Judicial AI: Expanding Uses, Unfinished Governance

Judicial AI and the question of accountability

How can courts preserve judicial accountability, procedural fairness and the parties’ ability to challenge the account of their case when an AI system helps produce that account? This question has become pivotal worldwide as artificial intelligence becomes part of how courts understand cases and prepare decisions. Taking the Colombian approach as its starting point, this article examines the Italian experience within the European regulatory framework and identifies lessons for institutional governance.

As in many other countries, the Judicial Council has set out measures for implementing (art. 11) and analysing (art. 12) institutional AI solutions or projects (Consejo Superior de la Judicatura, 2024). Yet the Colombian approach innovates, at least compared with European experience, by combining implementation guidelines with two levels of disclosure (Art. 10). A general duty requires judicial personnel to expressly report the use of generative AI to support institutional work, tasks, or activities. More detailed disclosure is required especially when those activities affect the preparation and issuing of substantive decisions. Officials must identify the tool, model and provider, record the date of use, confirm that they have read, understood and accepted its terms and conditions, and explain what it was used for and how, including the prompts or instructions given. At this reinforced level, they must also identify where AI outputs have been reproduced, transcribed, or incorporated, using quotation marks, bold, italics, or other means to clearly distinguish passages produced directly by the tool (Consejo Superior de la Judicatura, 2024). These requirements make both the use of AI and its contribution to the decision visible, providing a basis for scrutiny within the proceedings. The Council of State’s judgment from 10 August 2026 shows how this approach works in practice. The Third Section, Subsection C, acting as a constitutional judge, found the tutela action inadmissible because it did not meet the requirements of constitutional relevance and subsidiarity. The court also found no procedural defect in using AI to help organise claims in joined cases. The process involved manual organisation, then AI-assisted checking and systematisation, and finally human review and adjustment. This judgment demonstrates how the use of AI and human review can be examined through procedural scrutiny. (Consejo de Estado, 2026). Without going into the details of the case, the judgment matters because it makes the sequence of AI assistance and human verification visible and open to procedural scrutiny.

Italy: deployment ahead of governance

The Italian experience helps identify the institutional arrangements needed to make safeguards effective: responsibilities for selecting and introducing systems, testing their performance, preparing users and reviewing actual use. The comparison concerns different emphases within two approaches that combine institutional responsibilities with duties for individual judges.

The Italian experience examined here draws on a study of judicial AI initiatives between 2020 and mid-2026 (Contini, 2026) and analyses AI governance at the crossroads of the constitutional responsibilities of the Ministry of Justice and the Judicial Council. According to the Italian Constitution (Art. 110), the Ministry manages the organisation and services that support the courts and judicial activity, including administrative staff, infrastructure, budgets, and digital systems. The Judicial Council (CSM) safeguards judicial independence and governs the careers of judges and prosecutors, including appointments, assignments, transfers, promotions, and disciplinary decisions (Art. 104-105). With these powers, the Council also approves regulations required to ensure the organisation of courts while protecting independence. Under this arrangement, the Ministry first developed digital technologies for case registration and case and court management, under ministerial leadership. Later, the development of e-justice platforms involved the Council in various phases, reflecting both implementation difficulties and their growing effects on judicial activity. Digital systems determine which documents users can access, how information is organised, and which actions are available. They shape the conditions under which adjudication takes place. These effects, already visible in electronic justice platforms (Reiling & Contini, 2022), may extend with AI to the selection, classification, and synthesis of material, and to the extraction of key case-related information used in legal reasoning, among other possible uses. Hence, the Ministry and the Council should coordinate and agree on AI development and deployment (Castelli, 2024).

The Ministry’s limited results in developing applications for judicial work made this need for coordination more pressing. For example, AI-driven anonymisation of the judicial decisions database accessible to lawyers and citizens often rendered the text unintelligible. Lawyers challenged both readability and access arrangements. Judges retained access to a fuller information resource, while the public database offered altered texts. The controversy showed how a technical choice could affect legal knowledge and the practical conditions of defence (Contini, 2026).

Alongside these limited institutional results, Italian magistrates began experimenting independently with commercial large language models (LLMs), creating dispersed practices of shadow AI. Later, in spring 2025, the Copilot icon appeared on judges' and prosecutors' computers. Without any official notice, generative AI became available through the commercial work environment already provided to judicial users. Its availability in spring 2025 preceded clear communication about acceptable uses, functional limitations and generalised training. In January 2026, a ministerial communication addressed the situation. It announced the introduction of the application as an experiment, clarified that use was permitted only for administrative and ancillary tasks, and anticipated the launch of dedicated training (Contini, 2026).

