Artificial intelligence in judicial systems: promises and pitfalls

Artificial intelligence (AI) is viewed as the hope for humanity by many and as its eventual destroyer by others. Attitudes regarding AI use in judicial systems are similarly divided. However, despite legitimate concerns, AI is already being employed by courts, judges, prosecutors and lawyers, sometimes with official sanction and guidance, and sometimes by individuals on an ad hoc basis.

As scholars have observed, the term “artificial intelligence” is “less a technically precise descriptor than an aspirational project that comprises a growing collection of data-centric technologies”.² AI refers to a “constellation” of processes and technologies enabling computers to complement or replace specific tasks otherwise performed by humans.³ In the present report, consideration is given to all forms of AI, including predictive AI (which involves using statistical analysis and machine learning to identify patterns, anticipate behaviours and forecast upcoming events)⁴ and generative AI (which can create original content, such as text, images, video, audio or software code in response to a user’s prompt or request).⁵ It is recognized that different risks arise based on the type of technology used, but the associated risks also depend on other factors, including the specific task that AI is tasked to perform and the manner in which the technology has been developed, procured and deployed.

2. Meredith Whittaker and Lucy Suchman, “The myth of Artificial Intelligence”, The American Prospect, 8 December 2021. ↩
3. A/73/348, para. 3. ↩
4. Tim Mucci, “What is predictive AI?”, IBM, 12 August 2024. ↩
5. Cole Stryker and Mark Scapicchio, “What is generative AI?”, IBM, 22 March 2024. ↩

States express a desire to employ AI as a solution to numerous problems, including in judicial systems.⁶ The current justice architecture is far from perfect, and new approaches are needed.⁷ However, as with all technologies, AI use must comply with human rights. AI is valuable only when it improves justice in concrete ways; it should not be pursued as an end in itself. Those considering using AI should assess which justice problems it can help to solve, whether it is fit for purpose and when it may make things worse. When do the impacts of AI deployment, in particular its environmental impacts, outweigh the benefits? Climate-related and other types of harm must be assessed, and the experiences of justice system actors and those experiencing justice problems should be taken into account, recognizing that these are often not the same populations. Importantly, human rights can only be protected if human beings – and particularly human judges – have realistic and clearly defined responsibilities in relation to AI tools.

6. Submissions from Latvia, Morocco, the Bolivarian Republic of Venezuela, and the Attorney General’s Office of Uruguay. ↩
7. See A/78/171. ↩

States should be aware that AI use entails a concentration – and often a transfer – of power. AI technology is focused in a minority of States and a handful of companies. It is unlikely that this unequal landscape will be brought into equilibrium in the near future, due to the vast data sets, enormous computing power and access to semiconductor chips and energy required to build and run advanced AI models. Another kind of power is also concentrated as AI use advances, namely, epistemological power, or the capacity to define what counts as knowledge in a given field. As States increasingly rely on AI, power will become more condensed and inequalities will grow. 1 The Special Rapporteur is grateful for research and analysis undertaken by Katarina Sydow, her students at New York University School of Law and her colleagues at University of California, Los Angeles. 2 Meredith Whittaker and Lucy Suchman, “The myth of Artificial Intelligence”, The American Prospect, 8 December 2021. 3 A/73/348, para. 3. 4 Tim Mucci, “What is predictive AI?”, IBM, 12 August 2024. 5 Cole Stryker and Mark Scapicchio, “What is generative AI?”, IBM, 22 March 2024. 6 Submissions from Latvia, Morocco, the Bolivarian Republic of Venezuela, and the Attorney General’s Office of Uruguay. 7 See A/78/171. 4/21

The aim of the present report is to provide a road map on how to navigate the risks and benefits of AI. First, the impact of AI on numerous human rights within the Special Rapporteur’s mandate is examined, followed by an assessment of how the promise of AI can be harnessed, and the associated pitfalls avoided, in relation to those rights. The report draws on submissions to the Special Rapporteur⁸ and data from online consultations. The Special Rapporteur is grateful to the Africa Judges and Jurists Forum, the Council of Bars and Law Societies of Europe, the Cyrus R. Vance Center for International Justice, the Hague Institute for Innovation of Law, the International Association of Judges and the United Nations Educational, Scientific and Cultural Organization (UNESCO) for their support in organizing regional and thematic consultations.

8. Submissions are available at www.ohchr.org/en/calls-for-input/2025/call-input-special-rapporteur-independence-judges-and-lawyers-her-next. ↩

The Secretary-General has stressed that the use of technology by Member States must be consistent with international human rights standards,⁹ while the United Nations High Commissioner for Human Rights has recommended that States ban AI applications that cannot be operated in full compliance with human rights law and impose moratoriums on the sale and use of AI systems that carry a high risk for enjoyment of human rights, unless and until adequate safeguards are in place.¹⁰

9. A/79/296, para. 3. ↩
10. See A/HRC/48/31. ↩

AI has the potential to affect a wide range of human rights, many of which have been analysed by other special procedures mandate holders and human rights treaty bodies. The present report is focused on numerous international human rights norms central to the independence of judges and lawyers that are likely to be affected by AI, specifically the rights to access justice, to equality before the law, to a fair trial and to a competent, independent and impartial tribunal. The ways in which AI use by lawyers may affect these rights are considered, but the regulation of AI use by lawyers, which has been the subject of extensive study by independent bar associations, is not examined.

The right to access justice is derived from article 8 (right to an effective remedy) and article 10 (right to a fair and public hearing by an independent and impartial tribunal) of the Universal Declaration of Human Rights, as well as from article 2 (3) (right to an effective remedy) and article 14 (right to a fair and public hearing before a competent, independent and impartial tribunal) of the International Covenant on Civil and Political Rights. Effective access to justice requires citizens to be aware of and able to use the law. To that end, under principle 4 of the Basic Principles on the Role of Lawyers, Governments and professional associations are required to promote programmes to inform the public about their rights and duties under the law.

