1. Principles
Organizations and individual members of the judiciary should follow thirteen principles when adopting and using any type of AI system, including generative AI tools:
1.1. Protection of human rights: Adopt AI systems grounded in respect for human rights that allow the judiciary to respect, protect, and promote human rights while administering justice. 16 While all human rights that may be potentially affected by AI systems should be considered along the life cycle of these tools, the following four points are particularly important in the context of judicial proceedings:
a. Fairness: Adopt AI systems that aim to attain their goals through processes that safeguard fairness and ensure inclusive technology access.17
b. Non-discrimination: Prevent biased applications of AI systems and outcomes that reproduce, reinforce, perpetuate, or aggravate discrimination. 18
c. Procedural fairness: Assess the implications of AI systems for procedural fairness throughout the AI system’s life cycle and prevent deployments that breach rights to procedural fairness.
d. Personal data protection: Adopt AI systems that protect personal data treated for the administration of justice and deploy tools that contribute to anonymizing udicial decisions. 19 The judiciary should avoid using AI tools in ways that generate risks of disclosing such data or enable unauthorized access by third parties.20
1.2. Proportionality: Adopt AI systems that aim to achieve legitimate and proportional ends in the context of their use. 21
1.3. Safety: Adopt AI systems that avoid, address, prevent, and eliminate unwanted harm.22
1.4. Information security: Adopt AI systems that protect confidential information in line with international standards for access to information. 23
1.5. Awareness and informed use: Understand the functionalities, types of uses, potential impacts, limitations, and risks of available AI systems to make informed decisions about their implementation and be aware of the intended purpose for using a specific AI system to carry out judicial activities. 24
1.6. Transparent use: Inform in a proper and timely manner when and how AI systems are used and how these tools work, especially when decisions made with or based on such tools can affect the rights and freedoms of individuals or communities. 25
1.7. Accountability and auditability: Ensure accountability by informing and explaining why certain AI tools were adopted by the judiciary and ensuring traceability of the AI system’s processes and outcomes, mainly when such tools are used for decisionmaking processes. 26 Undertake administrative, legal, and human measures to ensure that the AI systems may be audited during and after their deployment.27
1.8. Explainability: Adopt AI systems that are transparent in terms of how the system operates, how it was developed, its training data, its limitations (including its margin of error), its capabilities, and the purpose of the systems.
28 Explainability refers to making intelligible and providing insight – to deployers and users – into the inputs, outcomes, and functioning of AI systems, making these elements understandable and traceable for humans.29
1.9. Accuracy and reliability: Adopt and use accurate AI systems, meaning AI systems that can provide useful and pertinent information and produce correct outputs and predictions, 30 and reliable AI systems, meaning AI systems that work “properly with a range of inputs and in a range of situations”.31
1.10. Human oversight: Do not rely exclusively on AI systems to adopt decisions or automate entire processes that may negatively impact the rights and freedoms of individuals or communities, and maintain an appropriate level of human control and involvement concerning all AI systems. 32
1.11. Human-centric design: The development, deployment, and use of AI systems should follow human-centric design principles to complement and augment the judiciary's capacities and respect human dignity and autonomy.33
1.12. Responsibility: Organizations that deploy AI systems and individuals that use AI systems must assume responsibility for the decisions and actions taken with the support of AI tools without prejudice to the provider's potential liability in case the AI system is defective. 34
1.13. Multi-stakeholder governance and collaboration: The organizations that are part of the judiciary should consult diverse stakeholders throughout the AI system's life cycle, especially those that may be directly or indirectly affected by its deployment.
Allow meaningful participation of marginalized groups and incorporate their feedback in the development and use of AI tools used to make judgments or with the potential to impact any other significant legal issue. 35
2. Specific guidance for organizations that are part of the judiciary These recommendations apply to the bodies that govern the judiciary, courts, and tribunals that intend to adopt and use AI systems.
