AI in Legal Processes and Prosecution Preparedness

AI in Digital Evidence Trials and Faster Justice

Authors

Executive Summary

Firms across almost all industries have been attempting to integrate artificial intelligence (AI) into their practices, as there has been demonstrable evidence that there are massive potential gains in terms of streamlining processes to increase efficiency. The Indian legal industry is no different, and despite traditionally being hesitant to adopt change, it has a prime opportunity to work towards developing and integrating new technologies. This policy brief has identified four components of the judicial process where bottlenecks exist: legal filings, the rostering of cases, the adversarial actions that play out in the courtroom and the final judgement.

This brief is an output of the second Hackathon of Law Enforcement and Policing Fellowship 2024-2025.

The authors of this policy brief would like to thank Devina Sikdar, Dr Mohit Nayal, Dr Naveen Kumar Singh, Rahul Dev Sharma, IPS and Dr Syed Mustafa Hashmi (IPS) - fellows of the Law Enforcement and Policing fellowship (LEPF) who had an input on the content of this paper.

Easing these would help alleviate the problem of a year-on-year exponential increase in the backlog of cases, which has led to an unnecessarily traumatic, prolonged trial process, burdening all stakeholders. The proposed solution recommends an incremental augmentation of the trial process through a four-phase plan with extensive use of pilot programmes. This would help ensure the functioning of the proposed technology, because the repercussions of a mistrial could be grave, not just for the individuals involved, but also in terms of the precedent it sets for its future adoption. Thus, the proposed solution would utilise AI to allocate state resources to hasten the judicial process efficiently without compromising the principles of justice.

The Challenge

The term “clearance rate” is used to measure the efficiency of courts. It is the number of cases resolved by a particular court divided by the number of cases instituted at the same court. In 2024, the district subcourts in only sixteen of the thirty-six states and union territories have a case clearance rate that is equal to or greater than 100 per cent. When extended to high courts, the figure drops to thirteen, and for the purposes of this policy brief, cases from union territories that get elevated to their respective high court are considered independent of the cases high courts handle from their respective states. This totals 23.46 million cases disposed of from the 24.46 million cases filed at a subcourt level and 2.26 million out of 2.40 million at the high Court level. This 95.8 per cent clearance rate pushes more than 1 million cases to the next year, an action that has led to a backlog of over 52 million cases. This is an alarming statistic as it demonstrates that the 2024 average backlog rate of 4.2 per cent is almost double the average backlog rate since the end of 2022 (2.37 per cent).

Present Bottlenecks

Transcription and Translation

Every court proceeding in India needs stenographers to note down statements by participants in the trial process. India has very few skilled stenographers who can accurately transcribe spoken words in real-time and without errors, forcing speakers to slow down or even repeat statements so that stenographers can write accurate court records.

Another issue that is raised is that Article 348 of the Constitution mandates the use of English in proceedings unless approved by the President and Chief Justice of India. Supreme Court proceedings are always conducted in English, but the same does not necessarily hold true for high courts and district courts. Presently, only four high courts: Rajasthan (1950), Uttar Pradesh (1969), Madhya Pradesh (1971) and Bihar (1972), have been given permission to conduct proceedings in their regional language (Hindi). However, since then, other high courts that have sought permission have seen their requests turned down. This trend has not trickled down to district courts, as state governments have the prerogative to declare any regional language as an alternative for court proceedings under section 137 of the Code of Civil Procedure. This is problematic when appellate courts do not use the same regional language.

Roster Management

Cases in high courts are allocated to judges by the Chief Justice of the respective high court or according to the Case Flow Management Rules. These exist to ensure that cases do not exceed their given timeframe, but these guidelines are often flouted. A reason for this is that judges, especially at the district level, are given points for disposing of cases and thus have an incentive to clear easier cases first. This leads to parties being called to the court complex on days their case is not even going to be heard, which could lead to the victim getting disheartened and withdrawing their case or witnesses not showing up, as prolonged uncertainty causes disruptions in their daily lives.