Article 26 of the International Covenant on Civil and Political Rights guarantees equality before the law and equal protection of the law, while article 14 provides for equality before courts and tribunals. Equality before the law is also protected by the core non-discrimination treaties. Non-discrimination is just one aspect of the right to a fair hearing before a court or tribunal. The right to equality of arms is also included, which requires the same procedural rights and protections to be provided to all parties (Human Rights Committee, general comment No. 32 (2007) on the right to equality before courts and tribunals and to a fair trial). Other elements of a fair hearing 8 Submissions are available at www.ohchr.org/en/calls-for-input/2025/call-input-specialrapporteur-independence-judges-and-lawyers-her-next. 9 A/79/296, para. 3. 10 See A/HRC/48/31. enshrined in article 14 of the Covenant and general comment No. 32 (2007) include the right to a hearing without unreasonable delay and the right to be assisted by an interpreter. Furthermore, under article 14 of the Covenant and principle 1 of the Basic Principles on the Role of Lawyers, an individual’s right to receive assistance from a lawyer of one’s choosing is guaranteed. In article 14 (3) of the Covenant, it is established that States must provide an interpreter for free in criminal cases, as well as free legal assistance to those without sufficient means. In general comment No. 32 (2007), it is clarified that States are also encouraged to provide free legal aid in civil cases for individuals who do not have sufficient means to pay and that States may be required to provide the free assistance of an interpreter where otherwise an indigent party would be unable to participate in the proceedings on equal terms.

In general comment No. 32 (2007), it is explained that the right to a hearing before a competent, independent and impartial tribunal constitutes an absolute right from which there can be no exception. Accordingly, States are required to protect judges from political influence in their decision-making. Independence requires that judges decide matters “without any restrictions, improper influences, inducements, pressures, threats or interferences, direct or indirect, from any quarter or for any reason”.¹¹ Impartiality requires that judges not allow their decisions to be influenced by personal bias or prejudice and also that they appear to be impartial to a reasonable observer.¹²

11. Basic Principles on the Independence of the Judiciary, para. 2. ↩
12. Human Rights Committee, general comment No. 32 (2007), para. 21. ↩

The Special Rapporteur considers that the right to a hearing before an independent and impartial tribunal requires access to a human judge and that the right to counsel of one’s choosing entails access to a human lawyer.

The Secretary-General has examined a number of specific concerns relating to the rights outlined above.¹³ First, the use of biased AI in criminal legal systems entails a risk of discriminatory outcomes. Second, the use of AI may result in an inequality of arms, particularly in criminal proceedings where defendants remain unaware of how AI systems affect their arrest and prosecution. Third, if AI is used to automate judicial decisions, the “black box” nature of AI tools may render the decision-making process so opaque and incontestable that the right to a fair trial is violated. In particular, there is a risk that automation bias – the tendency to consider uncritically solutions offered by technology, such as AI, as correct – may render human input ineffective. In addition, judicial independence may be undercut by the influence exerted by political branches of government and even by private companies in the design, development, training and deployment of AI solutions used in judicial systems. Judges may face reprisals for exercising judicial discretion and deviating from recommendations generated through assisted or automated decision-making, while court administration tools may be used to impose efficiency-based performance indicators and targets. Moreover, where the judiciary is targeted by autocratic leaders, organized crime or powerful economic actors, AI could supercharge intrusive surveillance, manipulation or undue influence.

13. A/79/296, paras. 6–22 and 37–39. ↩

The UNESCO Global Toolkit on AI and the Rule of Law for the Judiciary contains an analysis of human rights affected by AI, including rights that fall within the Special Rapporteur’s mandate.¹⁴ It is noted that, since many AI systems are opaque, individuals may be unaware of how decisions affecting their rights were made or whether the process was discriminatory. Often, the judicial operator using AI may be unable to explain the automated decision-making process, which complicates the justification and appeal of AI-based decisions. UNESCO describes automation bias as “one of the greatest threats generated using AI systems in the administration of justice”. Uncritical acceptance of AI-based decisions may lead to a change in judicial reasoning aimed not at explaining the reason for the decision but rather at justifying why the result offered by the system has not been followed. Such a change would seriously degrade the core principle of reasoned decision-making.

14. United Nations Educational, Scientific and Cultural Organization (UNESCO), Global Toolkit on AI and the Rule of Law for the Judiciary (Paris, 2023), pp. 128–145. ↩

AI guidelines and regulation are aimed at responding to these concerns. In the Global Digital Compact, adopted in 2024 as an annex to the Pact for the Future, Member States stress the need to enhance international governance of AI, while calling upon standards development organizations to promote the development of AI standards that uphold safety, reliability, sustainability and human rights.¹⁵

15. General Assembly resolution 79/1, annex I, para. 5. ↩

In some regional and national contexts, general purpose legislation that is broadly applicable to AI, including the use of AI in justice systems, has been passed. Such legislation may be specifically directed at regulating AI¹⁶ or at establishing protections, such as protections related to data privacy,¹⁷ that have become necessary owing to the use of AI. In many States, however, there is no regulatory regime applicable to AI.¹⁸

16. For example, see European Union, Artificial Intelligence Act, regulation No. 2024/1689 (13 June 2024). ↩
17. For example, see European Union, General Data Protection Regulation, regulation No. 2016/679 (27 April 2016). ↩
18. Herbert Smith Freehills Kramer, “AI tracker: tracking AI law and policy globally”, available at www.hsfkramer.com/en_US/insights/reports/ai-tracker. ↩

The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law is the first international legally binding treaty on AI. It has been signed by 16 countries, including several outside Europe. Article 4 provides that each Party shall adopt or maintain measures to ensure that the activities within the life cycle of AI systems are consistent with obligations to protect human rights. Article 5 stipulates that Parties shall adopt or maintain measures that seek to ensure that AI systems are not used to undermine democratic institutions, including respect for judicial independence and access to justice.

While general purpose regimes may provide an overarching framework for the regulation of AI in justice systems, it is vital that judiciaries govern their own use of AI. Nevertheless, even though a survey conducted by UNESCO in 2023 on AI use in judicial systems revealed that up to 44 per cent of judicial operators had used AI for work-related activities, only 9 per cent of respondents said that their organization had issued guidelines on the use of AI.¹⁹ Judicial councils and judges’ associations have commenced the vital work of preparing rules and guidelines at the domestic level.²⁰ The Special Rapporteur hopes that the guidance furnished in the present report serves to ensure that human rights remain at the core of the development and application of such rules.