This sequence exposed a gap between deployment and the preparation required for responsible use. It also sat uneasily with the AI literacy obligation already in force from February 2025 under the EU AI Act. The evidence supports criticism of deployment before adequate preparation. Assessing specific legal breaches would call for further evidence about the measures taken for particular users.

The wider institutional consequence was growing dependence on a general-purpose commercial platform in the absence of dedicated applications. The provider controls product design and functions. Meeting the specific requirements of judicial work therefore depends on the scope for configuring the system and on the working procedures adopted by judicial offices. When a system is not adaptable to users' requirements, users must adapt their tasks to the system.

These deployment problems reflected an earlier weakness in the allocation of institutional responsibility. The Ministry initially introduced AI into the judicial working environment through its own initiative. The Observatory on AI in judicial activity, established in July 2024, initially lacked Council representatives. The designation of two Council members in June 2025 partially addressed this gap, while leaving the wider allocation of responsibilities open.

The Council, in turn, was neither proactive nor reactive. The Council developed its position progressively, with its first dedicated recommendations in October 2025. These recommendations appeared several months after Copilot became available and after at least two years of independent experimentation by magistrates.

These first recommendations focus on permissible uses and institutional safeguards, without mentioning disclosure. Their emphasis provides a basis for comparison with the Colombian approach, where disclosure connects AI use to scrutiny in individual proceedings.

Shared governance and expanding uses

The developments mentioned in the previous section made the need for shared governance increasingly evident. The Ministry provides infrastructure, security, procurement and technical capacity. The Judicial Council is responsible for judicial independence and the organisation of judicial work. The legal profession contributes knowledge of defence rights, procedural practice and the needs of those using the courts. Their responsibilities intersect wherever technology shapes case handling (Contini, 2026).

The Council’s role therefore extends to the conditions under which systems should be selected, tested and introduced. Rules addressed to individual judges depend on earlier technical and organisational choices. A requirement to check sources, for example, becomes workable when the application provides reliable access to the authentic source: while judges have access to the complete, non-anonymised database, lawyers have access only to its anonymised, unintelligible version. Hence, the Bar should also be active in AI governance.

The October 2025 Council recommendations recognised an intermediate area between organisational assistance and judicial activity. They identified uses including research, summaries of publicly accessible decisions, administrative support, language revision, and translation. They also called for cooperation with the Ministry, controlled experimentation, periodic audits, and discussion with the legal profession. The Council emphasised the time required for critical review and the judge’s independent assessment (Consiglio Superiore della Magistratura, 2025).

In 2026, cooperation between the Ministry and the Council eased deployment. Cooperation proceeded through the Joint Committee (Cavallini, 2024)¹, and the Ministry provided technical assurances about the Microsoft environment available to judges, including data storage in Italy, separation from other customers’ environments, and exclusion of judicial data from model training. Despite how easily these technical safeguards can be circumvented under the US Cloud Act (Rojszczak, 2022), the Council authorised experimental uses involving judicial documents, subject to cumulative conditions (Consiglio Superiore della Magistratura, 2026).

The expanded activities include classifying procedural documents, analysing the content of parties’ submissions and reconstructing their positions. They also include comparing those positions to identify the issues for decision and the facts requiring proof. These functions place AI close to the intellectual preparation of adjudication and allow comparison with the assistance examined in the Colombian judgment.

Conditions include verifying every output before use, mandatory approval points that halt automatic progression in structured workflows, structured prompts where feasible, and prior technical training. Judicial case assessment and judgment remain the magistrate's responsibility. Unlike the Colombian experience, the Council guidelines do not mention a duty to disclose AI use, leaving lawyers unable to challenge its use. This development contains a significant tension. The same document records weaknesses in Copilot’s connection to documentary sources, invented information and failures to signal gaps. The Council also considers standard Copilot inadequate as the long-term solution for judicial AI (Consiglio Superiore della Magistratura, 2026).

The expansion therefore rests on individual users' capacity to implement safeguards and manage recognised weaknesses. In the new guidelines, data protection assurances help justify processing documents within the institutional environment. Accurately reconstructing those documents requires additional controls.

