reluctant. He was convinced that he was not suited to technological change. Nevertheless, at Mr Chitale's insistence, he adopted the AIR digital platform, with my uncle, G.N. Shrihari (an advocate himself), helping him set it up and learn its usage. Since then, there has been no looking back.

Having witnessed how previous generations in my own family eventually embraced technological change despite initial reservations, I suspect AI is simply the next chapter in that continuing story.

AI AND THE LEGAL PROFESSION

What was once limited to keyword searches has evolved into tools capable of summarising judgments, identifying relevant precedents, analysing contracts and even generating first drafts of legal documents.

This shift is already visible across the legal ecosystem. Dedicated legal AI platforms such as Harvey and Jurisphere, amongst others have emerged with the promise of streamlining legal research and professional workflows. Traditional legal research providers have also recognised the inevitability of this transition. Platforms such as Manupatra, SCC Online and Indian Kanoon have progressively integrated AI-powered functionalities into their services, seeking to move beyond simple database retrieval towards more intuitive and contextual legal assistance.

At the same time, general-purpose AI systems such as ChatGPT, Claude and Gemini, amongst others while not designed exclusively for legal use, have become increasingly common companions for lawyers, law students and researchers seeking to simplify complex information, obtain preliminary research assistance or even fine tune drafts.

In fact, hardly a day passes without headlines predicting that AI will transform our profession or some dystopian, far-fetched headings even claiming that it would render our job obsolete.

However, the benefits of AI are difficult to ignore. It can reduce hours of repetitive research, assist in navigating vast databases of case law, improve access to legal information and allow lawyers to devote greater attention to strategic and analytical work.

I am personally of the opinion that the legal profession's adoption of AI is not a question of if, but rather, a question of when and how.

WHEN IT GOES SOUTH

For all its promise, artificial intelligence is not without its imperfections. The same technology capable of analysing thousands of pages within seconds, identifying relevant authorities and simplifying complex legal issues is also capable of making mistakes.

Unlike traditional legal databases, which merely retrieve information, generative AI creates information. AI-generated responses can occasionally contain inaccuracies, omissions or entirely fabricated information delivered with a level of confidence and sophistication that makes them appear genuine.

An AI system may produce information that appears entirely credible but is, in reality, inaccurate, incomplete or wholly fabricated. In a legal context, the following are the common forms of errors;

1. Fabricated Citations (Phantom Cases) - The AI generates a case citation, party names or legal authority that does not exist.
2. Mischaracterisation of Holdings - The AI cites a genuine judgment but inaccurately states the legal principle or conclusion reached by the court.
3. Confusion Between Ratio and Obiter - The AI presents a passing observation or obiter dictum as though it were the binding ratio decidendi of the case.
4. False Attribution of Legal Principles - The AI attributes a proposition of law to a judicial decision that never articulated or endorsed that principle.
5. False Confidence in Uncertain Areas - The AI presents unsettled, disputed or emerging legal issues as though the law were clear and settled.

The aforesaid outputs are more often than not presented in clear, confident and persuasive language, giving them an appearance of authority that may not always be deserved. A fabricated citation may resemble a genuine precedent, while a mischaracterised judgment may appear entirely consistent with the proposition for which it is cited. Unless every authority is independently verified, there is a real risk that inaccuracies may find their way into legal research, pleadings, opinions and even judicial decision-making.

Unfortunately, these errors have already found their way into courtrooms. In recent months, lawyers and litigants have submitted pleadings and legal arguments containing AI-generated citations that were either wholly fictitious or materially inaccurate and to make matters worse, there have been instances whereby judgments were passed and authored placing blind reliance on AI.  

The following are a few select instances;

• Supreme Court of India - Gummadi Usha Rani v. Sure Mallikarjuna Rao

In this case, a trial court in Andhra Pradesh relied upon several non-existent precedents. To name them, T Subramani v. M.Natarajan, (2013) 14 SCC 95; Chidambaram Pillai v. SAL Ramasamy, (1071) 2 SCC 68; Lakshmi Devi v. K. Prabha, (2006) 5 SCC 551 and Gajanan v. Ramdas (2015) 6 SCC 223.

