Judicial AI

What is Judicial AI?

The term 'Judicial AI' refers to the usage of Artificial Intelligence (AI), and other machine learning tools in judicial functioning and court-related activities. It is the use of AI in assisting administration of justice, research or analysis of any of the aspects involved in judicial decision-making.

Official Definition of Judicial AI

There exists no internationally accepted, explicit or verbatim definition for the term 'judicial AI'. However, several national and international policy documents define the term 'Artificial Intelligence' contextually based on which a general understanding could be deduced. Although no internationally accepted definition presently exists, emerging comparative research increasingly conceptualises Judicial AI as AI systems deployed within judicial institutions that directly or indirectly affect the administration of justice, adjudication, court administration or access to justice. The Oxford Institute of Technology and Justice similarly frames AI in courts through the lens of AI-assisted proceedings and fair trial rights rather than through purely technological classifications, suggesting that Judicial AI should be defined by its constitutional impact rather than its underlying technical architecture.

As defined in policy frameworks

Draft Regulations for Use of Artificial Intelligence (AI) in Court [Supreme Court][2026]

The draft regulations defined the term 'artificial intelligence' as a machine based system that is capable of inferring information from data, generate recommendations, predictions or decisions with varying degrees of autonomy that are being deployed for court processes. This definition specifically excludes any general-purpose software or digital tools that do not operate or functionally depend on artificial intelligence. It is seen that the definition aims to provide a contextual understanding of AI tools in judiciary rather than a general understanding of AI itself.

The draft regulations separately defines “Adjudicatory Function” and “Administrative Function” (Regulation 3(1)(a) and 3(1)(b) respectively). That distinction is important and carries weight. Adjudicatory functions, such as hearing evidence, framing issues, pronouncing judgments and sentences, will attract a higher level of human oversight, whereas in terms of administrative functions, such as filing, scheduling, notice issuance, and record maintenance, they may allow for somewhat greater automation, though objectively always within sanctioned limits.

There are several definitions that address distinct risks and concepts. For example, Regulation 3(1)(z) defines “Hallucination” as the phenomenon by which an AI system produces outputs that appear plausible but are factually wrong, fabricated, or unsupported, including fabricated citations to cases or statutes. Regulation 3(1)(n) defines “Black Box” as an AI system whose internal reasoning is not transparent and cannot be explained in accessible terms. Regulation 3(1)(zg) defines “Risk Scoring” as the use of AI to assign a score estimating the probability of a person’s future behaviour, such as commission of an offence, recidivism, or failure to appear before a court. Regulation 3(1)(j) defines “Algorithmic Decision-Making” (‘ADM’) as any use of algorithmic output to inform, recommend, or arrive at a decision affecting a person or process.

“Human-in-the-Loop” (‘HITL’) is defined under Regulation 3(1)(zb). It is perhaps the most critical operational concept in the entire document. It means a governance process in which every AI output is subject to mandatory human review and where the final decision-making authority always vests with a human.

The definition of “explainability” as defined under Regulation 3(1)(w) notes that an AI system is explainable if it can, upon request, generate a comprehensible account of the reasoning and factors behind its output “in terms that can be understood by the judicial officer, court staff, or litigant” without requiring specialist technical knowledge from the recipient. This sets a demanding standard. Explainability is measured not by what the system can produce in principle, but by what a non-expert person can actually understand.

The Draft Regulations’ definitional architecture also defines both basic and advanced concepts, such as anonymisation, artificial intelligence, generative artificial intelligence, large language model, machine learning, sensitive judicial data, synthetic data/information, data minimisation, and more.

Policy on the Use of Artificial Intelligence in Judicial and Court Administration[Gujarat HC][2026]

The Gujarat HC released a policy whereby it defined and clearly demarcated the permitted and prohibited uses of AI tools in the administration of justice. The usage of the tools in administrative tasks, legal research support, drafting assistance, translation and case management were deemed permitted subject to verification of the veracity of the information and active human oversight. AI usage intending to replace the application of mind or decision making is strictly prohibited thereby upholding the autonomy of the ultimate judicial process. The policy clearly establishes that AI is to be used only as an assistive tool that minimises workload and streamlines processes and that it cannot be misconstrued as a replacement to judicial reasoning, thinking and ultimate decision making.

Policy regarding the use of Artificial Intelligence (AI) Tools in District Judiciary [Kerala HC][2025]

The Kerala HC's policy was the first document that established guidelines for the responsible and restricted use of AI in judiciary. The policy enlists the dangers of unregulated AI usage such as violation of confidentiality, erroneous or biased results. The Kerala HC's cautious approach of compulsory human supervision and exclusion of AI from decision-making is reflected across subsequent policy frameworks. The policy calls for judicial training in the ethical, technical and practical aspects of using AI to ensure proper management and acknowledgment of the risks involved.

Official Government Reports

White Paper on Artificial Intelligence and Judiciary [Supreme Court][2025]

The Supreme Court's White Paper highlights the disparity between AI systems and judicial process, observing that the scale, speed and automation that makes AI desirable may compromise the integrity of the judicial machinery. Caution is advised in order to ensure that the ultimate decision making authority remains human i.e., the judges and that AI tools are only used to assist the process. According to the paper, technological integration in the judicial system is an ongoing process, from digitisation to the recent deliberations of AI tools, to increase the workflow efficiency of the judiciary. The perils of using AI in judiciary include AI hallucination, diminished human judgment, evidence fabrication, potential breach of privacy and AI bias. The paper calls for human oversight, verification of outputs and mechanisms to ensure confidentiality, accuracy and absence of bias. The paper proposes the formation of an 'AI Ethics Committee', a policy framework for ethical use of AI, accountability and mandatory AI training for judicial officers.