Like lawyers, judges have areas of expertise, but there is no guarantee that the presiding judge for a case will be an expert in the dispute in question. The two problems that manifest because of this are that of speed and quality. Judges’ ruling outside their area of expertise would need to spend much longer conducting judicial research for their rulings, and are likely to set precedent in their rulings that are later cancelled. These are addressed in the next section of this policy brief. While there is no quantitative evidence related to India specifically in this matter, a study in Kazakhstan revealed that a judge presiding over a case in their area of expertise is more likely to come to a ruling that sets good precedent 67 per cent of the time and at a rate that is 75 per cent faster than one that is presiding over a case in a legal field they are not as conceptually familiar with.

Judicial Research

In the trial process, judges consider arguments from both parties before delivering a verdict. In some trials, the judge gives an ex tempore verdict, which means the judgement immediately follows the closing arguments instead of being written first. However, there has been an increase in instances of judges taking months or even years to write out a judgement, which leaves the parties in the dark as to the outcome of the trial. Additionally, the writing of judgements takes a significant amount of time, and a judge performing such administrative duties is a judge who is unable to oversee trials.

AI as a Solution

Transcription and Translation

This study proposes an AI tool that would convert speech to text, not just in the courtroom, but also in places where legal filings are initiated. This tool would do away with the need for repetition of speech and would likely be able to record statements accurately. Indian regional courts sometimes use the local language instead of English, which necessitates quality bilingual stenographers, an even rarer commodity. This problem also arises when cases with FIRs lodged at police stations in a certain language get escalated to higher courts that cannot understand the language.

Such a tool was first used in 2023 in Subhash Desai v Principal Secretary, Governor of Maharashtra & Ors. Chief Justice of India, Dr Dhananjaya Y Chandrachud, directed the court to make note of an AI-based transcribing tool that displayed live arguments, sent transcripts to any stakeholders and published them on the Supreme Court’s website. The tool, Teres, was developed by Bangalore-based Nomology Technology Private Limited, and works by feeding audio recorded by already prevalent recording services such as Zoom and Microsoft Teams into its transcribing service with humans monitoring input data and making corrections, if necessary, all at the cost of US $100 per hour of audio. Astute Dispute Resolution, a Geneva-based law firm, has touted Teres’ hybrid system as highly impressive, as there was very “minimal human correction” despite the input audio speech being a “strong French accent” with “pronounced stuttering”. However, in the Indian context, CJI Chandrachud said there were issues caused when there were multiple people speaking at once, but that there were personnel who would “clean up the errors by the evening”. This demonstrates that despite presently displaying a few minor flaws, the tool has practical feasibility in the future as the rigorous training of AI would eliminate errors.

Roster Management

The utility of this tool is not purely restricted to expediting cases that require urgent resolution. AI can be an effective tool for roster management purposes as it could reduce dependence on the present allocation system and Case Flow Management Rules, which have been previously mentioned as being flawed. Without human bias present, cases can proceed to hearing in a more efficient manner. The tool can be used to issue particular cases to judges who are experienced in the specific law related to the issue of the dispute. Similarly, this can be used to manage workloads as the software would not assign cases that it forecasts to be drawn out to judges with upcoming transfers. This would help eliminate the delays caused by new judges having to be reacquainted with the facts of the case.

The proposed tool would analyse all relevant variables used in case roster preparation, such as the number of witnesses, the nature and complexity of the case, travel requirements for parties involved, and other logistical considerations to generate an optimised hearing schedule. The output should remain strictly advisory, with the final decision on whether to adopt the suggested timetable resting entirely with the Master of the Roster, the judge who decides case allotment. Given that court hearings are inherently dynamic and subject to unforeseen disruptions, the tool must be capable of incorporating real-time deviations from the planned schedule and recalculating future hearing dates accordingly. Such adaptive rescheduling would help maintain overall efficiency while preserving judicial discretion and flexibility.