19. UNESCO, “UNESCO global judges’ initiative: survey on the use of AI systems by judicial operators”, 2024. ↩
20. For examples see International Association of Judges, “General report of the 1st Study Commission of the International Association of Judges: the effects of artificial intelligence on the judiciary”, 2024. ↩

There is a global crisis in access to justice. According to a survey conducted in 2019, 49 per cent of respondents had experienced at least one legal problem in the preceding two years; less than a third of those who had experienced a legal problem had sought advice to understand or resolve their problem; and those who had sought assistance preferred to turn to family or friends.²¹ The Special Rapporteur examined this crisis and its impacts in a previous report,²² observing that the justice gap has multiple causes, including the cost of hiring a lawyer, a scarcity of lawyers in some areas, practical difficulties in reaching courts, and justice systems that are intimidating or ill-equipped to meet the needs of marginalized people. Where hiring lawyers is not feasible, non-lawyers experience additional barriers to self-representation. Technical legal language can make it challenging to perform even simple tasks, while a lack of knowledge means that many non-lawyers are not even aware that their problems are legal in nature.²³

21. World Justice Project, Global Insights on Access to Justice: Findings from the World Justice Project General Population Poll in 101 Countries (Washington, D.C., 2019), pp. 6 and 7. ↩
22. A/78/171. ↩
23. World Justice Project, Global Insights on Access to Justice, p 7. ↩

States must uphold the right of an individual to receive assistance from a lawyer of one’s choosing. However, a multitude of legal problems currently fall outside the scope of this right, leaving people without support. In addition, in cases where people are unaware that their problems are legal in nature, access to intelligible information about the law could motivate those individuals to seek the assistance of a lawyer or engage with the formal justice system.

AI can be used to extract information and collate and convey it in accessible language and user-friendly formats. It can be used to give life to repositories of information by helping the public to access information and to better understand the law and their rights. For example, AsyLex, a non-profit organization, operates the “Rights in exile” digital platform, which, with the assistance of AI, compiles country-specific resources for asylum cases related to sexual orientation and gender identity and LGBTQI+ rights.²⁴ In India, Haqdarshak, a social enterprise, uses AI to extract information from government databases and long administrative documents in order to promote awareness of welfare entitlements.²⁵

24. Consultation with Access to Justice Innovators. ↩
25. Consultation with Access to Justice Innovators. ↩

AI can be used to transform technical legal information or judicial decisions into plain language to support broader understanding. In Spain, the “Carpeta justicia” system can be used to produce a summary or translation of any legal document into simplified terms. In Mexico, the Supreme Court has launched the “Sor Juana” system, 21 World Justice Project, Global Insights on Access to Justice: Findings from the World Justice Project General Population Poll in 101 Countries (Washington, D.C., 2019), pp. 6 and 7. 22 A/78/171. 23 World Justice Project, Global Insights on Access to Justice, p 7. 24 Consultation with Access to Justice Innovators. 25 Consultation with Access to Justice Innovators. 8/21 which uses AI to translate publicly available court rulings into plain language, generate press releases and answer queries from the public.²⁶ Providing legal information

26. Submission from Mexico; and consultation with judges from Latin America. ↩

Building on the capacity of AI to collate and translate legal information, people who are not well-versed in legal technicalities can use AI as an intermediary to receive queries in plain language, identify applicable laws and translate them into a comprehensible format to facilitate understanding and action by the justice-seeker. There has been a proliferation of such AI-powered assistants around the world. Applications brought to the Special Rapporteur’s attention include a virtual wizard in Latvia that provides basic legal information on common topics and offers possible solutions,²⁷ a chatbot on the official website of the Malaysian court system that provides answers to basic questions regarding sharia procedures,²⁸ and a digital assistant service in Saudi Arabia that can be used by litigants to gain an understanding of legal options and obtain answers to inquiries related to judicial systems.²⁹ In Nigeria, the OpenLawsNig platform has an AI assistant targeted at low-income and underserved communities,³⁰ which is aimed at democratizing access to legal information by making a library of laws, rights and public legal data accessible, understandable and actionable.

27. Submission from Latvia. ↩
28. Submission from Malaysia. ↩
29. Submission from Saudi Arabia. ↩
30. Consultation with Access to Justice Innovators. ↩

In the light of the great inequalities in technology literacy and access to computers and the Internet between States, communities and individuals, AI assistants will be most effective when they leverage interfaces that have been tailored to enhance accessibility, such as chatbots that can be operated on smartphones through messaging services or voice assistance in multiple languages. In Nigeria, the Podus AI tool, which is integrated with WhatsApp, provides free legal first aid in three major local languages, with functionality for voice prompts and audio responses. Users can also escalate their issues by creating a case and connecting directly to a lawyer.³¹

31. Interview with Nelson Olanipekun. ↩

Some AI assistants go beyond facilitating access to relevant legal information and provide guidance about where and how to take action. The Australia-based National Justice Project operates Hear Me Out, a free AI-powered, direct-to-consumer platform that provides recommendations on where and how to file complaints.³² 32 AI can be used to automate document drafting if a user inputs relevant information. AI assistance has certain benefits: it can be accessed 24 hours a day, in multiple languages, without users having to travel or navigate often intimidating interactions. AsyLex reports that its AI assistant helps to reach asylum-seekers who struggle to obtain in-person support from its staff based in Switzerland. After collecting an applicant’s information, the AI assistant can be used to generate a case file and draft complaints to United Nations bodies, saving time and effort.³³ In appropriate circumstances, when there is no right to free legal aid and litigants regularly represent themselves, AI might even facilitate interactions with the court system. AI tools can assist non-lawyers with the preparation and filing of documents in court. For example, 26 Submission from Mexico; and consultation with judges from Latin America. 27 Submission from Latvia. 28 Submission from Malaysia. 29 Submission from Saudi Arabia. 30 Consultation with Access to Justice Innovators. 31 Interview with Nelson Olanipekun. 32 Dean Moutopoulos and others, “AI-powered platforms for access to justice: the case of hear me out”, UNSW Law Research, No. 25-13 (February 2025). 33 Consultation with Access to Justice Innovators. Michigan Legal Help offers a free document assembly program called MiFILE for low-income and self-represented litigants. Analysis of trends in access to justice

32. Dean Moutopoulos and others, “AI-powered platforms for access to justice: the case of hear me out”, UNSW Law Research, No. 25-13 (February 2025). ↩
33. Consultation with Access to Justice Innovators. ↩

AI can be harnessed to analyse existing legal rules and procedures to better meet people’s needs. The Ministry of Justice of Colombia proposes using AI to process and analyse judicial data to support the development of strategies in the justice sector. This could include identifying patterns in demand for court services and trends in court cases, which enables the optimization of decisions on resource allocation and the development of inclusive public policies to improve access to justice for the neediest populations.³⁴ Data, however, do not always present a straightforward picture on access to justice; the underrepresentation of some populations in data may reflect lack of access rather than a lack of need for services. As long as States take a nuanced approach to data interpretation, the use of AI to identify trends holds promise. Furthermore, if AI assistants help more people to understand the law, this may encourage communities to use the law, participate in debates about the law and advocate for changes to make it better serve their needs. Facilitating the widest possible engagement with the law and legal processes may result in collective benefits.