Reports by International Organisations

AI for Justice: Ethical, Fair and Robust Adoption in India's Courts [United Nations Development Programme][2026]

The report examines the incorporation of AI in the Indian Judiciary through already existing AI deployments, institutional readiness and possible governance gaps. The potential consequences of judicial AI tools on due process, independence of judiciary, privacy, fair trial and public trust is evaluated to understand the different concerns that may arise. The report proposes comprehensive AI governance models that prioritise protection of rights, carry out assessments of institutional readiness, possible risks, technical issues and ensure continuous evaluation post-deployment. Mandatory human oversight, dedicated technical machinery and standardisation of tools is also recommended to ensure that transparency in the judicial decision making process is ensured.

The report recommends courts begin with the institutional readiness assessment (tool to assess) Courts that score optimally, reflected in positive response to all essential questions a score of at least 60%, may proceed to the next step. If the score falls below this threshold, the court must remedy limitations before proceeding to the risk assessment. (tool to assess) The risk assessment enables courts to classify the intensity of risk associated with the tool. The subsequent action should be calibrated accordingly. For low-risk tools, deployment may proceed with basic safeguards in place. Medium-risk tools should trigger non-negotiable assurances from vendors in the technical assessment. 

In the case of high-risk tools, the report recommend that courts undertake the full technical assessment (tool to assess) and proceed only in case of a reasonably high score (>=60%). In case of high risk with prohibition, courts should refrain from use AI, as the risks outweigh potential benefits. Finally, once AI is deployed and ready for adoption, the monitoring assessment tool (tool to assess) provides supervisory questions to keep the AI system in check and mitigate harms over time.

p.57 DAKSH, Digital Futures Lab and UNDP India, AI for Justice: Ethical, Fair and Robust Adoption in India’s Courts (UNDP India, Feb 2026) <https://www.dakshindia.org/wp-content/uploads/2026/02/AI-for-Justice-Ethical-Fair-and-Robust-Adoption-in-Indias-Courts-Final-Report.pdf> accessed 29 June 2026. https://www.dakshindia.org/wp-content/uploads/2026/02/AI-for-Justice-Ethical-Fair-and-Robust-Adoption-in-Indias-Courts-Final-Report.pdf

Responsible AI in Justice: Regional Policy Guidance for UNDP Programming in Asia-Pacific [United Nations Development Programme][2026]

The report establishes a regional governance framework for the responsible adoption of artificial intelligence across justice systems in the Asia-Pacific region. Recognising that AI deployment within justice institutions directly affects access to justice, due process, judicial independence and public trust, the report advocates a human rights-based and risk-sensitive approach to AI governance. Rather than encouraging technology adoption for efficiency alone, it emphasises that AI systems must remain subordinate to constitutional principles, the rule of law and internationally recognised human rights standards.

The report identifies several governance priorities, including human oversight, explainability, transparency, accountability, data governance, privacy protection, institutional capacity building and continuous monitoring throughout the AI lifecycle. It further recommends that justice institutions undertake multidisciplinary risk assessments, establish clear governance structures, strengthen technical expertise and ensure meaningful stakeholder participation before introducing AI into judicial processes. Importantly, the guidance recognises that the appropriate regulatory response should vary according to the nature of the AI application and its potential impact on fundamental rights, thereby supporting a proportionate, risk-based model of governance. The report serves as a regional normative framework that complements country-specific initiatives such as the UNDP's AI for Justice: Ethical, Fair and Robust Adoption in India's Courts and provides an important comparative benchmark for developing India's Judicial AI Governance Framework.

The Algorithm in the Courtroom: How Artificial Intelligence Is Reshaping Justice and the Rule of Law Across Asia and the Pacific [United Nations Development Programme][2026]

The report provides one of the first comprehensive comparative studies of AI adoption across justice systems in the Asia-Pacific region. Examining developments across multiple jurisdictions, the report analyses the growing use of AI in judicial administration, legal research, case management, language translation, evidence analysis and decision-support systems. Rather than treating AI solely as a technological innovation, the report evaluates its implications for the rule of law, judicial independence, procedural fairness, access to justice and public confidence in judicial institutions. It identifies common governance challenges including algorithmic bias, opacity, data protection, cybersecurity, unequal digital capacity and institutional readiness, while emphasising that judicial AI should augment rather than replace human adjudication.

"India 
NATIONAL AI REGULATION AND JUSTICE SECTOR POLICIES 
India’s strategic vision for artificial intelligence is articulated in NITI Aayog’s National Strategy for Artificial Intelligence: #AIforAll, which frames AI as a cross-cutting enabler of inclusive growth and publicsector transformation. Within the justice sector, this vision aligns with the e-Courts Mission Mode Project (Phase III)—a large-scale digital transformation programme with a budget exceeding ₹7000 crore to modernise case management, improve service delivery, and build the data foundations needed for responsible AI. Two flagship judicial tools embody the augmentation-not-automation stance: SUPACE (Supreme Court Portal for Assistance in Court’s Efficiency), an AI-assisted legal research environment intended to help judges and registries sift and synthesise large volumes of material; and SUVAS (Supreme Court Vidhik Anuvaad Software), which uses AI-enabled translation to render judgements into multiple Indian languages, expanding accessibility and aligning with language rights. Together, these initiatives situate AI as a support to human adjudication and as an access-to-justice accelerator, rather than as a substitute for judicial decision-making. The Kerala High Court became the first state in India to issue an AI-use policy for the district judiciary, allowing only court-approved tools as strictly supervised assistive aids (not a substitute for legal reasoning), requiring verification of all AI outputs, mandating detailed audit trails, and warning that indiscriminate use can trigger disciplinary action and risks privacy, data security, and public trust. 