It was later identified as AI-generated and the petitioner preferred an appeal with the Andhra Pradesh High Court. The Hon’ble High Court proceeded to decide the matter on its merits; it acknowledged that the authorities cited were fabricated.

The issue eventually reached the Supreme Court, which observed that reliance on such fake judgments directly impacts the integrity of judicial decision-making and went on tostate "a decision based on such non-existent and fake alleged judgments is not an error in the decision making but may amount to misconduct and entail legal consequences.

• Karnataka High Court

In this case, a trial court order in Karnataka which relied upon four Supreme Court judgments that did not exist. The matter ultimately reached the High Court of Karnataka, whereby a probe was initiated and it was identified that the order was made based on AI output and subsequently, action was taken against the judge.

• Delhi High Court

In this case, A petition before the Delhi High Court was withdrawn after it emerged that several of the authorities relied upon were AI-generated and could not be traced to any recognised legal source.

• Bombay High Court - Deepak v. Heart & Soul Entertainment Ltd

In this case, the Hon’ble Bombay High Court identified several indicators that made it obvious that the submissions made by the respondent had been generated using AI, including repetitive language, unusual formatting and most significantly, reliance on a case that did not exist. The incident required the Court to spend valuable time verifying authorities that should have been checked beforehand by the counsels. The court, however, recognised the utility of AI as a research tool but reiterated that while it can assist legal work, the responsibility for its accuracy remains firmly with the lawyer who signs and submits it.

The Bar Council of India has also adopted a similar stance. Responding to a parliamentary query on artificial intelligence in the legal profession, the Bar Council of India clarified that advocates cannot evade professional responsibility by attributing errors to artificial intelligence.  The exact words of the Bar Council were, ‘If a false case, fake quote, or misleading proposition is filed, the advocate cannot avoid accountability by saying that a machine produced it. The machine neither signs the pleading nor owes duties to the court. The advocate does.’

I find myself in complete agreement with this position. In my view, AI should be treated no differently from a junior associate or an intern tasked with carrying out preliminary work. It can assist in identifying authorities, summarising material and providing a useful starting point, but its work product is only as valuable as the scrutiny applied to it by the supervising lawyer. No prudent advocate would blindly reproduce a junior's research without reviewing the underlying authorities; the same standard must apply to AI-generated outputs.

Ultimately, the duty of diligence is non-delegable. Whether an error originates from an intern, a junior associate, a legal database, or an artificial intelligence system is immaterial. The advocate whose name appears on the pleading bears the responsibility of ensuring that every proposition, citation and authority placed before the court is accurate and reliable. AI may be a powerful assistant, but it cannot be a substitute for professional judgment.

THE SUPREME COURT'S DRAFT AI REGULATIONS

With the waves that AI was making, it became hard for the Supreme Court of India to ignore. The Supreme Court's AI Committee has recently released a draft regulatory framework governing the use of artificial intelligence within the bar and the bench for public consultation inviting comments from stakeholders, legal professionals, technology experts and the public at large until 20th June 2026 (Send in your opinions about the draft regulations via email to office.regcc@sci.nic.in).

PRELIMINARY ASSESSMENT

The draft is designed with a broad scope of application. Once brought into force, they will govern the use, deployment and integration of artificial intelligence not only within the Supreme Court of India, but also across the High Courts, subordinate courts, tribunals and statutory commissions performing adjudicatory functions throughout the country.

Their implementation, however, will be phased, with the Chief Justice of India empowered to notify their commencement in relation to the Supreme Court, while the respective Chief Justices of the High Courts may determine the date and manner of their adoption within their jurisdictions, including bringing different provisions into force at different stages.

Amongst the most significant provisions of the Draft Regulations is Regulation 4, which establishes the foundational principle of the entire framework;
Artificial intelligence is to remain an assistant, not a decision-maker. It is made clear that ‘the usage of Artificial Indulgence is strictly subservient to human judgment and judicial authority’ and may function only in an assistive capacity.