34. Submission from Colombia. ↩

The promise of AI is dependent on its capacity to carry out tasks with a high level of accuracy. If AI applications summarize or translate a document inaccurately, misinterpret a user’s query, identify irrelevant legal rules or pursue improper avenues for redress, they may do more harm than good. Legal terms of art pose a particular challenge, as words that might be synonymous in ordinary use can have very different meanings in the context of the law. Careful thought should be given to the types of issues that might be suitable for being addressed by AI, while recognizing the risk that inaccurate information may be supplied and relied upon without any straightforward method of redress for the user, especially when lawyers are not available to help users with legally related AI tools.

Errors may be minimized, although not fully avoided, when AI is trained on custom data sets and subjected to legal review. When done properly, this process is expensive and time-consuming. Developers of these technologies describe the need to gather and digitize relevant legislation, court decisions and other information. For one virtual assistant, the focus of which is limited to housing repairs law in New York City, that process alone took a team of five people 13 months.³⁵ Lawyers should be deeply involved in supervised learning and ongoing review of any legal AI application, checking to ensure that it is providing accurate answers, restricting itself to matters that fall within its knowledge base and not “hallucinating”. Ongoing quality control also requires continuous monitoring to ensure that the knowledge base remains up to date, which is another significant financial burden. 34 Submission from Colombia. 35 Marco Poggio, “NYU law professor on his AI-powered tenants’ rights bot”, Law360, 7 February 2025. 10/21 Legal system not digitalized

35. Marco Poggio, “NYU law professor on his AI-powered tenants’ rights bot”, Law360, 7 February 2025. ↩

Many States lack national legal information systems, fully digitized resources and legal databases. This paucity of data may require the allocation of significant resources for identifying, compiling and digitizing relevant laws, cases and guidance, before effective AI solutions can be developed. All digitalization and data collection processes must be human rights-compliant³⁶ and should result in the creation of free, open-access legal resources.

36. Office of the United Nations High Commissioner for Human Rights, “A human rights based approach to data”, 2018. ↩

AI-based solutions are expensive. Profit-making enterprises, non-profit actors, academia and States are all involved in developing AI solutions for access to justice. Many of the projects described above were developed using funds from grants or investors, with developers reporting that they had not identified a sustainable financing model.³⁷ Another reported problem pertains to interoperability between private sector and public technologies. The utility of self-help tools will be reduced if their outputs (such as automatically generated forms or complaint letters) are incompatible with the platforms used by the State (such as court e-filing systems). Furthermore, beneficial AI solutions may be abandoned because funds run out or new regulatory barriers are imposed. Regulation is an important method of securing safety and quality control for consumers and preventing the improper use of data. However, a patchwork approach to both AI adoption and regulation is unlikely to lead to optimal results.

37. Consultation with Access to Justice Innovators. ↩

People-centred justice is required to narrow the justice gap. Data on the legal needs that matter most to people should be gathered, and solutions should be identified from the perspective of those experiencing justice problems. Where appropriate, justice solutions may involve AI, but the right to access courts or receive assistance from a lawyer of one’s choosing should never be diminished or diluted. The way that applications are selected and implemented matters greatly. Even if AI outputs are not inaccurate or misleading, AI may nevertheless fail to provide adequate benefit. Some States reported deactivating chatbots due to poor user uptake.³⁸ Given the costs of developing and maintaining AI, design should be driven by users, reflecting their diverse needs.³⁹ States should always consider whether AI or a low- or even no-tech tool is best for ensuring access to justice.

38. Submission from Austria. ↩
39. Sherley Cruz, “Coding for cultural competency: expanding access to justice with technology”, Tennessee Law Review, vol. 86, No. 347 (2019). ↩

Offering new paths for filing complaints will be futile if dispute resolution services lack the capacity to address those complaints. Overreliance on technological solutions must not lead to the defunding of legal services and existing efforts to close the justice gap, especially since many populations are excluded from digital solutions. Nevertheless, in some circumstances, AI can be used to free up human advisors to do more, in addition to putting the law in the hands of the people more effectively, allowing communities to advance their rights.

Around the world, there are failures to achieve guarantees of due process, including the right to equality before the law, the right to access a lawyer of one’s choice, the right to interpretation and the right to a speedy trial. The Special Rapporteur has received many allegations concerning these issues and has repeatedly observed such problems when undertaking country visits. When AI is used for the administration of justice, it should enhance, and never undermine, the right to a fair trial. This suggests that some AI uses will be more appropriate than others.

AI is particularly useful for the translation of legal documents and interpretation during proceedings, including in States with multiple official languages.⁴⁰ This may help to improve equality before the courts for different language users at a lower cost than human translation services. However, AI remains significantly less reliable in its translation of some languages, in particular so-called low-resource (including Indigenous) languages, which lack large online data sets or are spoken by relatively small populations. Given their limited digital resources, States in the global South are likely to be particularly affected by this issue.

40. Submissions from India, Saudi Arabia and Spain. ↩

A variety of AI functionalities may help to speed up the administration of justice, due to the potential of AI to perform some tasks more swiftly than humans. It can be deployed to analyse and classify or cluster cases when they first enter the court system, suggesting an order of importance or urgency, or recommending an assignment to the appropriate court. Examples include the optimized AI-managed inbox described in the submission of Austria or the Minerva procedural management system described in the submission of Spain. The VICTOR system used in Brazil is one of the most prominent models, speeding up the process of admissibility assessments before the Supreme Court. Where a human would take an average of 44 minutes, the VICTOR system spends only a couple of seconds.⁴¹

41. Daniel Becker and Isabela Ferrari, “VICTOR, the Brazilian Supreme Court’s artificial intelligence: a beauty or a beast?”, 15 March 2020. ↩

In many States, AI-powered speech recognition tools are being used or piloted to create automatic transcriptions of live or recorded court hearings.⁴² These can be reviewed for accuracy by court staff more rapidly than if staff were required to carry out manual transcription. AI can also be used to anonymize case documents swiftly.⁴³

42. Submissions from Colombia, India, Latvia, Lithuania, Malaysia, Spain and the Mongolian Judges’ Association. ↩
43. Submissions from Austria, Colombia, Romania and Spain. ↩

Avoiding unreasonable delay is of particular importance to individuals held in pretrial detention, who constitute approximately one third of the global prison population.⁴⁴ In Nigeria, the non-governmental organization Citizens’ Gavel is attempting to leverage AI to address this problem. Using optical character recognition, prosecutors can scan a paper case file and use AI trained on Nigerian law to generate a rapid legal opinion to aid in assessing the possibility of pretrial release. The aim is to decrease pretrial detention and reduce the time between arrest and release from months to days.⁴⁵

44. United Nations Office on Drugs and Crime, Prison Matters 2024: Global Prison Population and Trends – A Focus on Rehabilitation (Vienna, 2024), p. 13. ↩
45. Consultation with Access to Justice Innovators. ↩

Some States have begun using or contemplating automated decision-making under the supervision of a human judge to speed up the processing of certain highvolume cases, such as air passenger complaints.⁴⁶ Enhancing the capacity of legal aid lawyers

46. Consultation with judges from Europe and North America. ↩

AI may also be used to support the right to assistance from a lawyer. In a recent field study, 90 per cent of legal aid professionals provided with free access to AI products for legal research and drafting assistance reported increased productivity.⁴⁷ Notably, the reported use of AI programs to, for example, carry out confirmatory research, draft cease and desist letters, and translate documents into simple language in no way replaced the human lawyer. However, reducing the time spent performing lower stakes tasks freed up resources for more sophisticated work. Arguably, this increased productivity could allow legal aid lawyers to support more low-income clients, leaving fewer people without representation.