A forecast function can also play a role in informing all the stakeholders in a given case about a tentative timeline of the proceedings. This is especially crucial for witnesses who, as previously mentioned, tend to get discouraged and skip court appearances because of a lack of clarity as to the days they are required to appear.

Judicial Research

Judicial research is a critical component of the judicial process, enabling judges to interpret and apply the law accurately. It is the duty of the judiciary to ensure that legal interpretation remains consistent across all levels of the court system. Under the principle of judicial hierarchy, lower courts are bound by the decisions of higher courts, making it essential for judges to identify and apply relevant precedents. Accordingly, a judge presiding over a case must thoroughly examine past rulings in similar matters to support their reasoning and ensure that their judgment aligns with established legal principles.

However, this research process can be extensive and time-consuming. When carried out manually, navigating the vast repository of past judgments poses a significant challenge, particularly in complex cases. The sheer volume of case law increases the risk of overlooking critical precedents, which may lead to inconsistencies in the application of law. Moreover, manual research is prone to cognitive biases such as confirmation bias or selective reading, potentially affecting the objectivity of judicial reasoning.

AI is already being used in the Supreme Court to some extent, but its adoption remains limited and fragmented. To ensure consistency and wider impact, these capabilities should be extended to lower courts through a unified tool that can be implemented consistently across all levels of the judiciary. For instance, the Supreme Court has introduced tools like SUPACE (Supreme Court Portal for Assistance in Court Efficiency), which helps judges analyse facts and relevant precedents for complex matters, and Supreme Court Vidhik Anuvaad Service (SUVAS), which supports translation of judgments across Indian languages. While promising, these tools are still in early stages and have not yet been scaled to district or subordinate courts, where they could have the greatest impact in reducing pendency and improving access to justice.

One of the key strengths of AI is that it can be trained on vast amounts of legal information, and when asked a question, it can understand the context based on its training and generate a response that is relevant, accurate, and complete. This ability can be used to build tools that support judges in carrying out judicial research more efficiently. An AI system trained on millions of past judgments can quickly find and present all relevant cases, helping judges avoid missing important precedents. It also reduces the time spent on research and ensures that decisions are based on comprehensive and objective information. With such tools, judges can write judgments faster and with greater confidence in the legal foundation of their rulings.

Concerns and Mitigation

Accountability and Accuracy

One key caveat in the use of AI in judicial processes is the need for complete accuracy and explainability. AI tools must not generate fabricated or unverifiable information, as even a single instance of such an error could have serious consequences, not only for the outcome of the case but also for the public’s trust in the judicial system. Ensuring that all outputs are traceable to actual legal sources and the case facts is therefore critical.

Given the current capabilities and limitations of AI, its role in the judiciary should be narrowly defined. For judicial research, the tool must be designed solely to retrieve relevant case records from an authenticated database, without attempting to generate new legal reasoning or conclusions. For legal filings, it should generate charge sheets and plaints based on actual evidence and facts provided. To avoid the risk of hallucinating non-existent information, a known issue with general-purpose AI models, the system should include strict safeguards that constrain its responses entirely to verifiable legal texts. This level of control is essential to maintain both the accuracy of judgments and public trust in the judicial process.

Another common criticism of AI is that it can amplify existing biases present in the data on which it is trained. This concern is relevant to the judicial research tool as well, given that historical case data may reflect systemic or unconscious biases in past judgments. Identifying and mitigating such biases is a complex and resource-intensive task. One possible approach could be to commission an independent study to analyse historical case patterns, with AI itself being used to identify such patterns. However, a detailed exploration of this issue and its potential solutions is beyond the scope of this brief.

Data Gathering

One major challenge in developing an AI tool for judicial research or for automated legal filing is the creation of a comprehensive and standardised database of historical case records for its training. The National Judicial Data Grid (NJDG), established under the e-Courts Phase I initiative, provides access to case data dating back to 2011. Additionally, the Supreme Court and some high courts maintain their own digital archives, though these are not uniformly structured or standardised. Several private legal databases also offer extensive case law collections. Before an AI model can be effectively built and trained, a critical first step is the collection, collation, cleaning, and standardisation of case data from these diverse sources.