2.1. On the adoption of AI tools:
2.1.1. Conduct algorithmic impact assessments before AI systems are deployed, including human rights and ethical assessments. AI systems that support decision-making processes should be examined through algorithmic impact assessments that identify implications for access to justice and individual rights and potential risks. Such reports may inform whether the systems should be deployed and recommend the appropriate risk prevention, mitigation, redressal, and monitoring measures.36 The algorithmic assessments should be available for public consultation. Algorithmic impact assessment guidelines and instruments, such as UNESCO’s “Ethical Impact Assessment: A Tool of the Recommendation on the Ethics of Artificial Intelligence”, may be used to examine potential impacts on human rights such as fair trial and due process, access to justice and effective remedy, privacy, and data protection, equality before the law, and non-discrimination. 37 These assessments should be done in consultation with the relevant stakeholder groups.
Finally, other emerging frameworks could be considered for conducting AI algorithmic impact assessments (including some focusing on human rights). 38
2.1.2. Data transparency, quality, and integrity. Adopt AI systems that offer greater transparency of their training data and allow deployers and users to assess the quality and integrity of such data.39
2.1.3. Necessity, proportionality, suitability, and alignment. The necessity, proportionality, and appropriateness of using an AI system to perform tasks should be assessed and established from the outset and aligned with the objectives of the organization using the AI system (e.g., that protects human rights and ensures the judiciary’s independence and autonomy).
2.1.4. Consult with relevant stakeholders. Before adopting AI systems that are used for decision-making or tools that support management processes and that can significantly affect fundamental rights, such as access to justice, the organization should consult with relevant stakeholders about the opportunities, potential risks, and impacts of introducing such tools.
2.1.5. Compliance with the law and compatibility with organizational standards.
Adopt systems developed and offered in compliance with general principles of international human rights law and local laws, such as data protection and privacy, antidiscrimination, procedural fairness, consumer, competition, and intellectual property laws. Furthermore, when an external provider supplies the AI system, ensure the tool’s terms of use are compatible with local laws and organizational standards of the organizations that govern the judiciary.
2.1.6. Access to key information about AI systems. Obtain information from the AI system’s developers and providers about the tool's adequate uses, the requirements for maintaining its optimal operation, and the limits and risks associated with its adoption. This point may require developers and providers to disclose sufficient information about how the model was developed (including the data used to train it) and how it works. Different transparency instruments, such as model cards, may be considered for providing access to such information. 40 Furthermore, the organizations that are part of the judiciary should adopt explainable AI tools, which refer to systems with the “ability to explain their rationale for decisions, characterize the strengths and weaknesses of their decision-making process, and convey an understanding of how they will behave in the future.”41
2.1.7. Ensure the viability of the algorithmic audits conducted while AI systems are deployed. In the acquisition process and during the implementation process, ensure the developer or provider of the AI system agrees to allow and collaborate with algorithmic audits carried out or commissioned by the organization to external parties (e.g., firms specialized in conducting algorithmic audits that are contracted by the organization of the judiciary). If the AI system was developed internally, ensure the development of the internal capacity to monitor and audit the tool (with external guidance if needed), for example, by establishing “mechanisms that facilitate the system’s suitability, such as ensuring traceability and logging of the AI system’s processes and outcomes.”42 Furthermore, “ establish processes for third parties (e.g., suppliers, consumers, distributors/vendors) or workers to report potential vulnerabilities, risks or biases in the AI system.” 43
2.1.8. Proactively disclose key information about the AI systems used by the judiciary. Publish information about the adopted AI systems, how they operate, how they are used, and the implications when such tools are deployed to automate decisions or support decision-making processes. For example, the bodies that govern the judiciary can publish an online repository with key information about the AI systems adopted to administer justice. The repository can include data on how the AI systems work, who uses the tools, how they are used, and the implications for judicial decision-making processes.44 The information on the repository should be updated periodically, at least on a yearly basis, and it should be clear when the last time that the repository was updated.
2.1.9. Conduct algorithmic impact evaluations of AI systems that have been deployed. Evaluate the effects of the deployed AI systems on users, affected populations, and society. The latter should include evaluating the impact on human rights and identifying the system’s main effects (intended and not intended) over diverse user groups and populations.45
2.2. On internal procedures and standards:
2.2.1. Human intervention. Ensure that the system allows for human intervention. In other words, human control and supervision, or at least monitoring, should occur during all the implementation stages and usage of the AI system.