47. Colleen V. Chien and Miriam Kim, “Generative AI and legal aid: results from a field study and 100 use cases to bridge the access to justice gap”, Loyola of Los Angeles Law Review, vol. 57, No. 4 (2025). ↩

Procedural and administrative acts can affect human rights as much as substantive decision-making, and there may be a lack of meaningful avenues for appeal. AI may introduce delays if it mis-categorizes a case during triage. AI-powered case allocation also risks facilitating capture of systems. For example, an algorithm could be biased towards assigning cases against the Government to pro-government judges, or cases against businesses to pro-business judges. Any allocation system is vulnerable to exploitation, but AI raises particular concerns due to the lack of transparency around its operation. A preliminary ruling has been requested from the Court of Justice of the European Union regarding the Random Allocation of Judges System (SLPS) in operation in Poland, due to failures to publish the tool’s source code and difficulties ascertaining its vulnerability to errors and manipulation. Particular concerns were raised in the context of the tool’s allocation, in a single draw, of 56 cases to one judge and no cases or just a few cases to other judges.⁴⁸

48. Poland, District Court of Warsaw, Case C-159/25 (Rowicz), February 2025. ↩

The black box nature of AI is especially problematic when AI-generated evidence is used in court proceedings, particularly criminal trials. The principle of equality of arms demands that each side has the opportunity to contest evidence adduced by the other party.⁴⁹ However, in many States, black-box AI systems produce evidence that is relied on by prosecutors in court while being protected from meaningful challenge on the basis that it is proprietary technology. In the United States of America, evidence produced by the AI-powered gunfire detection system ShotSpotter, which is the subject of concerns regarding its performance accuracy, has been admitted in 200 criminal cases, sometimes without admissibility hearings,⁵⁰ while its developers shielded its source code from scrutiny by claiming trade secret protections.⁵¹ Furthermore, there is no way to challenge the propriety of technologies if they do not appear in the court’s record.⁵² For example, police might identify suspects using AI facial recognition, but a subsequent positive identification through a standard photo array might be relied on at trial. This practice is concerning in view of overwhelming evidence that AI systems replicate and even enhance human biases.⁵³

49. Human Rights Committee, general comment No. 32 (2007), para. 13. ↩
50. Brendan Max, “SoundThinking’s black-box gunshot detection method: untested and unvetted tech flourishes in the criminal justice system”, Stanford Technology Law Review, vol. 26, No. 2 (2023). ↩
51. Rebecca Wexler, “Life, liberty, and trade secrets: intellectual property in the criminal justice system”, Stanford Law Review, vol. 70 (May 2018), pp. 1365–1366. ↩
52. Alissa Marque Heydari, “AI and prosecution: mapping the current and future roles of artificial intelligence in prosecution”, 11 December 2024, pp. 10 and 11. ↩
53. A/79/296, para. 16. ↩

AI-generated evidence may also affect the fairness of trials through deepfakes and digital enhancement technology. Such destabilization of the reliability of audiovisual evidence risks increasing scepticism⁵⁴ and harming public trust.

54. Rebecca A. Delfino, “Deepfakes on trial: a call to expand the trial judge’s gatekeeping role to protect legal proceedings from technological fakery”, Hastings Law Journal, vol. 74, No. 2 (2023), pp. 310–313. ↩

The Special Rapporteur has noted the potential benefits that AI could bring to legal aid lawyers. However, in practice, the prohibitive cost of many specialized legal research tools limits these programs to big law firms that can afford enterprise-level investments, relegating less wealthy lawyers and their clients to free or low-cost models that may be less accurate or effective. This uneven adoption risks creating a stratified legal ecosystem.⁵⁵ These disparities can be significant to case outcomes, particularly when specialist AI is used to create granular predictions regarding the rulings of individual judges, permitting highly calculated choices of forum and litigation strategy.⁵⁶ Arguably, this allows certain litigants to rig the system in their favour. In 2019, France was the first country to outlaw the use of such analytics to predict the practices of individual judges.⁵⁷ To mitigate the lopsided benefits of specialist AI, the Paris Bar developed an innovative scheme to offer subsidized or free access to legal AI tools to solo practitioners and two-person firms.⁵⁸ However, unequal benefits arise not only in access to AI tools, but the uses that can be made of them. Particular advantages accrue to wealthy repeat players, primarily business actors, with access to massive volumes of private settlement data that can be used to train specialist AI.⁵⁹ Repeat players may even stand to gain in relation to AI systems used for court administration.⁶⁰ Large language models are vulnerable to manipulation,⁶¹ suggesting that sophisticated users could game triage or case management processes.

55. Drew Simshaw, “Access to A.I. justice: avoiding an inequitable two-tiered system of legal services”, Yale Journal of Law and Technology, vol. 24 (2022), p. 156. ↩
56. Global Newswire, “Pre/dicta expands its AI-powered litigation prediction platform: now includes motion for summary judgment, class certification, case timelines, and more”, 14 November 2023. ↩
57. Michael Livermore and Dan Rockmore, “France kicks data scientists out of its courts”, Slate, 21 June 2019. ↩
58. Submission from the Paris Bar. ↩
59. David Freeman Engstrom and Nora Freeman Engstrom, “Legal tech and the litigation playing field”, in Legal Tech and the Future of Civil Justice, David Freeman Engstrom, ed. (Cambridge, 2023). ↩
60. Submission from the Supreme Court of Brazil. ↩
61. Abhinav Rao and others, “Tricking LLMs into disobedience: formalizing, analyzing, and detecting jailbreaks”, 2024. ↩