AI SYSTEMS AND TOOLS IN JUSTICE 
Operational deployments reflect a split between court-centred augmentation and policing-centred surveillance and prediction. In courts, SUPACE enhances research and workflow efficiency, while SUVAS directly addresses language barriers by translating judgements at scale. Another innovation is Adalat AI, a non-profit providing affordable real-time transcription and translation in trial courts, which has already scaled to over three thousands courtrooms, cutting case durations by an estimated 30–50% in those pilot courts. India’s National Informatics Centre has also developed a retrieval-augmented generation (RAG) judicial search assistant for Supreme Court justices, which combines information retrieval with language generation to produce more accurate, source-grounded answers. The tool is intended to provide a specialised search service that helps judges quickly locate specific information, and it will initially be made available to all Supreme Court judges, with plans to scale it across the wider judiciary. In law enforcement, multiple agencies have adopted facial recognition technology (FRT), including the Delhi Police Automated Facial Recognition System used to identify suspects and trace missing children; similar deployments exist in Chennai and in states such as Telangana and Punjab. Predictive policing initiatives leverage the Crime and Criminal Tracking Network & Systems (CCTNS) to map hotspots and anticipate criminal activity, supporting patrol allocation and investigations. 

DEPLOYMENT PATTERNS AND INSTITUTIONAL ADOPTION 
Judicial deployments are carefully staged, confined to research, translation, and administrative efficiency. By contrast, policing deployments are operational, fast-moving, and data intensive, with FRT and predictive analytics integrated into routine practice across several jurisdictions. This asymmetry underscores the need for harmonised safeguards across institutions implementing similar classes of AI risk. 

ETHICAL AND GOVERNANCE CONSIDERATIONS 
A vibrant scholarly and civil-society debate foregrounds constitutional protections under Article 14 (equality before the law) and Article 21 (right to life and personal liberty). Concerns include the legality and proportionality of mass surveillance, potential profiling of marginalised communities, privacy violations, error rates and demographic bias in matching, and the opacity of vendor systems. For predictive policing, critics note the danger of historical bias reinforcing over-policing where crime data are already concentrated, risking feedback loops that entrench inequality. In this context, accessenhancing tools (e.g., SUVAS) are broadly welcomed, while surveillance-oriented deployments face calls for statutory basis, human-rights impact assessment, transparency, and redress. 

JUSTICE STAKEHOLDER ENGAGEMENT AND PUBLIC PERCEPTION 
Engagement spans the Supreme Court, High Courts, bar councils, legal academia, and civil society. Priorities converge on procurement safeguards, accuracy testing, explainability, privacy compliance, and training. Public trust hinges on demonstrable benefits (faster, more intelligible services) paired with credible accountability for high-risk uses, especially FRT."

The report further observes that jurisdictions across the region have adopted divergent regulatory approaches, demonstrating the absence of harmonised standards governing judicial AI. By documenting comparative experiences and identifying emerging governance trends, the report provides an important empirical foundation for developing context-specific regulatory frameworks. It also establishes the need for regionally informed governance models, which are subsequently developed in the UNDP's Responsible AI in Justice and AI for Justice reports.

The following principles draw on leading international frameworks, including the UNESCO Recommendation on the Ethics of AI (2021), the UNESCO Guidelines for the Use of AI Systems in Courts and Tribunals (2025), the OECD AI Principles (2019), and the Council of Europe CEPEJ European Ethical Charter on the Use of AI in Judicial Systems (2018) as well as guidance from judiciaries, recommendations of national human rights institutions, and analyses by digital rights organisations across Asia–Pacific. These principles are organised under three complementary normative frameworks that underpin UNDP’s approach to governance and justice: the human rights-based approach (HRBA), rule of law, and people-centred justice. These are supplemented by essential technical safeguards. Together, they provide an integrated foundation for the responsible design, governance, and deployment of AI in the justice sector.

1st AIAB report on the use of Artificial Intelligence (AI) in the Judiciary based on the Information contained in the resource centre on cyber justice and AI [European Commission For The Efficiency of Justice][2025]

The report maps deployed AI systems across courts to assess their compliance with the 2018 European Ethical Charter on AI Judicial Systems. Empirical mapping of the deployed judicial technologies signals increasing efficiency, non-negotiable human oversight and ensured fundamental rights protection. The report reinforced ethical evaluation pre-deployment and continuous monitoring post-deployment to ensure conformity with ethical guidelines as AI tools in judiciary are categorisable as High-Risk.

Artificial Intelligence and the Administration of Justice [United Nations General Assembly][2025]

The report examines the increasing integration of AI into justice systems worldwide and evaluates whether such incorporation strengthens or undermines access to justice, fair trial and judicial independence. The report focuses on the possible pressure points of AI integration such as replacement of legal aid lawyers and the resultant inaccessibility of justice to poorer litigants, automation bias that judicial authorities begin to confirm with, de-skilling and growing dependency amongst judges and increasing global inequality in access to justice. The preservation of the right to a human judge and lawyer is proposed as the strongest recommendation of the report, which proposes AI education, formulation of AI guidelines and maintaining transparency.

Use of AI Systems in Courts and Tribunals [United Nations Educational, Scientific and Cultural Organisation][2025]

The UNESCO document proposes a global framework to guide responsible adoption of AI in justice systems. It reframes AI Governance as a human rights issue rather than a purely technological one and propagates transparency in legal decision making to reduce algorithmic bias, violation of privacy and ensure procedural fairness. Preservation of judicial autonomy, legal reasoning and accountability is supreme whilst deliberating integration of AI in judiciary. A mandatory algorithmic impact assessment before deployment is proposed to help identify possible harm. Further, periodical continuous monitoring to regulate post-deployment functioning is also deemed crucial.