2.2.2. Risk management systems. Regarding the AI systems that are used in decisionmaking processes, establish a risk management system that enables the organization to identify, detect, monitor, classify, diagnose, control, and prevent risks as well as mitigate harm. The management system should assign clear roles, responsibilities, and procedures along the life cycle of the AI system. In case of defective AI systems or systems that appear to be producing negative effects, the deployment should be suspended while an investigation is conducted.
2.2.3. Cybersecurity-enhancing measures. Adopt technical, managerial, and human measures to prevent, control, and mitigate cybersecurity risks and incidents. When the deployment and use of the AI system entails accessing cloud services, ensure that the level of protection offered by the provider matches local legal standards as well as organizational security standards.
2.2.4. Data governance. Establish a robust data governance framework and infrastructures to protect personal data and promote responsible data-sharing practices for personal and non-personal data. The organization should also establish “data protocols governing data access. These protocols should outline who can access data and under which circumstances. Only duly qualified personnel with the competence and need to access individual’s data should be allowed to do so”46
2.2.5. Publish impact evaluations and performance reports. Carry out continuous and regular risk assessments of AI systems and publish periodic reports on the impactsand performance (e.g. effectiveness and efficiency) of the adopted AI systems in meeting the organization’s objectives.
2.2.6. Enhanced privacy protections. Given the sensitive nature of personal and legal data that is handled in judicial processes, it is essential to have stringent privacy protections in place. (i) Data minimization. To mitigate the risk of data breaches, AI systems that require minimal personal data to function, particularly in cases involving sensitive personal information should be used. Hence, the organizations of the judiciary should deploy AI systems that are based on privacy enhancing processes.
(ii) Consent protocols. Pursuant with local data protection and privacy laws, draft and implement consent protocols for treating personal data with AI tools within the judiciary. This will ensure that parties are informed and can control how their data is used.
(iii) Data anonymization. Anonymization techniques should be employed when AI systems process personal data, especially in creating databases used for legal analytics or precedent study (e.g., anonymizing the parties’ personal data from court decisions before the rulings are published). 47 However, this should be balanced with the need for having open datasets as well as the right to freedom of expression and access to information so as not to impede the development of digital justice systems.
How these rights are balanced in practice depends on how each jurisdiction regulates the right to access courts documents and anonymization rules derived from data protection and privacy laws.
2.3. On human talent:
2.3.1. Provide access to training. Offer training opportunities to the members of the judiciary who will use the AI tools and ensure that they develop core competencies of AI literacy, particularly to identify the suitability of AI tools that can be used to carry out different types of tasks, understand how to use these tools and their potential impacts and risks in relation to the international and regional human rights law and national fundamental rights law, and assess their outputs critically.48
2.3.2. Adopt organization-specific guidelines. Adopt context and organizational-specific guidelines on which AI systems can be used, under what conditions, and the accepted and prohibited uses for specific processes and tasks. Moreover, the guidelines should include protocols for incident reporting and clear information on how each actor is accountable within the organization. The guidelines should ensure members of the judiciary acknowledge and take responsibility for any materials they produce using AI.
2.4. On the use of generative AI systems:
2.4.1. Awareness of functions and limitations. Ensure that the members of the judiciary who will use the generative AI tools are aware of the tool's adequate uses, limitations and the risks associated (e.g. biased, incorrect outputs) with their adoption for the drafting of legal documents and supporting participation in legal proceedings, as well as supporting court management activities (e.g. translation).