Whatever strides have been made in developing specialist legal AI, technology cannot offer the full services of a lawyer. However, in the worst-case scenario, States may attempt to replace free legal aid with lower cost AI programmes, undermining the right to counsel for poorer litigants. This is unacceptable: although AI may offer productivity benefits, this should never be used as an excuse to deprioritize funding for legal aid lawyers. Similarly, States should exercise caution regarding automation in legal proceedings. Even so-called small claims may be brought by litigants with a high level of financial dependence on the outcome; they should not be required to forfeit the right to human adjudication.⁶²

62. Consultation with UNESCO experts. ↩

Approximately one third of individuals worldwide still do not have access to the Internet.⁶³ Furthermore, only 32 countries host the powerful data centres used for developing complex AI systems. The majority of these are located in the United States and China, while Africa and South America have almost no AI computing hubs.⁶⁴ Even when States in the global South gain access to AI tools, they can be ill-suited for deployment in local judicial systems, given the vastly different contexts in which they were developed. For example, commonly used large language models are more proficient and accurate in English and Chinese.⁶⁵ As already described, States with limited digital data sets will not see their national laws and precedents represented, reducing the utility of AI tools. This risks the development of two speeds between the global North and South, one digital and the other analogue.⁶⁶

63. International Telecommunication Union (ITU), ITU DataHub, available at https://datahub.itu.int/. ↩
64. Adam Satariano and Paul Mozur, “The global AI divide”, The New York Times, 21 June 2025. ↩
65. Ibid. ↩
66. Submission from Mozambique. ↩

States tempted to resort to public-private partnerships to facilitate rapid digitalization and the introduction of AI justice tools should be aware of the risk of abuse by companies motivated by profit, which are likely to want to charge for their services or monetize data extracted from judicial systems.⁶⁷ In particular, States must ensure that they retain ownership of data and that they are not bound to pay an ongoing licencing fee or risk losing access to their own data and bespoke systems.

67. United Nations Development Programme, e-Justice: Digital Transformation to Close the Justice Gap (New York, 2022), p. 24. ↩

In section III of the present report, the Special Rapporteur considered how AI could be used to close the justice gap. However, there is another possibility: that AI will widen that gap. Researchers have observed the frequent overlap between individuals affected by the digital divide and those with unmet justice needs.⁶⁸ Given the specialized skills, data, and financial and technical resources required to build and adapt AI, this gap may be supercharged if AI technologies are favoured over low-tech improvements. 61 Abhinav Rao and others, “Tricking LLMs into disobedience: formalizing, analyzing, and detecting jailbreaks”, 2024. 62 Consultation with UNESCO experts. 63 International Telecommunication Union (ITU), ITU DataHub, available at https://datahub.itu.int/. 64 Adam Satariano and Paul Mozur, “The global AI divide”, The New York Times, 21 June 2025. 65 Ibid. 66 Submission from Mozambique. 67 United Nations Development Programme, e-Justice: Digital Transformation to Close the Justice Gap (New York, 2022), p. 24. 68 World Justice Project and World Bank, Advancing Access to Justice via Information and Communication Technology: A Literature Review (2025), p. 36.

68. World Justice Project and World Bank, Advancing Access to Justice via Information and Communication Technology: A Literature Review (2025), p. 36. ↩

Nevertheless, some uses of AI offer positive trade-offs that may justify the potential risks. States must weigh the rewards and risks of a particular use case with the practical problems that the justice system faces. Brazil is a good example: with over 80 million cases awaiting judgment,⁶⁹ many believe that the risks of adopting an AI tool to speed up adjudication offers a fair trade. It is vital that such decisions are taken following multi- stakeholder consultation, with full and transparent information regarding the AI tools to be deployed and their anticipated risks, as well as the consideration of low- or no-tech solutions that may avoid such harms entirely.

69. Daniel Becker and Isabela Ferrari, “VICTOR, the Brazilian Supreme Court’s artificial intelligence: a beauty or a beast?”, 15 March 2020. ↩

Around the world, independent judges are under attack from powerful political⁷⁰ and economic⁷¹ actors who seek to capture justice systems and wield them to advance their own ends. Moreover, judges in many States are overwhelmed and under-resourced, facing ever-increasing backlogs. In this environment, AI promises productivity gains to overworked judiciaries, while also creating significant novel risks of capture and even techno-authoritarianism.

70. See A/HRC/56/62. ↩
71. See A/79/362. ↩

Section IV contains instances of AI use by judicial systems; the present section considers more direct use of AI by judges in judicial tasks. Two separate categories must be taken into account: use of AI pursuant to an official judiciary-wide (or more localized court-specific) policy; and unregulated use of AI by individual judges. The primary uses of AI by judges are likely to be the same in both categories, and could encompass any of the following: summarizing legal submissions and evidence; finding similarities between cases and offering insights; searching for and retrieving relevant legal documents and precedents; conducting legal research; predicting legal outcomes; conducting risk assessment; providing templates or initial drafts of opinions or drafting routine orders; and editing, proofreading and checking spelling and grammar in draft opinions.

Certain judges advised the Special Rapporteur that AI promised to reduce their administrative workload, for example by summarizing parties’ positions for inclusion in written opinions.⁷² Similarly, some judges hearing cases involving vast quantities of evidence have used AI to summarize and search within that evidence, saving time.⁷³ 73 In States where the official language of the court is not the first language spoken by many judges, AI support with spelling, grammar and syntax can also reduce the time taken to finalize written judgments. An important randomized controlled trial is currently ongoing in Pakistan, comparing the performance of judges provided with a bespoke AI-based tool called JudgeGPT with the performance of judges who do not 69 Daniel Becker and Isabela Ferrari, “VICTOR, the Brazilian Supreme Court’s artificial intelligence: a beauty or a beast?”, 15 March 2020. 70 See A/HRC/56/62. 71 See A/79/362. 72 Consultation with judges from Asia and Oceania. 73 Consultation with judges from Europe and North America. 16/21 have access to the tool.74 Studies of this kind, when administered under judicial supervision, could objectively demonstrate the impact of AI on judges’ productivity and the quality of their rulings. Crucially, such studies could also reveal unexpected risks or problems. Counteracting bias

72. Consultation with judges from Asia and Oceania. ↩
73. Consultation with judges from Europe and North America. ↩

Innovative AI programs aim to mitigate unintentional biases and enhance impartiality. In Pakistan, Project Mehfooz is developing AI agents to provide multidisciplinary support, including in the fields of child psychology and neuroscience, designed to help judges resist patriarchal norms and adopt trauma-informed decision-making in cases of domestic violence. AI may identify and mitigate other patterns of discrimination, for example by flagging racially inflected language in court transcripts or auditing sentencing patterns for racial disparities. As ever, the utility of such tools depends on their accuracy and on ensuring that they are themselves transparent and free from improper influence and biases. The Special Rapporteur stresses that such uses must be at the behest of and under the governance of the judicial branch.