European Union AI Act [2024]

The European Union AI Act represents the first comprehensive statutory framework regulating artificial intelligence through a risk-based approach. Rather than regulating AI uniformly, the Act classifies AI systems according to the degree of risk they pose to health, safety and fundamental rights. AI systems intended for use in the administration of justice are designated as high-risk AI systems, recognising that judicial decision-making directly affects access to justice, procedural fairness and the rule of law. The Act imposes extensive obligations on providers and deployers of such systems, including requirements relating to risk management, high-quality datasets, technical documentation, record-keeping, transparency, human oversight, cybersecurity and post-market monitoring. By treating judicial AI as inherently high-risk, the legislation establishes that efficiency gains cannot outweigh the protection of fundamental rights.

Global Toolkit on Artificial Intelligence and the Rule of Law for the Judiciary[United Nations Educational, Scientific and Cultural Organisation][2023]

Envisioned as a practical guidance framework to help judges and judicial institutions understand how artificial intelligence affects justice systems and how courts should govern AI in ways consistent with democratic constitutional principles. The primary objectives of the report is to improve judicial understanding of AI, equip and train judges on AI foundational principles, identify risks associated to judicial AI and establish governance principles in order to ensure that AI remains compatible with the larger human rights obligation of justice.

European Ethical Charter on the Use of Artificial Intelligence in Judicial Systems and their Environment [European Commission for the Efficiency of Justice][2018]

The charter was one of the 1st international governance framework that specifically dealt with AI tools in the judiciary. The central aim is to ensure technological innovation remains compatible with human rights and rule of law principles by ensuring that AI deployment is not merely efficiency driven, but focuses on fairness, judicial independence and access to justice. Respect for fundamental rights, non-discrimination, security and impartiality are deemed the foundational pillars of Judicial AI tools.

International Initiatives

Judicial Use of Generative AI: Lessons Learned [National Center for State Courts][2026]

An empirical study conducted across US to understand how generative AI tools are being used in judicial practice in order to develop a practical guide for responsible use of AI. The study found that whilst AI could significantly improve efficiency, judges remains skeptical about AI hallucination, privacy and confidentiality concerns and the erosion of human thinking/ skill. Mandatory verification of AI outputs, AI literacy training and restricting AI usage to low-risk administrative tasks is recommended.

AI for Judges [Judicial AI Consortium]

A resource platform that aims to educate judges and the judicial machinery on the safe, ethical and effective incorporation of AI in the judiciary. It centres initiatives to educate judges on how AI tools can support judicial functions whilst preserving due process, fairness and utmost judicial independence. By focusing on AI literacy, increased efficiency by reducing repetitive tasks, necessitating human judgment and formulating institutional governance frameworks, AI tools can be effectively integrated into judicial functioning.

Research that engages with 'Judicial AI'

Smart Courts and Digital Justice: A Governance Framework for Responsible Artificial Intelligence in Judicial Decision-Making[2026]

The research focuses on building governance frameworks that propagate responsible AI usage in courts. It aims to strike a balance between technology integration and preserving institutional values of the judicial machinery. Studying the court documents published between 2018-2026, the aim is to identify how AI is currently being deployed, distinguish between high risk and low risk AI, examination of the various challenges in governance and proposing a framework for governance. In order to examine this, the paper carries out an extensive review of existing literature including:

• Re and Solow-Niederman framing of AI adjudication as a challenge to the values that courts are expected to serve
• Floridi and Cowls ethical framework based on beneficence, non-maleficence, autonomy, justice, and explicability
• Dressel and Farid's study on COMPAS recidivism tool and its limits in accuracy and fairness [5].
• Green and Chen findings that that adding an algorithm to a decision process does not automatically make the result better or fairer. They found that courts should avoid equating prediction with adjudication. Prediction estimates a future event; adjudication requires public reasons based on law, facts, procedure, and context.
• Krištofík's argument that AI-supported judicial decision-making can reproduce older institutional biases through new technology
• Završnik warnings that big-data criminal justice may weaken case-specific narratives.
• Lopes argument that AI in adjudication can affect perceptions of procedural fairness even when it is introduced as a response to human bias.
• Socol de la Osa and Remolina's finding that GenAI can support access and efficiency but can also misinform judicial reasoning through hallucinations, opacity, and weak accountability.

The study proposes a strong recommendation to restrict AI tools to assistive functions in courts rather than involvement in decision-making, it prohibits autonomous adjudications and proposes a system of audits, reviews and testing before deployment of judicial AI tools. Much of the contemporary scholarship on Judicial AI governance increasingly relies upon the European Union AI Act as the benchmark regulatory model owing to its structured risk-based approach. Rather than treating judicial AI as an ordinary technological innovation, the Act recognises AI deployed in judicial administration as a high-risk application affecting fundamental rights, thereby requiring heightened standards of transparency, accountability and human oversight. This legislative approach has significantly influenced subsequent academic proposals on judicial AI governance.

Access to Justice in the Age of AI: Evidence from U.S. Federal Courts [2026]

This paper examines how generative AI has transformed access to the United States federal civil justice system. Using administrative data from over 4.5 million federal civil cases (FY2005–FY2026), the authors find a marked increase in pro se litigation following the widespread adoption of generative AI. While AI appears to lower barriers to filing lawsuits by enabling litigants to draft legal documents without counsel, it also places greater pressure on courts, as pro se cases generate substantially more docket activity and do not conclude more quickly. The authors further validate growing AI adoption by showing that AI-generated text appears in an increasing share of court filings, rising from virtually none before 2023 to over 18% of sampled complaints in 2026. The paper concludes that generative AI has expanded access to justice but simultaneously increased judicial workload, highlighting the need for procedural reforms to manage AI-enabled litigation.