Human Training

Presently, only 32.6 per cent of students and professionals in legal and related fields are “very familiar” with the use of AI. This lack of knowledge can only be addressed by educational courses and would require present stakeholders to undergo training to understand the methods by which they can leverage AI tools to augment their work. This reform would affect stakeholders ranging from law students looking to enter the industry right up to seasoned veterans who have spent years in the legal world and might be used to archaic methods of working, where a lot of documents were handwritten. These manual methods discount the efficiency gains brought about by the integration of new technology.

However, it is not possible to determine if this is a pressing issue or one that will solve itself as time passes, as a different study found that 70 per cent of Indian legal professionals are open to the integration of AI in their practice. It is also likely that they will be keen to integrate AI in their practices as it would increase productivity by 13 per cent. Some of the steps they could take are the creation of guidebooks and workshops for their employees, as this would help them get ahead of the curve. It is also recommended that all stakeholders in the legal process that fall under the banner of the legislation undergo rigorous cybersecurity hygiene training, as data scams through online mediums are becoming increasingly prevalent today.

Data Security

The Indian legal framework that governs AI has been subject to extreme scrutiny. The two pertinent pieces of legislation are the Information Technology Act of 2000 and the Digital Personal Data Protection Act (DPDP) introduced in 2023. Some of the concerns raised include the lack of explicit legislation targeting AI, especially automating decision-making and model accountability. Section 7 of the DPDP, for example, states that even when a person denies consent for the collection of their data, firms can still process the data if they can justify using listed items, such as in the “interests of sovereignty”. Such ambiguous phrasing has led to concerns amongst consumers. An example of this could be that information from sensitive trials will reach domains that it isn’t supposed to reach, which makes it accessible to malicious third parties. However, statewide mandates such as the High Court of Kerala’s ‘Policy Regarding Use of Artificial Intelligence Tools in District Judiciary’ which require every single judicial worker to “meticulously” verify any content that has been produced with AI.

The primary method by which India could address these concerns is by introducing new legislation or amending the Digital Personal Data Protection Act, so its scope includes firms that develop AI tools or the firms that store data liable for any potential breaches. This incentivises them to ensure that steps are taken to minimise any chances of a breach. An alternate method of mitigating this concern is by introducing legislation that mandates that firms that develop the AI and firms that collect data from these trials should be separate entities that have no links to each other. This way, data can be scrambled when sent to servers, with only the judiciary having a decryption key.

Unauthorised Access and Tampering of Records

Inadequate security of backend databases or AI-generated outputs can result in sensitive legal data falling into the hands of malicious actors. Even archived case data stored on servers could be vulnerable to tampering, which may in turn alter AI research results, translations, or other judicial processes that rely on such data. To address this risk, judicial AI systems should implement robust role-based access controls and multi-factor authentication for all users, along with mechanisms to detect alterations in stored data. Regular data security audits and vulnerability assessments should be mandated to identify and address potential weaknesses before they can be exploited.

Data Sovereignty

Many AI tools are built on open-source large language models that may process data on servers located outside India. This raises concerns regarding compliance with domestic data protection laws and the potential for foreign jurisdiction over sensitive judicial information. While open source models are great for data sovereignty, to mitigate this data security risk, all judicial AI tools should be hosted on secure, government-controlled servers within India. Strict compliance with the Digital Personal Data Protection Act and other relevant legal frameworks must be ensured, and the routing of sensitive judicial data to offshore processing environments should be expressly prohibited.

Insecure Data Transmission

Data transmitted over insecure networks is vulnerable to interception during upload, processing, or download, exposing sensitive legal documents to unauthorised access. This risk can be addressed by enforcing end-to-end encryption for all data in transit, using secure communication protocols. In addition, judicial AI tools should operate over a dedicated VPN or within a private judicial network infrastructure to further safeguard data transmissions from potential interception or compromise.