The Italian Ministerial monopoly is now eased by communication and discussion with the Council. At the same time, permitted use has expanded to include case analysis and is therefore very close to adjudication, subject to check and revision by judicial officers.

Making safeguards work in practice

The question becomes how to obtain evidence that actual use remains within the authorised limits. A judge may retain final responsibility, relying on a summary that has already changed the apparent importance of an argument. An omitted claim or a contested allegation presented as an agreed fact can influence the later analysis. This influence occurs during the construction of the account of the case in factual and legal terms, where parties present their claims, arguments and evidence to the court.

Appropriate governance must therefore link technical safeguards, professional competence and procedural scrutiny. The recommendations examined leave important implementation questions open. The following proposals develop the Council’s requirements into practical arrangements that could also inform institutional implementation in Colombia.

Independent audits must examine two levels. A technological audit should assess the system as configured for judicial use: access permissions, data processing, available sources, links between outputs and documents, operational records, and the functioning of mandatory approval points. Tests should examine omissions, incorrect attribution, and invented references using realistic case materials.

An audit of use should assess whether users follow the approved procedure, consult the originals, identify errors and complete the required checks. Evidence may include exercises, interviews and appropriately protected samples of operations.

The audit mandate should focus on systems and working practices while safeguarding judicial independence, confidential information and deliberation. This assessment should not affect the status or careers of the judges involved, nor should it trigger a disciplinary investigation. Findings should lead to corrections, additional training or restrictions on specific uses. Audit is not a one-off step. Because Copilot (like any LLM) evolves quickly, audits ought to be periodic.

Structured and validated prompts can define the task, permitted sources, output format and how uncertainty is handled. For example, an instruction can require the separate identification of each party’s claims, supported by the original quote and a page indicator. It can also require the system to distinguish allegations from facts and to flag materials not taken into account.

The practical objective is an institutionally validated collection of instructions, tested for specified tasks and system configurations. This would give practical effect to the Council’s recommendation to use structured prompts. Institutional validation is proposed here as a way to implement that recommendation. Validation should cover the entire procedure, including the prompt text, inputs, outputs and checks.

Judges and technical specialists should develop these procedures together. Lawyers can support by assessing cases where a summary is correct and cases where it might distort a party’s position. Substantial changes to the prompt, system or configuration should trigger further testing.

Mandatory approval points should prevent a structured workflow from proceeding until the magistrate has reviewed its output. The software should enforce this requirement. Records should identify the version reviewed and any corrections. These records support accountability, while the substance of the review requires further evidence.

Completeness warrants particular attention. Verifying the accuracy of passages selected by the system leaves open the possibility that an important argument was omitted. Review must therefore remain directly connected to the authentic documents. The judge needs sufficient time and practical means to compare the automated account with the case file.

Training could develop the judge's ability to compare the automated account with the case file. Exercises could include omitted claims, confused parties, invented sources, and misleading summaries. Competence should be assessed for the authorised uses. It should also develop the ability to correct outputs, recognise the limits of a task, and stop using the system when reliable verification becomes impracticable.

The Judicial Council’s call for cooperation with the legal profession can be translated into responsibilities at national and local levels. The National Bar Council (Consiglio Nazionale Forense, CNF) should work with the Ministry and the Judicial Council to define requirements for systems that process submissions and evidence. It should also contribute to testing procedures and independent audits, focusing on the accurate representation of claims, arguments and disputed facts. This participation would bring defence rights into decisions on the design, deployment and evaluation of judicial AI.

Local bar associations and judicial offices could develop shared practices through joint working groups and jointly assess outputs. These exercises should examine omissions, incorrect attribution and changes in meaning, including whether a system turns an allegation into a fact or overlooks a subsidiary claim. Their findings would inform the development of structured prompts, verification procedures and training.

These arrangements would also foster an ongoing exchange between professional experience and institutional oversight, helping courts improve their procedures and correct weaknesses as they arise.

Connecting institutional governance with procedural scrutiny

The Colombian and Italian experiences address the common problem of integrating AI into judicial work while protecting fair proceedings. Both combine institutional responsibilities with duties for individual judges, with different emphases that offer complementary lessons.