Persuadability and LLMs as Legal Decision Tools [2026]

The research primarily focuses on testing the presumed neutrality of AI adjudicators against varying degrees of persuasion in arguments. The research reflected that majority of the LLMs were significantly persuadable and that AI cannot be presumed as objective adjudicators. If proper guardrails are not imposed, AI could worsen the inequality, inaccuracy and further limit access to justice. The AI tools slated to be deployed in the judiciary ought to be tested for persuadability and assessed for robustness against manipulation in addition to legal accuracy.

AI as a tool to facilitate the presentation of evidence in proceedings? [2026]

The article examines the the feasibility of integrating AI into evidentiary process to improve collection, organisation, analysis and presentation of evidence in litigation and ancillary adjudication mechanisms. It explores the ethical and procedural limitations of placing reliance on AI for evidence processing and deliberates as to whether efficiency of the process can be improved without compromising on fairness, reliability and procedural justice. Preserving procedural fairness is deemed dependent on verification mechanisms, human evaluation and strict authentication/ regulatory standards.

Increasing Use of AI Across the Justice System [2025]

The recently established Oxford Institute of Technology and Justice represents one of the first dedicated global research centres examining the intersection between artificial intelligence, technology and justice systems. Unlike traditional legal AI research that primarily focuses on automation or efficiency, the Institute approaches Judicial AI through a rights-based framework centred upon fair trial rights, due process and judicial independence. Its work is organised around three interconnected pillars:

• Access – examining how AI may responsibly improve access to justice through legal information systems and digital legal assistance;
• Accountability – studying digital evidence, cyber accountability and mechanisms for ensuring technological accountability within justice systems; and
• Advancement – developing international standards for AI-assisted court proceedings while preserving human rights and the right to a fair trial.

One of the Institute's most significant initiatives is the AI Justice Atlas, which comparatively maps the deployment and regulation of AI in criminal proceedings across jurisdictions. Rather than advocating technological adoption, the Institute emphasises governance mechanisms that ensure AI remains compatible with constitutional guarantees, transparency and judicial accountability. These principles closely align with emerging Indian policy initiatives, particularly the Supreme Court Draft Regulations and the UNDP's risk-based governance framework.

Rudimentary AI in Judicial Decision-Making: Lessons From Colombia [2025]

The paper argues that generative AI should be viewed as an intellectual assistant rather than a judicial substitute. Drawing on the author's experience as a magistrate of Colombia's Special Jurisdiction for Peace (JEP), he explains that AI improved judicial work by assisting with legal research, comparative analysis, drafting, interdisciplinary learning, maintaining institutional consistency, and reducing time spent on repetitive tasks. Importantly, every AI-generated output required independent judicial verification and critical reasoning. The paper contends that responsible AI use enhances not replaces judicial deliberation and can strengthen access to justice, particularly in complex transitional justice settings. The paper highlights that judges have a professional duty to understand and responsibly use AI while retaining full accountability for legal reasoning and final decisions.

The Perils and Promises of Artificial Intelligence in Criminal Sentencing [2024]

The paper undertakes a risk-benefit analysis of using AI tools in criminal sentencing, more specifically the veracity of its risk assessment decisions. The deployment of AI tools in systems where the users do not understand its algorithm or working is said to derail the transparency and reliability of the decision making process. It is observed that AI in judiciary, especially one that can deliver proper sentencing scores, can significantly reduce disparity and arbitrariness in decision making, improve process efficiency. However, the research also highlights the threat of opacity and bias in AI decision making and proposes guardrails to ensure fairness such as human oversight, transparency, accountability and most importantly, a provision to challenge the automated decision making.

Judicial AI tools in use

India

Supreme Court Vidhik Anuvaad Software (SUVAS)

The Artificial Intelligence Committee of the Supreme Court developed SUVAS as a translation tool that provides vernacular translations of Supreme Court judgments in the official languages of the country in order to ensure inclusivity and access to justice. The deployment and usage of this tool is consistent with the permitted uses of AI in judiciary as per the draft regulations provided that there exists verification of the accuracy and fidelity of the translated version.

Supreme Court Portal for Assistance in Court Efficiency (SUPACE)

The tool generates concise case summaries and highlights the relevant issues to be considered by analysing the repository of judgments, documents and ancillary records. The tool also provides the judges with the key precedents in the relevant subject matter thereby significantly reducing time spent on research and review of documents. It is the automation of the administrative and fact-finding process without overstepping into decision-making or legal reasoning of the judicial authorities.

Technology Enabled RESolution (TERES)

TERES is a transcription tool used by the Supreme Court to capture and transcribe oral arguments in real time. This tool was particularly developed in order to be deployed before Constitution Benches, and such transcriptions are subsequently uploaded in the Court's website enabling enhanced public access to the judicial process.

Legal Research Analysis Assistant (LegRAA)

LegRAA was developed under the guidance of the SC's eCommittee to provide assistance to the judges in legal research, analysis of documents and support judicial decision-making. The tool, which has a repository of over 36,000 Supreme Court judgments is being used to distill the factual matrix, identify relevant precedents and legislative developments.

International

Intelligent Adjudication System [China]

One of the earliest known uses of Generative AI directly into judicial work, the Intelligence Adjudication System was trained on Chinese laws, judgments and academic legal literature in order to mimic and support judicial decision making with solid legal reasoning. The concern of bias amplification and hallucination remained with no overt mechanisms in place to avoid it. Though the ultimate decision making lies with a human, studies observe that it may not be sufficient to eliminate governance risks.

VICTOR, SIGMA, ATHOS [Brazil]

VICTOR was launched by the Brazilian Supreme Federal Court in 2018 to assist in handling of the high volume of cases by implementing a constitutional filtering mechanism that determines whether the Supreme Court ought to hear a case or not. SIGMA is a judicial assistance tool that automates repetitive adjudicatory tasks such as precedent identification, document filtering and generating decisional templates. ATHOS is a tool designed to detect repetitive litigations involving similar subject matter, especially those that require standardised treatment. It was developed and deployed in an attempt to improve the consistency of judicial decision making.