Further, the Regulation further provides that the Bangalore Principles of Judicial Conduct shall continue to govern judicial officers even when they engage with artificial intelligence.

If Regulation 4 establishes that judicial power must remain human, Regulation 8 ensures that accountability remains human as well.

The Regulation unequivocally provides that where a decision is taken with the assistance of artificial intelligence, responsibility for that decision continues to rest solely with the officer making it. The officer cannot seek refuge behind an AI-generated output, the opacity of a "black box" (An AI system whose internal processes and decision-making logic are not transparent) system, or even an AI hallucination to justify an incorrect or unlawful decision. After all, an AI system neither takes an oath of office, nor signs a judicial order, nor owes duties to the court or litigants. The individual exercising authority remains accountable for the consequences of that exercise.

Equally important is the Regulation's insistence that AI-generated outputs be treated as advisory rather than determinative. As a general rule, reasonable efforts must be undertaken to verify the accuracy of AI-generated information before it is relied upon.

In its essence, regulations 8 states, while technology may evolve, accountability within the justice delivery system must always remain traceable to a human being.

Regulation 16 adopts a position that is neither hostile nor overly cautious towards artificial intelligence. Rather than treating AI as an inherently suspect technology, the Regulation establishes a presumption in favour of its responsible adoption wherever it can improve access to justice, reduce delays or enhance administrative efficiency but cannot be deployed for dispute-outcome prediction.

Interestingly, the refusal to permit the use of any AI System or AI Tool must be recorded in writing the restriction must be reasonable and to such extent so as to address the concern identified.

One of the more progressive aspects of the Draft Regulations is the emphasis placed on transparency in the use of artificial intelligence. Recognising that AI can influence the preparation of pleadings, evidence, research and administrative processes, Regulation 43 requires disclosure whenever AI materially assists in court proceedings. Courts are obligated to inform parties when AI tools are used in case management, document analysis or judicial administration, while litigants and legal representatives must expressly disclose the use of AI in the preparation of pleadings, documents or evidence through prescribed declarations and certificates.

The Regulations further empower courts to seek additional information regarding the AI system used, the extent of assistance provided and the verification measures undertaken before relying upon AI-generated content. Significantly, any person relying on synthetic data or synthetic information in judicial proceedings is also required to disclose its use.

Regulation 43(6) once again reiterate that whereby the AI-generated content is found to be false, fabricated or misleading, responsibility remains with the person who submitted it and the fact that the content originated from an AI system cannot be invoked as a defence.

To strengthen verification mechanisms, the regulations also seeks to establish an AI Content Verification Authority, tasked with developing and updating standards, tools and protocols for verifying AI-generated content used in court processes.

CAN AI BE TRULY FAIR?

Regulation 6 seeks to address one of the most widely discussed concerns surrounding artificial intelligence, algorithmic bias. The Regulation mandates that AI systems used within the judicial ecosystem must be designed, trained and deployed in a manner that promotes fairness and avoids discrimination. It further prohibits the deployment of systems that perpetuate or amplify bias on grounds such as race, religion, caste, sex, gender, disability, language or economic status, while specifically emphasising the need to protect vulnerable and marginalised groups.

At a conceptual level, the objective is difficult to fault. Given the judiciary's constitutional commitment to equality and non-discrimination, any technological tool deployed within the justice delivery system must necessarily be held to the same standards.

That said, this is perhaps one of the provisions that raises the greatest practical questions. My admittedly lay understanding of artificial intelligence is that AI systems do not develop views of their own; they learn from the data on which they are trained and existing. AI does not operate in isolation from society's prejudices; it often learns from them. If historical data reflects disparities based on caste, gender, socio-economic status or other factors, there is a risk that those patterns may be mirrored in the outputs produced by the system. In that sense, bias may not be an exception to AI systems but an inevitable risk inherent in their design.

To its credit, however, the Draft Regulations do not treat fairness as a mere aspirational principle. They attempt to address the issue through a multi-layered oversight framework.