2.4.2. Content authenticity and integrity. Ensure the authenticity and integrity of content in the judiciary. To achieve this, several measures should be taken. (i) All AIgenerated legal documents, evidence presentations, or judicial opinions should be clearly labeled as AI-assisted, so that judges, lawyers, and parties involved are aware of the nature of the content. (ii) Robust systems should be implemented to track the development and modifications of AI-generated legal content. This is important for evidentiary purposes, as it ensures that all materials used in court can be verified for authenticity. For instance, if an AI tool has been used to create or modify a legal document, there should be a clear record of when and how it was used. (iii) Certification protocols should be developed for AI tools used in the judiciary, verifying that they meet international and local ethical guidelines as well as each jurisdiction’s legal standards for accuracy and reliability. An example of this could be a certification process for an AI tool that is used to analyze evidence, ensuring that it is accurate and reliable enough to be used in a court of law.
2.4.3. Usage restrictions. Specific guidelines should be issued to govern the use of generative AI in the judiciary to prevent misuse and protect the integrity of the legal process. Certain applications of AI should be banned or limited considering their impact on human rights. For example, when the terms of use of a generative AI tool indicate that the user’s prompts will be used by the provider to train its models or that third parties can access these prompts, then the use of such tool should be prohibited or restricted to prevent that the judiciary loses control of who can access confidential information or personal data. Moreover, for instance, the use of AI in certain sensitive areas, such as the unilateral generation of binding legal decisions or the creation of fabricated evidentiary material, should be prohibited. Special consideration should be given to intellectual property rights’ compliance by ensuring that any AI-generated content respects the intellectual property of the original text.
3. Specific guidance for individual members of the judiciary These recommendations are applicable to individuals that are part of the judiciary, including magistrates, judges, justices, judicial officers, and judicial support staff.
3.1. Preparation for the use of AI tools:
3.1.1. AI awareness, AI literacy and capacity building. Individuals should be aware of the functionalities, strengths, and accuracy of AI systems and their limitations, biases, and risks in the context of legal activities. The latter also includes awareness of the liabilities arising from these tools' negligent use. The organizations of the judiciary should ensure that individuals have access to education programs and ongoing training to develop critical AI literacy skills. These skills should allow individuals to understand AI fundamentals, identify the functionalities and purposes of diverse AI systems (particularly those deployed in their organization), understand the limits and risks of these tools, learn how to use the AI systems ethically and responsibly, and critically assess the outputs and performance of the tools.49
3.1.2. Use AI tools that have been tested through algorithmic impact assessments.
As explained in a previous section, the organizations that are part of the judiciary should test AI systems through algorithmic impact assessments, including human rights and ethical assessments. The assessment of AI tools that are meant to support decision-making processes is needed to identify implications for access to justice and individual rights and potential risks. However, if your organization does not assess its AI systems, ensure that you use tested tools (by the provider or expert third parties) and select those that have been the subject of human rights impact assessments.50
3.2. On the use of AI tools:
3.2.1. Avoid overreliance on AI systems to make substantive decisions. Do not rely entirely on AI systems to make decisions on the merits of a case nor to decide on procedural issues that may affect human rights; instead, use AI tools’ outputs to complement the legal analysis undertaken with other methods and sources of information.
3.2.2. Aligned use. If you are allowed to use AI systems that have not been formally vetted by your organization, make sure that the tool reflects intended values and goals of the administration of justice. Ensure that using these AI tools does not jeopardize the organization’s objectives and the human rights of any individual affected by the tool.
3.2.3. Adhere to the terms of use. Read and comply with the AI systems’ terms of use (published by the external supplier of the tool or provided by your organization when the system was developed internally). These terms usually indicate the adequate uses of the tools, the uses that are not allowed or prohibited, and the risks that should be avoided by deployers and users. However, consult with your organization about the terms of use of AI systems, especially if you consider that some terms may unduly restrict users’ rights, deny transparency or absolve developer of any liability.
3.2.4. Transparency. Provide meaningful information on when the AI tool is used and how its use could affect individuals who are part of a judicial proceeding or beneficiaries of the legal work. Inform what tools were used and their versions. Ensure that materials produced or published using AI are acknowledged and distinguished using quotation marks or citations.