Third Party Vendor Risks

When AI development, training, or hosting is outsourced to third-party vendors, there is a risk that vendor personnel may gain unauthorised access to sensitive judicial data. To prevent such incidents, all vendor contracts must include strict confidentiality and data-handling clauses. Furthermore, vendor employees with access to judicial data should undergo comprehensive security vetting and background checks. Technical safeguards such as access logging, real-time monitoring, and anomaly detection should be implemented to ensure any unauthorised activity is detected and acted upon promptly.

Integration With Other Systems

Judicial AI tools often integrate with other digital systems such as police databases, the e-Courts platform, or evidence management systems. Such integrations can create cross-system vulnerabilities, where a breach in one system could compromise the security of the entire judicial data ecosystem. This risk can be mitigated through comprehensive security audits of all integrated systems prior to deployment, strict API access controls, and limiting data sharing to the minimum necessary. Network segmentation should also be employed to isolate systems and contain the impact of any breach.

Lack of Clear Accountability

In the event of a data breach involving an AI tool, determining liability between the judiciary, vendors, and hosting providers can be complex, potentially delaying corrective measures. To avoid such ambiguity, a clear governance framework should be established, assigning responsibility for each stage of data handling. Detailed audit logs must be maintained for all system interactions and data access, and incident response protocols should be predefined. All contracts should also include explicit liability clauses to ensure swift and accountable action in the event of a security incident.

Changing Contemporary Perceptions

Only 4.1 per cent of stakeholders in the Indian legal process fully trusted AI, and just under half of the people trust the integration of AI with human oversight. This leaves a majority of people who do not trust the addition of AI to the judicial process. It is integral to change this perception of the utility of AI amongst the majority if it is to be accepted as a tool in the judicial process. Furthermore, as per a 2022 report, only 41 per cent of district court complexes have the hardware that enables studio-based video calling. As Teres utilises this technology, the adoption of AI is dependent on the construction of these facilities at the district court level. The concerns about the use of AI tools also exist at a judicial level, as demonstrated by the High Court of Kerala, which issued an official memorandum outlining its stance.

As already mentioned in this paper, the Supreme Court of India has already taken baby steps by attempting to integrate AI in some of its hearings. Efforts have already begun from the top, as the Supreme Court of India has urged High Courts to follow suit and adopt AI in its proceedings as well. So far, only the High Court of Delhi has attempted to do this. While the Kerala memorandum does not explicitly outlaw the use of AI, it states that any use of AI by law interns and clerks right up to the actors in the courtroom must be “meticulously verified” by judicial officers, which is a practical impossibility. It says that “most AI tools produce erroneous, incomplete or biased results”, which is a generalisation that might hold water when applied to infant models of such tools, but contradicts the modern findings about tools such as Teres. Thus, educational campaigns and workshops that correct such misinformed stances could be effective in adding weight to the argument that supports the implementation of AI in the legal process. It is also discussed in the policy recommendation section as to why a phased integration of AI into judicial processes could be beneficial to all stakeholders.

Cost

There are five broad categories across which costs can be spread. They are as follows: - Pay for members of each of the three committees, - Development of the tool, - Building digital infrastructure in all courts across the country to support these new tools, - Training of stakeholders about the utility of the tools, - Maintenance of the tools and servers that store the information

Bridging the digital infrastructure gap in terms of the provision of software and hardware, as well as the training of humans, is probably going to cost the most money as both are revolutionary changes to the present system. However, it is important to note that these are one-time costs. On the other hand, once the basic groundwork is set, the costs associated with maintaining such a system will recur. As the integration of all proposed tools is unprecedented, the onus will fall on the government to cover costs in terms of ensuring funds are not cut amid the integration process. There has been a lot of support for the digitisation of the legal system over the past decade, with Rs. 7,210 crore being allocated for the four-year period commencing 2023, indicating that the government is likely to be willing to bear the costs of such a project.