Aisha [UAE]

Aisha was developed by the UAE Ministry of Justice as its first AI-powered virtual judicial employee in order to support judicial operations. The functioning of the tool was designed in a way to enable assistance to all the stakeholders in the judicial infrastructure including litigants, lawyers and judges. The tool was fed with millions of past legal cases in order to support judges in identifying relevant legal precedents, extracting comparable historical rulings and assisting in the decision making process.

Challenges and Way Forward

The various discussions and reports signal significant challenges in the effective incorporation of AI tools in judiciary. Post-deployment analysis and considerations may prove futile if prior risk assessment mechanisms do not exist. The Indian judiciary has increasingly incorporated AI tools within its justice delivery machinery, warranting a transition from the existing fragmented AI frameworks towards holistic AI Governance framework for Judiciary. The most immediate challenges to judicial AI remain:

1. Absence of a Uniform Definition of "Judicial AI"

Perhaps the most fundamental challenge is that Judicial AI itself remains an undefined legal and regulatory concept. While the Supreme Court Draft Regulations, the Gujarat High Court Policy and the Kerala High Court Policy regulate the use of AI within courts, none of them expressly define what constitutes "Judicial AI" as a separate legal category. Instead, they define "Artificial Intelligence" and regulate its contextual use within judicial administration.

This conceptual ambiguity has significant governance consequences. Without a precise definition, it becomes difficult to determine:

• which technologies qualify as Judicial AI;
• whether ordinary automation software should be regulated alongside generative AI;
• whether Large Language Models, predictive analytics, recommender systems and translation software should all be subjected to identical safeguards;
• the level of judicial oversight required for different categories of AI.

Judicial AI should therefore not be understood merely as AI deployed inside courts. Rather, it ought to refer to AI systems that directly influence judicial administration, legal reasoning, evidentiary assessment or adjudicatory processes. A functional definition distinguishing AI according to its impact upon judicial functions would create regulatory certainty and enable proportionate governance.

2. Failure to Categorise Different Types of AI Systems

Most existing judicial policies proceed on the assumption that AI constitutes a single technology. In reality, AI systems possess vastly different capabilities and corresponding risks. For example:

• Optical Character Recognition (OCR) merely digitises documents.
• Machine learning classification models organise files.
• Predictive analytics estimate probabilities.
• Large Language Models generate natural language.
• Generative AI creates entirely new text, summaries or legal reasoning.

Treating these systems identically ignores their differing capacity to influence judicial reasoning. Future judicial governance should therefore classify AI according to both technical architecture and legal function, for example:

• Administrative AI
• Research AI
• Decision-support AI
• Generative AI (LLMs)
• Predictive AI

Each category warrants different standards of explainability, transparency, auditability and human oversight. The Draft Regulations already separately define Generative AI, Large Language Models, Machine Learning, Risk Scoring and Algorithmic Decision Making, providing an initial taxonomy that could be expanded into a complete governance framework.

3. Ambiguities within the existing AI Policies

Although the Gujarat High Court Policy represents one of India's most comprehensive judicial AI policies, several governance gaps remain. The policy successfully distinguishes between permitted and prohibited uses of AI and reiterates that AI cannot replace judicial decision-making. However, it leaves unanswered several practical questions:

• What degree of human verification satisfies the requirement of "active oversight"?
• Are judges required to disclose AI-assisted drafting?
• How should AI-generated hallucinations discovered after pronouncement be addressed?
• Which authority audits AI tools before deployment?
• What standards govern procurement of third-party AI systems?
• What liability follows if judicial officers rely upon incorrect AI outputs?

Similarly, while legal research assistance is permitted, the policy does not distinguish between conventional search engines and generative AI capable of producing entirely new legal analysis. This omission becomes increasingly problematic as judges begin using foundation models that generate persuasive but potentially fabricated reasoning. The policies therefore requires supplementary procedural standards governing verification, disclosure obligations, audit mechanisms and institutional accountability.

4. Legislative Fragmentation

India presently regulates judicial AI primarily through judicial policies rather than legislative enactments. The Supreme Court Draft Regulations, Gujarat Policy and Kerala Policy collectively establish valuable safeguards, yet none possess the force of legislation. Consequently different High Courts may adopt inconsistent standards, the procurement practices vary considerably, accountability mechanisms remain unclear and the rights available to litigants differ across jurisdictions.

Unlike the European Union AI Act, India lacks a comprehensive statutory framework regulating high-risk AI systems within justice administration. Future legislation should therefore provide:

• statutory definitions
• nationally applicable governance principles
• mandatory algorithmic impact assessments
• independent auditing obligations
• vendor accountability
• appeal mechanisms against AI-assisted administrative decisions.

5. Unclear Regulatory Objectives

Current judicial policies largely identify prohibited conduct without clearly articulating the objectives of Judicial AI governance. A governance framework should explicitly identify its objectives before regulating technology.

The main objectives of such regulation ought to include:

• preserving judicial independence
• protecting procedural fairness
• strengthening public confidence
• improving access to justice
• increasing administrative efficiency
• safeguarding privacy
• ensuring explainability
• preventing discrimination
• maintaining accountability

Clearly identifying these objectives enables regulators to assess whether a proposed AI deployment actually advances constitutional values rather than merely improving efficiency.

6. Risk Scoring and Predictive AI in Criminal Justice

One of the most controversial uses of AI globally concerns predictive risk assessment. The Supreme Court Draft Regulations recognise "Risk Scoring" as AI systems estimating the likelihood of future criminal behaviour, recidivism or failure to appear before court. The controversy surrounding the COMPAS algorithm demonstrates that predictive systems frequently reproduce historical inequalities while creating an appearance of scientific objectivity. Research on such predictive algorithms note that prediction should never be equated with adjudication and there exists empirical concerns regarding fairness, opacity and bias in such systems. Specifically, in India, the severe judicial backlog may incentivise increased reliance on AI-generated summaries, precedents and reasoning and lead to erosion of independent judicial reasoning/ discretion.