3.2.5. Responsibility and accountability. Take responsibility for using AI tools in legal tasks and adhere to standards for accountability in the use of AI developed by your organization. This recommendation implies both an ex-ante duty to disclose the use of an AI system and an ex-post duty to provide further information when required to determine responsibility
3.2.6. Opportunity to review decisions and contestability. Provide parties, interested parties, or clients with an adequate opportunity to challenge and contest decisions taken with or supported by AI systems (e.g., decisions that are informed by the outputs produced by AI predictive tools), as well as basic information of how the AI system works, how it was trained, what inputs were used to operate the system, and how the outputs produced by the AI tool informed the decision.
3.2.7. Proactive reports for preventing harm. Inform the organization when there is suspicion of malfunctioning or potential or likely negative impacts and stop using the AI system if you notice that it creates potential harms on human rights.
3.3. On the use of generative AI systems:
3.3.1. Protect personal and confidential data. Do not include personal data or confidential information in prompts when using external generative AI tools. 51 Be aware that any information that you input into a public AI chatbot (e.g., as part of the prompt) should be treated as being made available to everyone, among others, because the terms of service of most of the AI companies that provide free access to these AI systems indicate that the inputs will be used to train future models. 52
3.3.2. Main uses for LLMs. LLMs may be used for different tasks including, but not limited to, drafting basic legal documents, writing speeches and presentations, summarizing, translating, making grammatical corrections, modifying the tone of a text (e.g., informal to formal), improving its readability, exploring specific topics, and carrying out administrative tasks (e.g., drafting emails). 53 However, all the previous tasks involve verifying the AI system’s output and cross-checking with reliable sources. AI tools can assist in locating material that you are familiar with and can evaluate on your own. On the contrary, these tools are ineffective means of researching to find information you cannot assess independently. 54
3.3.3. Unreliability as search engines and for legal analysis. Commercial generalpurpose LLMs are not reliable sources of information or adequate means for conducting legal analysis or carrying out mathematical tasks. “Even with the best prompts, the output may be inaccurate, incomplete, misleading, or biased.”55
3.3.4. Awareness of LLMs’ limitations and risks. Be aware that the outputs generated by current LLMs may include incorrect, imprecise, or fictitious information about factual, legal (laws and case law), and technical issues. LLMs can produce answers that lack any reference to the real world and nonsensical text. 56 Moreover, members of the judiciary should be conscious of potential bias in the text produced by LLMs: “Imbalances in the training data can also lead to biases in the model.”57 Additionally, certain LLMs may have been trained on data up to a certain point in time (e.g., data available up to the point when they were last trained). Therefore, information about recent legal cases, laws, and facts may not have been included in the training data; thus, the output produced by the LLM-powered chatbot may be outdated or inaccurate.
3.3.5. Verify outputs before using them. The convincing structure of the text produced by an LLM should not lead to excessive trust in the factuality and veracity of the output.58 “For various reasons, LLMs offer no guarantees regarding the factuality, quality, and desired formatting (e.g., specific code format) of their outputs.”59 Therefore, avoid excessive trust in AI tools by always ensuring that the output generated by generative AI systems is verified and contrasted with reliable sources of information before its use in legal materials and documents.
3.3.6. Transparent use. Disclose the use of generative AI systems for drafting text – rulings, opinions, and other documents that may have legal consequences – or when it is explicitly used in court hearings. For that purpose, distinguish the text produced by the AI chatbot used in a decision by employing quotation marks and a citation system.60
3.3.7. Integrity. Do not attribute synthetic text as your own. Additionally, the members of the judiciary should prevent potential infringements of copyright and intellectual property rights associated with the use of content produced by generative AI systems.
3.3.8. Responsibility. Take responsibility for using the output produced by generative AI systems to draft rulings and judicial decisions, as well as to inform your participation in court hearings.
-----------------
16 For an explanation of what does “respect, protection and promotion of human rights and fundamental freedoms and human dignity” entail in the context of AI technologies, refer to the Preamble and sections III.1 and III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence. For ethical impact evaluation guidance, see UNESCO’s “Ethical Impact Assessment: A Tool of the Recommendation on the Ethics of Artificial Intelligence”. For an overview of how human rights that may be affected by generative AI, see OHCHR (n 9).