Regulation 14 requires AI systems to be trained on data that is accurate, representative and, to the extent feasible, free from discriminatory bias.

Regulation 35 further mandates a Technical and Ethical Impact Assessment before any AI system can be approved for deployment, specifically requiring an evaluation of the nature and representativeness of training data as well as the risks of bias, error and misuse.

Beyond the initial approval stage, Regulation 9 subjects AI systems to continuous monitoring and periodic technical, legal and ethical audits, with mechanisms to detect, document and address biases throughout their lifecycle.

Finally, Regulation 39 establishes an AI Incident Database through which any bias, malfunction or error with potential legal consequences must be reported, investigated and shared across jurisdictions to facilitate corrective action.

While only time will reveal the effectiveness of these safeguards, the Regulations do appear conscious of the reality that fairness cannot simply be programmed into an AI system once; it must be continuously tested, monitored and enforced throughout its deployment.

AI, DATA PRIVACY AND INSTITUTIONAL SAFEGUARDS

The Draft Regulations recognise that the successful adoption of artificial intelligence within the judiciary is inseparable from the protection of judicial data. Regulations 47 and 48 therefore establish a dedicated framework governing privacy, data handling and cybersecurity obligations.

At the outset, all AI systems deployed in court processes are required to comply with the Digital Personal Data Protection Act, 2023, the Information Technology Act, 2000 and other applicable laws relating to personal data and judicial information.

Regulations 10 to 16 collectively seek to strike a balance between encouraging innovation and protecting the rights and interests of those interacting with the justice delivery system.

The regulations recognise that courts deal with some of the most sensitive categories of personal information and accordingly, the use of AI is made subject to the principles of purpose limitation, data minimisation and privacy by design, with specific reference to the Digital Personal Data Protection Act, 2023.

Similarly, personal information is required to be anonymised, wherever technically feasible, before being used for the training, testing or refinement of AI systems and the same is mandated by Regulation 48(4).

The framework further seeks to ensure that AI systems are used only for the purposes for which they have been approved. Regulation 11 embodies the principle of purpose limitation by requiring fresh approval before an AI system can be repurposed beyond its originally authorised function. This is an important safeguard against "function creep", where technology introduced for one objective gradually expands into areas for which it was neither designed nor assessed.

Equally noteworthy is Regulation 12 and the emphasis it places on proportionality. The Regulations recognise that not all uses of AI carry the same degree of risk. Consequently, applications affecting personal liberty, substantive rights or the integrity of judicial outcomes are subjected to heightened safeguards, including mandatory human oversight and independent scrutiny.

In addition, Regulations 14 and 15 address concerns relating to the quality of data used to train AI systems. AI tools deployed within courts must be based on accurate, representative and lawfully obtained data, and systems trained on unlawfully collected or demonstrably biased datasets are expressly prohibited. Coupled with the requirement for robust cybersecurity measures to protect the confidentiality, integrity and availability of court data, the framework attempts to address many of the concerns that have traditionally accompanied the deployment of AI in sensitive public institutions.

Regulation 48(5)  mandates periodic security audits, annual reviews of access-control protocols and continuous monitoring of security vulnerabilities and data incidents.

While Regulation 15 rightly emphasises the need for "robust, layered and continuously updated" cybersecurity measures, the provision remains largely aspirational in nature. The Regulations at the current stage do not prescribe any minimum technical standards, audit requirements or security benchmarks against which compliance can be assessed. Consequently, while the objective is laudable, it remains unclear what specific safeguards would satisfy the regulatory requirement in practice.

PERMISSIBLE USES AND PROHIBITED USES

Perhaps the most practical part of the Draft Regulations is the distinction it draws between what artificial intelligence may do and what it must never do. Regulations 19 and 20 effectively establish the operational boundaries of AI within the justice delivery system.