The Special Rapporteur believes that the right to an independent and impartial tribunal requires access to a human judge. Judges consulted for the present report agreed that AI must never replace a human in making final decisions. Even in States without an express constitutional prohibition against judges delegating their jurisdiction, significant concern was expressed about the risks, and what would be lost, if AI was used for decisions currently made by judges. For some, this was because of the sensitivity, nuance and human understanding required, for example in family courts,⁷⁵ or because the dignity of a criminal defendant could only be guaranteed if a human determined their guilt. Others pointed out that AI cannot reason and thus cannot provide reasoned decisions.

75. Consultation with Access to Justice Innovators. ↩

Judges are familiar with, and have acted to counteract, the problem of lawyers submitting pleadings prepared by AI that include so-called hallucinated cases. Comparable errors affect judicial use of AI. For example, automated summaries of judgments have been found to be fraught with inconsistencies and hallucinations.⁷⁶ Training AI tools on a closed corpus of domestic law may improve their performance,⁷⁷ but requires up-to-date local legal databases. Even specialized legal research tools continue to hallucinate at alarming rates, although they reduce errors when compared with general-purpose models. A recent study found that AI research tools made by LexisNexis (Lexis+AI) and Thomson Reuters (Westlaw AI-Assisted Research and Ask Practical Law AI) hallucinate between 17 per cent and 33 per cent of the time.⁷⁸

76. Aniket Deroy, Kripabandhu Ghosh and Saptarshi Ghosh, “How ready are pre-trained abstractive models and LLMs for legal case judgement summarization?”, 2023. ↩
77. Ibid. ↩
78. Varun Magesh and others, “Hallucination-free? Assessing the reliability of leading AI legal research tools”, Journal of Empirical Legal Studies, vol. 22, No. 2 (June 2025). ↩

In practice, it is extremely difficult to draw a line between AI tools that only assist judges and those that can interfere in decisions. For example, a summary of evidence that excludes or minimizes certain features may encourage a judge to disregard them. An AI tool that rewrites a judgment for clarity may alter its legal meaning. The answer to these concerns is often to retain a human in the loop. However, in many circumstances, this will be insufficient to ensure that a decision has in fact been taken independently by the judge. Therefore, judiciaries should proceed with caution when considering the use of an AI tool, having particular regard to the risks outlined here.

Verification is often identified as a corrective to AI risks, requiring a judge to check and confirm AI outcomes to spot errors and to ensure that a decision remains the judge’s own. However, some systems are not amenable to such review, as a judge cannot easily go into the black box of an algorithm and determine whether it was informed by flawed or discriminatory data. In addition, one of the key promises of AI – saving time – would be negated if judges were required to manually repeat the work of an AI tool to verify its outputs. Judiciaries must consider whether and when verification is feasible, and rigorously assess the risks of outsourcing tasks to AI in the light of that assessment.

Heightened risks to independence arise when AI suggests a particular approach or outcome to a judge. The human tendency to defer to algorithmic recommendations and rely on the ostensibly scientific quality of AI can, almost imperceptibly, erode judicial discretion and lead to automation bias.⁷⁹ Tools that assess risk or suggest or predict a legal outcome based on prior jurisprudence entail such dangers. However, any AI outputs may be uncritically trusted if judges do not receive adequate training and guidance regarding the problems and limitations of these technologies, undermining the capacity of the human-in-the-loop approach to maintain oversight over AI decisions.

79. Richard M. Re and Alicia Solow-Niederman, “Developing artificially intelligent justice”, Stanford Technology Law Review, vol. 22, No. 2 (Spring 2019), p. 268. ↩

The erosion of judicial discretion gives rise to a broader problem: that the mechanistic approaches promoted by AI will gain prominence over critical thinking and ethical and culturally embedded forms of judicial deliberation.⁸⁰ This poses a risk to skills – that judges relying on AI may lose their capacity for legal research, opinion drafting and even judicial reasoning – but also to the judiciary as a specialist epistemic institution tasked with discovering and delivering a certain kind of particularly valuable knowledge.⁸¹ AI systems are designed to draw conclusions based on patterns of aggregate data points; a judge ruling in the same way will be constrained by what the system has encountered before. Judges from common law jurisdictions question whether AI-inflected jurisprudence would stagnate.⁸² Judges from civil law jurisdictions query whether AI would inappropriately follow precedents, including from other jurisdictions.⁸³ Particular legal procedures and theories of justice – whether rehabilitative, retributive, redistributive or otherwise – should not be overturned or imposed through stealth or accident, via the operation of an ostensibly neutral AI tool. Such epistemic capture is particularly concerning when AI tools are dominated by State actors that seek to exercise control over judges. However, it is also a risk when tools designed by private companies, which have inbuilt preferences for the most dominant modes of problem-solving, are relied upon by judges who must, at times, depart from the path most commonly taken to achieve justice.

80. Submission from Ammar Younas. ↩
81. Oliver Milne, “Epistemic institutions: the case for constitutionally-protected academic independence”, Social Epistemology Review and Reply Collective (2019), p. 28. ↩
82. Consultation with judges from Africa. ↩
83. Consultation with judges from Europe and North America. ↩

Influence over AI tools is likely to be a focus for executive or legislative actors seeking to capture judicial systems for their own ends. Judges from many States expressed concerns about the potential for AI to be deployed to limit their independence and capacity to act as a check on government power. In China, courts have integrated AI into judicial processes, with the stated aim of promoting standardization. AI tools monitor and evaluate judges’ decisions, incentivizing conformity with the model’s outputs. Such systems may increase political oversight, rein in judicial autonomy and ultimately undermine independence.⁸⁴

84. Rachel. E. Stern and others, “Automating fairness? Artificial intelligence in the Chinese courts”, Columbia Journal of Transnational Law, vol. 59 (2021). ↩

The involvement of private sector actors in the design and introduction of AI tools also poses risks to judicial independence. The monopolistic nature of the technology industry reduces the number of businesses from which to procure AI tools, risking vendor lock-in and the privatization of public services.⁸⁵ Furthermore, due to claims of proprietary technology or trade secrets, judiciaries may lack adequate information about the development and training of AI programs to assess risks of error and bias. Principle 13 of the Guiding Principles on Business and Human Rights makes clear that private actors have their own duties to avoid causing or contributing to adverse human rights impacts. However, judges themselves must also be satisfied that AI tools procured from private actors do not infringe human rights.