Judicial sentencing requires proportionality, individualised assessment, consideration of mitigating circumstances and constitutional due process. Risk prediction instead relies upon statistical correlations derived from historical data, which may reflect entrenched social biases relating to caste, religion, gender or socio-economic background. The concern is not merely algorithmic bias but the transformation of sentencing from an exercise in legal reasoning into probabilistic forecasting.

Accordingly, India should considering adopting a prohibitory framework on deploying predictive AI for:

• criminal sentencing
• bail recommendations
• parole determinations
• preventive detention

AI may assist courts by organising information or identifying relevant precedents, but it should not be used to estimate future criminal conduct or influence punishment based upon statistical predictions. The European Union AI Act further strengthens this position by identifying certain AI systems involving social scoring and specific predictive uses in law enforcement and justice as unacceptable or strictly regulated practices. This reflects a broader international movement away from algorithmic prediction of future behaviour towards preserving individualized judicial assessment based upon evidence and legal reasoning. India's future Judicial AI framework should adopt a similarly cautious approach by expressly prohibiting AI-based criminal sentencing, recidivism prediction and behavioural risk scoring.

7. Over-Reliance on Large Language Models

Large Language Models introduce risks distinct from conventional automation. Unlike deterministic legal databases, LLMs generate probabilistic text based upon training data. Their persuasive language, fabricated citations, hallucinations and opaque reasoning may create automation bias whereby judges unconsciously accord greater credibility to AI-generated analysis. Emerging research further demonstrates that LLMs remain susceptible to manipulation through prompt engineering and persuasive framing, undermining assumptions of neutrality. Consequently, LLMs should never function as autonomous legal reasoning engines. Their permissible role should remain confined to summarisation, document organisation, language translation and legal research assistance subject to mandatory human verification.

United Nations Development Programme, Regional Bureau for Asia and the Pacific, Responsible AI in Justice: Regional Policy Guidance for UNDP Programming in Asia-Pacific (UNDP RBAP 2026).

8. Data Privacy and Confidentiality

The incorporation of AI into judicial administration inevitably involves the processing of highly sensitive judicial data, including pleadings, witness statements, medical records, commercially confidential information, personally identifiable information, and documents relating to national security and criminal investigations. Unlike conventional digital case management systems, many contemporary AI tools, particularly cloud-based Large Language Models process data through external servers, creating heightened risks of unauthorised access, secondary use of judicial information, cross-border data transfers, and cyberattacks.

The confidentiality of judicial proceedings forms an indispensable component of the right to privacy and fair trial, making judicial datasets substantially more sensitive than ordinary governmental records. Although the Supreme Court Draft Regulations recognise concepts such as sensitive judicial data, anonymisation, and data minimisation, comprehensive technical standards governing storage, localisation, encryption, retention periods, and vendor access remain underdeveloped. Consequently, the deployment of Judicial AI necessitates stringent data governance mechanisms that prioritise privacy by design, secure domestic infrastructure, strict access controls, and continuous cybersecurity audits to preserve public confidence in the justice delivery system.

9. Opacity and Explainability of AI Systems

Judicial legitimacy is fundamentally premised upon reasoned decision-making that is transparent, reviewable and capable of being scrutinised through appellate mechanisms. The increasing deployment of sophisticated AI models, particularly Large Language Models and other black-box systems, presents a significant challenge to these foundational principles. Where the internal reasoning of an AI system cannot be meaningfully explained, judicial officers may become unable to verify how particular recommendations, summaries or legal analyses were generated. This creates a tension between algorithmic opacity and the constitutional requirement that judicial decisions must be supported by intelligible reasons. Recognising this concern, the Supreme Court Draft Regulations expressly define both "Black Box" AI systems and "Explainability", requiring that AI-generated reasoning be comprehensible to judges, court staff and litigants without specialised technical knowledge. However, many generative AI systems remain inherently opaque despite advances in explainable AI.

10. Digital Infrastructure and Institutional Readiness

The successful integration of Judicial AI depends not only upon regulatory safeguards but also upon the technological readiness of judicial institutions. While pilot initiatives such as SUVAS, SUPACE, TERES and LegRAA demonstrate the judiciary's commitment to technological modernisation, the digital infrastructure across Indian courts remains uneven. Many subordinate courts continue to experience inconsistent digitisation of records, inadequate computing infrastructure, limited internet connectivity, shortages of technical personnel, and varying levels of digital literacy among judicial officers and court staff.

The introduction of advanced AI systems without addressing these structural deficiencies risks creating fragmented implementation, inconsistent standards of use, and unequal access to technological benefits across jurisdictions. International governance models similarly recommend that institutional readiness be assessed before AI deployment, with implementation proceeding only where courts satisfy minimum technical and organisational benchmarks. Accordingly, strengthening digital infrastructure, standardising e-court systems, investing in secure domestic computational resources, and providing continuous AI literacy and technical training should precede the large-scale deployment of Judicial AI across the Indian judiciary.

11. Ethical Problem of Automation Bias

Even where AI is formally restricted to an assistive role, judges may consciously or subconsciously defer to AI-generated recommendations because they are perceived as objective or technologically superior. This phenomenon, known as automation bias, risks shifting judicial reasoning from independent deliberation towards confirmation of algorithmic outputs. Over time, excessive reliance on AI-assisted research, summaries and recommendations may diminish critical legal reasoning and weaken the judiciary's constitutional obligation to provide independent and reasoned decisions.