Regulation 19 - Permissible Uses

Subject to prior approval and human supervision, AI may be utilised for a range of administrative and assistive functions, including:

     (i)         Case management, defect identification, cause list preparation and hearing scheduling;

    (ii)         Automated transcription of court proceedings, subject to verification;

   (iii)         Translation of judgments, pleadings and legal documents;

   (iv)         Legal research, precedent retrieval, citation verification and document summarisation;

    (v)         Filing assistance, record management and judicial resource allocation;

   (vi)         AI-powered chatbots and virtual assistants to assist litigants;

  (vii)         Accessibility tools such as speech-to-text, text-to-speech and Braille translation;

 (viii)         Document verification and fraud detection;

   (ix)         Anonymisation of judgments and court records;

    (x)         Court administration, performance assessment and backlog management; and

   (xi)         Generation of standardised notices, summons and administrative documents.

A common thread running through these permissible uses is that AI is largely confined to supporting judicial administration, improving efficiency and enhancing access to justice, while remaining subject to human oversight and verification.

Regulation 20 - Prohibited Uses

The following prohibitions are absolute and cannot be relaxed:

     (i)         Reaching judicial outcomes solely through algorithmic decision-making;

    (ii)         Allowing AI to adjudicate disputes or determine sentences independently;

   (iii)         Using AI for risk scoring, including predicting flight risk, recidivism or bail eligibility;

   (iv)         Assessing the credibility of parties or witnesses through AI systems;

    (v)         Predicting or profiling the future behaviour of litigants, accused persons or witnesses;

   (vi)         Deploying opaque or unexplainable AI systems in matters affecting rights or liberty;

  (vii)         Using AI for surveillance or continuous monitoring of judges, lawyers or litigants;

 (viii)         Submitting AI-generated material as evidence without full disclosure of its origin;

   (ix)         Using personal data to train AI systems without appropriate approval; and

    (x)         Employing AI in any manner that compromises judicial independence or the confidentiality of judicial deliberations.

Any violation of the aforesaid prohibitions must be reported immediately to the AI Secretariat and placed before the AI Committee for examination. Following an enquiry, the Committee is empowered to direct appropriate remedial measures, including the suspension of the relevant AI system where necessary.

Further, there is a requirement for Annual Transparency Reports, through which High Courts, Tribunals and Commissions must publicly disclose the AI systems in use, audit outcomes, reported incidents and compliance measures. Collectively, these provisions seek to ensure that the adoption of AI within the judiciary remains transparent, auditable and subject to public scrutiny.

GOVERNANCE AND INSTITUTIONAL OVERVIEW

Recognising that effective AI governance requires more than substantive principles, the Draft Regulations establish a detailed institutional framework for oversight and implementation. At the apex of this structure is a permanent, full-time Apex Body to be constituted at the Supreme Court of India, which will be responsible for AI governance, policy development, standard-setting, oversight and innovation across the judiciary.

To assist it in discharging its functions, the Apex Body is empowered to constitute the following specialised committees:

     (i)         Judicial Committee

    (ii)         Technical Committee

   (iii)         Committee on Infrastructure and Finance

   (iv)         Case and Data Management Committee

    (v)         Cyber Security Committee

The Regulations also provide for the establishment of a Centre of Research and Excellence on Artificial Intelligence (CoRE-AI), comprising experts from technology, law and academia. The Centre is intended to function as the judiciary's dedicated research and advisory arm, providing technical expertise and legal compliance support to the Apex Body.

At the operational level, the Supreme Court and every High Court are required to constitute an AI Committee to oversee, regulate and facilitate the responsible adoption of AI within their respective jurisdictions in accordance with standards prescribed by the Apex Body. These Committees are, in turn, supported by dedicated AI Secretariats, headed by an officer of the rank of District Judge and staffed by experts in law, technology, data science and judicial administration.

The AI Secretariats are entrusted with day-to-day implementation functions, including maintaining records, monitoring AI tools and granting expedited approvals for certain low-risk administrative AI applications. Collectively, these bodies create a multi-tier governance framework that seeks to ensure that the deployment of AI within the judiciary remains subject to continuous supervision, technical expertise and institutional accountability.