The Italian experience shows how difficulties in public development and the introduction of a commercial platform exposed weaknesses in concentrated ministerial control. Cooperation with the Judicial Council has gradually produced more detailed safeguards and a wider range of permitted activities. Making those safeguards effective requires tested procedures, competent users, independent assessment and participation by the legal profession. For Colombia, these developments identify institutional tasks that can support its existing implementation provisions and the verification duties placed on judges, and emphasise the need for a shared governance of judicial AI.

The Colombian experience adds a requirement that deserves explicit consideration in Italy: disclosure of judicial AI use. The Council of State judgment illustrates how information about the sequence of human work, AI assistance and subsequent verification can support procedural scrutiny. The Italian recommendations leave important implementation questions open and provide no comparable disclosure procedure. A further step would be to define when judges should disclose AI assistance, what information they should provide and how parties can raise concerns. These requirements should reflect the function performed and its possible influence on the proceeding, particularly when AI helps reconstruct disputed positions.

Disclosure and institutional safeguards would reinforce each other. Disclosure would enable lawyers to identify AI's role in a particular proceeding and question relevant omissions or distortions. Tested procedures, training and audits would provide evidence about the conditions under which the tool was used. Shared professional practices could connect these two levels of accountability while protecting confidential information and judicial deliberation.

A coherent approach would therefore link deployment decisions with approved uses, verified working methods, training, disclosure and periodic review. It should also assign responsibility for correcting or suspending practices when recurring weaknesses emerge.

For Colombia, the Italian experience highlights the institutional work required to support disclosure. For Italy, the Colombian experience shows how disclosure can link institutional safeguards to scrutiny in individual proceedings. Joined up, the two approaches support the judge’s direct and critical engagement with the case file and the parties’ ability to question how their positions are represented.

Note

¹ The Joint Committee (Comitato Paritetico Ministero della Giustizia - Consiglio Superiore della Magistratura) was established with the CSM deliberation of 18 May 2011. It deals with topics that requires the involvement of the two institutions, information technology and statistics, staffing levels, and best practices in organizational and judicial matters. https://www.csm.it/portale/web/csm-internet/comitato-paritetico

References

Castelli, C. (2024). L'agenzia: una nuova governance per i servizi informatici del Ministero. Questione Giustizia. Retrieved from https://www.questionegiustizia.it/articolo/agenzia-governance-it

Cavallini, D. (2024). Il CSM tra ordinamento e organizzazione della giustizia. L’ampliamento delle competenze nella gestione degli uffici giudiziari. Boogna: Bologna University Press.

Consejo de Estado. (2026). Sentencia de tutela, radicación 11001-03-15-000-2026-04632-00. Retrieved from https://codigoabierto.co/documentos/rad-2026-04632.pdf

Consejo Superior de la Judicatura. (2024). Acuerdo PCSJA24-12243. Retrieved from Bogotà: https://rm.coe.int/colombia-guidelines-for-the-use-of-artificial-intelligence-in-the-judi/1680b53484

Consiglio Superiore della Magistratura. (2025). Raccomandazioni sull’uso dell’intelligenza artificiale nell’amministrazione della giustizia. Roma: Consiglio Superiore della Magistratura. https://www.csm.it/portale/documents/21768/87316/linee+guida+intelligenza+artificiale+%28delibera+8+ottobre+2025%29.pdf/e0ed018a-43e2-9dd7-2420-e3d0632d658c?t=1772619014845

Consiglio Superiore della Magistratura. (2026). Aggiornamento delle raccomandazioni sull’uso dell’intelligenza artificiale nell’amministrazione della giustizia. Retrieved from Roma: https://www.csm.it/portale/documents/21768/8276330/NEWS+118-VV-2024+-+aggiornamento+uso+IA.pdf/75cc3c96-3e79-9ec0-037a-b5e2b69bbc11?t=1785311382253

Contini, F. (2026). Governare l’IA nella giustizia italiana: progetti, applicativi e responsabilità istituzionali. CNR-DSU Working Paper(1-26), 1-22. doi:https://doi.org/10.36134/CNRDSUWP-2026-1

Reiling, D., & Contini, F. (2022). E-Justice Platforms: Challenges for Judicial Governance. International Journal For Court Administration, 13(1). Retrieved from http://doi.org/10.36745/ijca.445

Rojszczak, M. (2022). CLOUD act agreements from an EU perspective. Computer Law & Security Review, 38. doi:https://doi.org/10.1016/j.clsr.2020.105442