Way Forward

A sustainable framework for Judicial AI in India must move beyond fragmented institutional policies and adopt a comprehensive, risk-based governance model rooted in constitutional values, judicial independence and procedural fairness rather than mere efficiency considerations. The most significant contribution in this regard is the framework proposed by the United Nations Development Programme (UNDP) in its report, AI for Justice: Ethical, Fair and Robust Adoption in India’s Courts (2026), which provides a structured methodology for assessing whether, and under what conditions, AI systems may be deployed in courts.

Unlike traditional technology policies that focus only on post-deployment safeguards, the UNDP model recognises that judicial AI governance must begin before procurement and deployment. It proposes a three-stage evaluation process consisting of:

(i) Institutional Readiness Assessment,

The first stage requires courts to evaluate their institutional preparedness before adopting AI systems. Courts should assess factors such as the extent of digitisation, technological infrastructure, cybersecurity mechanisms, data governance standards, availability of trained personnel, internal oversight systems and judicial AI literacy. Only courts that satisfy minimum readiness benchmarks, which is reflected through a score of at least sixty percent, ought to proceed to subsequent stages of evaluation.

This approach is particularly relevant for India, where disparities continue to exist between different levels of the judiciary in terms of digitisation, technical resources and institutional capacity. AI deployment without adequate readiness risks inconsistent implementation, unequal access and potential harm to litigants. Therefore, strengthening digital infrastructure, standardising e-court systems, establishing secure domestic computing facilities and providing AI literacy programmes for judges and court staff should constitute the first step in judicial AI governance.

(ii) Risk Assessment,

The second stage requires categorisation of AI systems according to the degree of risk they pose to rights, judicial independence and procedural fairness. The UNDP framework proposes that AI tools should not be regulated uniformly; instead, safeguards should correspond to the level of risk associated with the technology.

• Low-risk systems include translation tools, transcription software and document management systems such as SUVAS and TERES.
• Medium-risk systems include legal research assistants, precedent identification tools and summarisation systems such as SUPACE or LegRAA.
• High-risk systems include generative AI systems, recommendation engines, predictive analytics and any tool that influences judicial reasoning, evidentiary assessment, sentencing or outcome prediction.

The report further recognises a category of prohibited AI uses, where risks outweigh potential benefits. India should expressly place predictive AI systems used for criminal sentencing, bail decisions, recidivism prediction and behavioural risk scoring within this category. Such systems transform adjudication into statistical forecasting and undermines the constitutional principles of individualised justice, proportionality and due process.

United Nations Development Programme, Regional Bureau for Asia and the Pacific, Responsible AI in Justice: Regional Policy Guidance for UNDP Programming in Asia-Pacific (UNDP RBAP 2026).

(iii) Technical Assessment and Continuous Monitoring.

The third stage requires detailed technical evaluation before deployment and ongoing supervision after implementation. According to the UNDP framework, medium and high-risk AI systems must undergo rigorous technical assessment to evaluate explainability, accuracy, bias, privacy protections, cybersecurity safeguards and compliance with judicial values before approval for use. Only systems meeting prescribed thresholds ought to be deployed.

More importantly, governance cannot end at deployment. AI systems evolve over time, and their outputs may change depending on new data, updates and user interactions. Continuous monitoring is therefore necessary to identify emerging harms, hallucinations, biases and unintended consequences. Periodic audits, independent technical evaluations, vendor accountability mechanisms and public reporting obligations should form part of an institutional monitoring framework.

This model offers a practical framework capable of balancing innovation with constitutional protections.

Judicial AI Governance

Drawing from the UNDP framework, India should adopt a national Judicial AI Governance Framework that incorporates:

• a statutory definition of Judicial AI;
• categorisation of AI systems according to risk;
• mandatory institutional readiness assessments;
• pre-deployment algorithmic impact assessments;
• independent technical audits;
• strict confidentiality and data governance standards;
• mandatory human oversight;
• continuous post-deployment monitoring; and
• express prohibition of predictive AI in adjudication, sentencing and behavioural risk scoring.

The UNDP framework closely mirrors the regulatory philosophy adopted under the European Union AI Act. Both frameworks reject a one-size-fits-all regulatory model and instead advocate governance calibrated according to the level of risk posed by particular AI systems. While the AI Act classifies judicial AI as a high-risk application subject to stringent compliance obligations, the UNDP proposes institutional readiness assessments, risk assessments and technical evaluations before deployment. Together, these frameworks demonstrate an emerging international consensus that AI in courts requires significantly greater safeguards than AI deployed in ordinary commercial settings.

The framework is further reinforced by the work of the Oxford Institute of Technology and Justice, which conceptualises AI governance not merely as technological regulation but as the preservation of fair trial rights in AI-assisted proceedings. Through its AI Justice Atlas and comparative research on judicial AI, the Institute demonstrates that governance should focus not only on the technical reliability of AI systems but also on safeguarding due process, judicial independence and transparency. Consequently, India's judicial AI framework should be evaluated against constitutional guarantees of access to justice and fair trial rather than solely against measures of efficiency or case disposal.

Ultimately, the integration of AI into the judiciary should not be viewed as a project of technological optimisation alone. Courts are constitutional institutions tasked with preserving rights, ensuring fairness and maintaining public trust. Consequently, AI systems must remain assistive tools that augment judicial administration while leaving legal reasoning, evidentiary appreciation and the final exercise of judicial discretion exclusively in human hands. This approach aligns with the UNDP model and with emerging international consensus that judicial AI must remain subordinate to constitutional principles, human rights and the rule of law.

Contributors & review status

Curated / edited bynot yet recorded

Peer reviewpending

Project reviewpending

Expert reviewpending

Page created 2026-08-21  ·  Last updated 2026-08-23