OPERATIONALS

The Draft Regulations place considerable emphasis on continuous oversight even after an AI system has been approved for deployment. To promote transparency and institutional accountability, every court is required to maintain an AI Register documenting all approved AI systems, their intended purposes, service providers, approval conditions, audit records, impact assessments and any AI-related incidents. Subject to confidentiality and cybersecurity considerations, the Regulations also contemplate public dissemination of this information, thereby enhancing transparency regarding the judiciary's use of AI.

The framework further mandates periodic technical, legal and ethical audits of all AI systems at least once every year. Audit findings are required to be recorded in the AI Register and may influence whether a particular AI system continues to remain in operation. Interestingly, the Regulations require all audits to be conducted in-house and prohibit the sharing of source codes, algorithms and datasets with external entities. While this approach may safeguard confidentiality and judicial independence, one may question whether the absence of independent third-party audits could potentially limit external scrutiny and objective assessment of AI systems.

The Regulations also establish an AI Incident Database, requiring courts to systematically record and report all AI-related malfunctions, errors, biases and other incidents with potential legal consequences. Importantly, lessons learned from such incidents are to be shared across jurisdictions, enabling corrective measures to be adopted throughout the judicial system. This reflects an understanding that AI governance must be dynamic and capable of learning from operational experience.

Another welcome safeguard is the express preservation of human discretion. Regardless of the sophistication of an AI system, the officer supervising it retains complete authority to accept, modify or reject any AI-generated output and is required to exercise independent professional judgment. In essence, the Regulations reiterate that AI recommendations may inform decision-making, but they can never dictate it.

Finally, recognising that technology is not infallible, the Regulations require every High Court to establish emergency and fall-back protocols to ensure continuity of court functioning in the event of AI failure, malfunction or suspension. This is a practical inclusion, acknowledging that excessive dependence on technology can itself create vulnerabilities. By requiring courts to maintain alternative processes and contingency plans, the framework seeks to ensure that the administration of justice remains uninterrupted even when the technology supporting it does not.

AI LITERACY

The Draft Regulations rightly recognise that the effective governance of artificial intelligence depends not only upon rules and oversight mechanisms, but also upon the people expected to use the technology.

Regulations 49 and 51 therefore mandate regular and structured training for judges, advocates and court staff who interact with AI systems in the course of their duties. Importantly, the training framework is not limited to technical operation alone. It encompasses the capabilities and limitations of AI systems, the identification of bias and hallucinations, legal and ethical obligations, data protection principles, cybersecurity awareness and procedures for reporting AI-related incidents.

A particularly welcome feature is the emphasis on accessibility. The Regulations require training programmes to be available across all levels of the judiciary, including district courts, and to account for India's linguistic diversity.

The responsibility for developing these programmes has been entrusted to the AI Secretariats in consultation with domain experts and judicial training institutions, while High Courts are required to maintain annual training calendars and periodically review the effectiveness of such programmes. In doing so, the Regulations acknowledge an important reality: the challenges posed by AI are evolving rapidly, and the competence of those using these systems must evolve alongside them. After all, even the most sophisticated regulatory framework will have limited value if the individuals expected to implement it do not fully understand the technology they are supervising.

CONCLUDING REMARKS

On the whole, the Draft Regulations offers a thoughtful and balanced approach to the integration of artificial intelligence within the justice delivery system.

Rather than treating AI as either a threat to be resisted or a solution to be blindly embraced, the framework recognises both its immense potential and its inherent limitations.

The emphasis on human oversight, accountability, transparency and data protection ensures that technological innovation remains subordinate to the core values of the judicial process.

While certain provisions may benefit from greater clarity as the framework evolves, the Regulations undoubtedly constitute a significant step in the right direction. In an era where artificial intelligence is poised to become an integral part of legal practice, these Regulations provide a much-needed foundation to ensure that efficiency and innovation are pursued without compromising fairness, integrity and public confidence in the administration of justice.

My grandfather and father eventually adopted the AIR digital platform, and it transformed their practice. I suspect the profession's relationship with AI will follow a similar arc.