Policy Recommendation

This policy brief recommends the staggered introduction of AI into the judicial process in three phases, with a conservative timeline of ten years, despite certain technology already being available, as getting relevant stakeholders on board will take time. However, this is not a hard timeline recommendation, as achieving these goals in a shorter time frame is not realistically inconceivable.

Phase Zero

This phase consists of three smaller steps. The first would be the creation of a nationwide policy that outlines the method by which AI tools will be added to the judicial process. This commences with the appointment of a three-tiered action committee, starting with an oversight body to set the broad guidelines and vision, a steering committee that oversees implementation and a task force that takes care of the implementation of tools at the ground level. The composition of the three-tiered action committee with various stakeholders, practising and retired, would help set up these tools in terms of data collection and training. While it is difficult to alleviate all concerns about AI hallucinations and erroneous generations, the use of such an expert panel would reduce mistakes.

The second step would entail first providing hardware to facilitate video conferencing, especially at district-level courts, before moving on to the humans behind the machines. With only 33.6 per cent of legal professionals in India being unfamiliar with the use of technology in general, extensive workshops must be conducted to demonstrate and encourage this minority to augment their work with AI tools. This is also where law school curricula must be updated to make sure future legal professionals enter the industry with a strong understanding of available tools.

The next step is to identify a vendor to build the AI tools through a tender process. Considering the technical complexity and the sensitivity of the data and applications, a public-private partnership could be the way to go. A vendor shall have to be selected to build the applications, while the sole ownership of the data and the tools, once built, must remain with the judiciary.

One of the key activities in this phase is to gather, clean and standardise historical case data, which will then be used to train the AI models. This could be an extensive process requiring the cooperation of multiple stakeholders. A team of legal and tech experts must work with public and private sources of legal data to collect and prepare the data in the required format. Privacy and secured storage of this vast amount of data is of the utmost importance.

Phase One

The first phase would constitute the rollout of the tool for translation and transcription services in courts across the nation. As this technology is already available in some courts, it is likely that this would be the easiest of the three proposed tools to transition into widespread use across the legal system, as it is minimally invasive in the grand scheme of the judicial process.

This phase would also include introducing pilot programmes for AI tools that aid in legal filings in certain districts or states to test the viability and accuracy of such technology. It is imperative that the technology is implemented on a trial basis first, as this is the first tool that truly transforms the legal process instead of just augmenting the skills of human stakeholders. Despite the utility and massive productivity gains of such a tool, this policy brief has already recognised the hesitancy to accept transformative technology and recommends a trial programme to spread awareness about the upsides of this tool.

Phase Two

Phase two begins with an analysis of how well the transcription and legal filing tools are working. If there is demonstrable evidence that there are flaws in the functioning of the tools, it would be advisable to halt the use of these tools in any legal proceedings until the technology behind the tools is corrected and reapproved by the steering and overseeing committees. If the pilot programmes prove that these tools are working as intended, the transcription tool can continue to be used, and the status of the legal filing tool can be elevated from pilot programme to full-scale implementation.

This would entail making the legal filing tool available at all police stations and civil courts in the nation, as these are the locations at which complaints are lodged. As mentioned before, the legal community will have to be convinced about the utility of AI and, in terms of transforming the process, the roster management and judicial research tools are the most invasive. As there is already a level of trust that has been built thanks to the success of the other two phases, it is likely that the stakeholders will be willing to adopt the roster management and judicial research tools.

Phase Three

The final phase would consist of complete implementation and ensuring that the proposed tools are working harmoniously and are actually improving the efficiency of the Indian legal system in a way that does not compromise justice. The three-tiered committee system should not be disbanded, but its work should be changed to ensure the proper functioning of the reformed system. Once the systems are proven to be successful, the committees can look to integrate these systems with other law enforcement stakeholders. Extensive studies should also be conducted to determine if there is a quantifiable increase in courts’ clearance rates, which can only be done by the collection of feedback information and retraining of AI. While clearance rates are an extremely simplistic statistic that does not consider the complexity of each dispute, it is still the best metric to see if the system is working, as over the long run, the types of dispute that arise are the same.