85. A/HRC/59/53, para. 28. ↩

There is well-documented evidence of judicial AI tools that replicate and even exacerbate human biases. The most notorious example is the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) system deployed in some jurisdictions of the United States, which, when used to assist in sentencing decisions, was nearly twice as likely to misclassify Black defendants by predicting them to have a higher risk of recidivism than white defendants.⁸⁶ Judges must ensure that discrimination is not built into any AI tools they employ. Judiciaries should think carefully about the procurement of AI tools – whether purchased from commercial vendors or developed in-house – and their understanding and control over tools’ inbuilt assumptions, as well as the data sets they are trained on.

86. Jeff Larson and others, “How we analyzed the COMPAS recidivism algorithm”, ProPublica, 23 May 2016. ↩

AI is here. It is being used by judges. But no use of AI is without potential adverse human rights impacts. In a landmark 2025 decision addressing judicial AI use, the Supreme Court of Pakistan observed that “justice must be seen as legitimate by those it serves, and this legitimacy stems from the belief that a fellow human has truly listened, understood and acted with conscience”.⁸⁷ Judges consulted for the present report repeatedly voiced their concerns that AI could undermine public trust in justice systems by introducing errors, hallucinations and biases, by exposing or monetizing private data, or by subverting the right to a trial by a human judge.

87. Supreme Court of Pakistan, CPLA No. 1010-L/2022, Judgment, 13 March 2025, para. 13. ↩

The Special Rapporteur urges judiciaries to confront this issue as a matter of priority. To preserve judicial independence, judges must be the ones assessing and deciding on the adoption of any innovation that might affect their decision-making. Judges must therefore have access to training on AI, its promises and pitfalls, and the ability to confer with technologists, lawyers and the public about which AI systems, if any, to embrace.

States and justice professionals should not allow “techno-solutionism” or exaggerated hype about what AI can accomplish to propel the adoption of systems carrying significant human rights risks. States should identify, from the perspective of those experiencing justice problems, data-driven goals for the advancement of human rights. In some circumstances, AI may offer a path to achieving such goals. However, AI should not be adopted without careful assessment of its potential harms, whether these can be eliminated, and whether there are other solutions that are less risky and have fewer negative climate impacts. Whatever choice is made, training and guidance must be made available to mitigate the risks posed by the ongoing ad hoc use of AI by justice operators.

When pursued, AI use in judicial systems must be governed by judges, context-specific, and respectful of local needs, languages, culture and legal traditions. This necessitates prioritizing an enabling environment through the collation and digitization of data, the development of infrastructure and professional capacity-building.

The Special Rapporteur recommends that:
(a) Decisions about whether to use AI in judicial systems, and which tools to use, should be made by judges;
(b) Legal education should include detailed instruction regarding AI, including its practical uses and risks, and AI and digital literacy should be incorporated into continuing legal education requirements;
(c) Judiciaries should develop and adopt guidelines on the use of AI, having regard to international guidance such as that developed by UNESCO,⁸⁸ and States should make resources available to the judiciary for that purpose.

88. UNESCO, Guidelines for the Use of AI Systems in Courts and Tribunals (Paris, 2025). ↩

With regard to reimagining access to justice, States should:
(a) Support the sustainable development of people-centred justice technologies designed to close the justice gap, including by:
(i) Instituting sandbox environments to pilot AI programs and experiment with appropriate regulations.⁸⁹ Programs should focus on issues with high levels of unresolved problems, leverage interfaces tailored to enhance accessibility, and be interoperable with technology platforms used in the formal justice system;
(ii) Coordinating and sharing information about the results of such pilots with other States.
(b) Assess the added value and cost benefit provided by AI when compared with lower tech solutions;
(c) Maintain or enhance alternative, non-digital means of accessing justice.⁹⁰

89. Described in the submission from Azerbaijan. ↩
90. Open Government Partnership, “Justice and artificial intelligence”, 16 May 2025. ↩

With regard to increasing equality before the courts and bolstering fair trial rights, States should:
(a) Uphold the right to a human judge and a human lawyer;
(b) Prioritize the creation of an enabling environment for effective AI by compiling and digitizing relevant laws and creating open, free and public legal databases;
(c) Encourage open data policies in accordance with national data protection laws and support cross-border data-sharing through regional initiatives, to reduce regional inequality in data availability;
(d) Invest in national legal language models;
(e) Support the creation of a global fund for AI to put a floor under the AI divide⁹¹ and ensure that some of that funding is earmarked for legal uses.

91. Governing AI for Humanity (United Nations publication, 2024), p. 65. ↩

The Special Rapporteur recommends that in criminal justice systems:
(a) Prosecutors should avoid relying on AI evidence as the foundation for their pursuit of convictions, unless confident that the evidence is rigorously tested, not discriminatory and can be meaningfully challenged by the defence after mandatory prosecutorial disclosure;
(b) Predictive analytics should never be used as a basis for prosecution or detention;
(c) Judges and prosecutors should act as gatekeepers and only submit and admit reliable AI-generated evidence in criminal trials;
(d) Rules of evidence should provide no carve-out from admissibility requirements for privately owned AI tools and that the standard for admitting AI-produced evidence should be the same as for traditional forensic methods.

The Special Rapporteur also recommends that States and judiciaries should proceed cautiously when:
(a) Adopting AI case management tools, ensuring decisions are taken under judicial supervision and with the benefit of multi-stakeholder input and full and transparent information regarding the AI tool to be deployed and its anticipated risks;
(b) Entering into public-private agreements for the provision of AI services with for-profit private actors, which may charge for services or seek to monetize data extracted from judicial systems.

The Special Rapporteur recommends that:
(a) Businesses developing commercial AI tools for use by justice systems comply with their obligations under the Guiding Principles on Business and Human Rights to avoid causing or contributing to adverse human rights impacts, and address such impacts when they occur, including by carrying out human rights impact assessments and remediating any harms;
(b) Bar associations, States and technology companies consider offering subsidized or free specialist legal AI tools to smaller law firms and public interest lawyers.

With regard to enhancing judicial independence and impartiality, the Special Rapporteur recommends that:
(a) Judges be in charge of the adoption of technological innovations that might affect their work, so as to protect individual and judiciary-wide independence;
(b) When considering AI tools, judiciaries should engage in multi-stakeholder consultations and pilot programmes, move cautiously by testing, assessing and refining, collaborate with technologists, and seek transparency regarding tools’ inbuilt assumptions and the data sets on which they were trained;
(c) Key information about judicial AI systems be made publicly available, to permit legal challenges and oversight by civil society.