17 For an explanation of the principle of “fairness and non-discrimination”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
18 Ibid.
19 For examples on how AI tools have been use for automating the anonymization of judicial decisions, see UNESCO, ‘Global Toolkit on AI and the Rule of Law for the Judiciary’ (n 1); Terzidou, ‘Automated Anonymization of Court Decisions: Facilitating the Publication of Court Decisions through Algorithmic Systems’ (n 6).
20 For an explanation of the principle of “privacy and data protection”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
21 For an explanation of the principle of “proportionality and no harm”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
22 For an explanation of the principle of “safety and security”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
23 For a risk analysis published for authorities that aim at incorporating generative AI models into their workflows, as well as possible countermeasures, see BSI, ‘Generative AI Models. Opportunities and Risks for Industry and Authorities’ (Federal Office for Information Security - Bundesamt für Sicherheit in der Informationstechnik (BSI) 2024) <https://www.bsi.bund.de/SharedDocs/Downloads/EN/BSI/KI/Generative_AI_Models.pdf?__blob=p ublicationFile&v=4>. For more information about access to information laws, see UNESCO’s webpage: https://www.unesco.org/en/access-information-laws.
24 For an explanation of the principle “Awareness and literacy, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
25 For an explanation of the principle of “transparency and explainability”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence. For an overview of diverse algorithmic transparency instruments adopted by the public sector worldwide, see GPAI, ‘Algorithmic Transparency: A State-of-the-Art Report’ (Global Partnership on Artificial Intelligence (GPAI) 2024) <https://www.gpai.ai/projects>.
26 For a thorough discussion on what does “accountability” means in the context of AI governance, see Claudio Novelli, Mariarosaria Taddeo and Luciano Floridi, ‘Accountability in Artificial Intelligence: What It Is and How It Works’ [2023] AI & SOCIETY <https://doi.org/10.1007/s00146-023-01635-y>.
27 See AI HLEG, ‘Ethics Guidelines for Trustworthy AI’ (European Commission 2019) AI High-Level Expert Group on Artificial Intelligence (AI HLEG) <https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=60419>. For an overview of the scope and process of algorithmic audits, see Ada Lovelace Institute and DataKind UK, ‘Examining the Black Box: Tools for Assessing Algorithmic Systems’ (2020) <https://www.adalovelaceinstitute.org/report/examining-the-black-box-tools-for-assessingalgorithmic-systems/>.
28 For an explanation of the principle of “transparency and explainability”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
29 Ibid.
30 See AI HLEG (n 28) 17.
31 AI HLEG (n 28).
32 For an explanation of the principle of “human oversight”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence. See also ibid.
33 See ibid.
34 For an explanation of the principle of “Responsibility and accountability”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence.
35 For an explanation of the principle of “Multi-stakeholder and adaptative governance and collaboration”, refer to section III.2 of UNESCO’s Recommendation on the Ethics of Artificial Intelligence. See also, AI HLEG (n 28).
36 For an overview of algorithmic impact assessment tools and processes, see Dillon Reisman and others, ‘Algorithmic Impact Assessments Report: A Practical Framework for Public AgencyAccountability’ (AI Now Institute 2018) <https://ainowinstitute.org/publication/algorithmic-impactassessments-report-2>; Ada Lovelace Institute and DataKind UK (n 28).
37 Further, the potential implication of AI tools for the information ecosystem may require a more specific risk assessment based on UNESCO, ‘Guidelines for the Governance of Digital Platforms: Safeguarding Freedom of Expression and Access to Information through a Multistakeholder Approach’ (The United Nations Educational, Scientific and Cultural Organization (UNESCO) 2023).
38 Alessandro Mantelero, ‘Human Rights Impact Assessment and AI’ in Alessandro Mantelero (ed), Beyond Data: Human Rights, Ethical and Social Impact Assessment in AI (TMC Asser Press 2022) <https://doi.org/10.1007/978-94-6265-531-7_2>; Alessandro Mantelero and Maria Samantha Esposito, ‘An Evidence-Based Methodology for Human Rights Impact Assessment (HRIA) in the Development of AI Data-Intensive Systems’ (2021) 41 Computer Law & Security Review 105561; Government of the Netherlands, ‘Fundamental Rights and Algorithms Impact Assessment (FRAIA)’ (2021) <https://www.government.nl/documents/reports/2021/07/31/impact-assessmentfundamental-rights-and-algorithms>; Government of Canada, ‘Algorithmic Impact Assessment Tool’ (2023) <https://canada-ca.github.io/aia-eia-js/>.
39 AI HLEG (n 28).
40 “Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, […] skin type […] and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information.” Margaret Mitchell and others, ‘Model Cards for Model Reporting’ (2019). See also: GPAI (n 26); Matías Valderrama, María Paz Hermosilla and Romina Garrido, ‘State of the Evidence: Algorithmic Transparency’ (Open Government Partnership and GobLab (Universidad Adolfo Ibáñez) 2023) <https://www.opengovpartnership.org/documents/state-of-the-evidence-algorithmic-transparency/>; Open Government Partnership, ‘CHAPTER 8: Algorithmic Transparency’, The Skeptic’s Guide to Open Government (2022).
41 UNESCO, ‘Global Toolkit on AI and the Rule of Law for the Judiciary’ (n 1) 42.
42 AI HLEG (n 28) 31.
43 ibid.
44 For a review of diverse algorithmic transparency instruments, with an emphasis on repositories or registers of public algorithms, that aim at disclosing information about the AI systems adopted in the public sector, see GPAI (n 26).
45 For an overview of algorithmic impact evaluations, see Ada Lovelace Institute and DataKind UK (n 28).
46 AI HLEG (n 28) 17.
47 See Terzidou, ‘Automated Anonymization of Court Decisions: Facilitating the Publication of Court Decisions through Algorithmic Systems’ (n 6).
48 Impact evaluation instruments, such as UNESCO’s “Ethical Impact Assessment: A Tool of the Recommendation on the Ethics of Artificial Intelligence” and the human rights impact assessment framework proposed by Mantelero, may be used to evaluate the impacts on human rights of tools that are used by judicial operators. See, Mantelero (n 41).
49 Juan David Gutiérrez, ‘¿Están Los Jueces En Capacidad de Usar Modelos de Lenguaje a Gran
Escala (LLMs)?’ (2023) 7 Revista EXCEJLENCIA 10; David Bawden, ‘Origins and Concepts of Digital
Literacy’ in Colin Lankshear and Michele Knobel (eds), Digital literacies: concepts, policies and
practices (Peter Lang 2008) <https://search.ebscohost.com/login.aspx?direct=true&db=edshlc&AN=edshlc.011566630.3&site=e ds-live> accessed 17 April 2023.
50 Mantelero (n 41); Mantelero and Esposito (n 41).
51 For general recommendations on how judicial operators can assess matters of privacy and data protection, see UNESCO, ‘Guidelines for Judicial Actors on Privacy and Data Protection’ (The United Nations Educational, Scientific and Cultural Organization (UNESCO) 2022) <https://unesdoc.unesco.org/ark:/48223/pf0000381298>.
52 UK Courts and Tribunals Judiciary (n 3) 3.
53 See Courts of New Zealand (n 3); UK Courts and Tribunals Judiciary (n 3) 6.
54 UK Courts and Tribunals Judiciary (n 3) 3.
55 Courts of New Zealand (n 3) 2.
56 The Guidelines do not allude to the term “hallucination” because this use contributes to anthropomorphizing AI systems, particularly chatbots.
57 BSI (n 24) 9.
58 “LLMs generally generate linguistically correct and convincing text and are capable of making statements on a wide variety of topics. This can create the impression of a human-like performance, leading to excessive trust in the statements and the performance of the models (so-called automation bias). For users, this may result in drawing incorrect conclusions from the generated texts or accepting statements without questioning them” ibid 10.
59 ibid 9.
60 For example, for guidelines on how to cite the output of AI chatbots with the APA citation system, see Timothy McAdoo, ‘How to Cite ChatGPT’ (APA Style, 2024).
