The End of Judging and the Last Judge: Artificial Intelligence and the Conditions of Judicial Authority

The End of Judging and the Last Judge Artificial Intelligence and the Conditions of Judicial Authority Ian Bernaziuk Judge of the Administrative Cassation Court of the Supreme Court (Ukraine); Doctor of Law; Professor Scientific Expert, CEPEJ-GT-CYBERJUST, Council of Europe ORCID: 0000-0002-2353-4836 Date of manuscript: 14 October 2026 Correspondence: bernaziuk1979@gmail.com Abstract. The growing capacity of artificial intelligence (AI) requires courts to weigh the benefits of automation against the shortcomings of existing adjudication: delay, expense, inconsistent outcomes and weak public confidence. A faster or more consistent system may improve the judge's information without acquiring the legal authority to decide a case. This article proposes a test of substantive

judicial authorship with three cumulative conditions: the lawfully constituted judge forms the dispositive judgment and its decisive reasons, remains free in practice to reject a system's recommendation, and accepts accountability for reasons attributable to the court. Applying constitutional law, European judicial guidance, Ukrainian ethics rules and selected case law, the article distinguishes the lawful deployment of technical tools from the practical transfer of adjudication through default recommendations and nominal human approval. It argues that where these conditions fail, formal human approval cannot preserve judicial authority, and that such a failure may raise a further procedural question: whether, in substance, the case was heard and decided by the

legally empowered composition of the court. The test guides assessment of uses permitted under existing law and identifies the constitutional limit relevant to any future authorisation of more consequential AI assistance.

AI disclosure. OpenAI's ChatGPT and Anthropic's Claude were used for linguistic and structural editing, consistency checks and alternative wording. The author reviewed the AI-assisted material and remains responsible for the sources, legal analysis and conclusions. Further details are provided in the Declarations section.

Keywords: judicial independence; artificial intelligence; adjudication; judicial authorship; human oversight; reasoning; access to justice; Ukraine.

1. Introduction: the end of judging?

Francis Fukuyama's 1989 essay The End of History? and his 1992 book The End of History and the Last Man asked whether a political form could appear to have exhausted its principal alternatives.

The question here concerns a different institution: could a legal system continue to employ judges while progressively removing the exercise of judgment from their office [1; 36]?

Popular culture has already imagined the more visible version of this problem. Simón Casal’s Spanish-Portuguese political thriller Artificial Justice (Justicia artificial, 2024) depicts a referendum on replacing judges with an AI system promoted as a means of automating and depoliticising justice [58]. The question examined here is less dramatic but institutionally more difficult: can judging pass to a system while the judge, the courtroom and the judicial signature remain?

A recent United States executive order illustrates how quickly official terminology can change. Executive Order 14434 of 29 September 2026 directs the United States executive branch, to the maximum extent permitted by law, to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in non-statutory official materials, while for the time being defining SI by reference to the existing statutory definition of artificial intelligence [59]. Changing the terminology, however, does not itself alter the legal allocation of public authority.

This article was prompted by the European Judicial Training Network’s AI Summit “The Impact of AI on the Judiciary” (DIGI/2026/01), held in Riga on 24–25 September 2026. Its programme addressed AI in legal research, evidence assessment, judicial reasoning and court administration [2].

The discussions at the Summit sharpened a practical question: when a judgment has been prepared with AI assistance and is formally adopted by a judge, which part of the judgment has the judge actually decided?

Maria Giuliana Civinini, a CEPEJ expert and retired judge, asks in her presentation whether human oversight of a generated outcome can become mere approval; Tamara Ćapeta, Advocate General at the Court of Justice of the European Union (CJEU), juxtaposes AI-supported research with deciding a case and justifying the judgment [25, slides 18–19; 30, slides 8–9]. Presentation slides are treated here as evidence of individual speakers’ contributions, whereas adopted instruments represent the position of the issuing body; draft initiatives are identified as such and are not treated as binding or adopted standards.

There is a professional tendency to assess AI by its familiar failures while assessing adjudication by its normative ideal. Courts must instead examine both on equal terms. A system that helps identify relevant precedent, detects an unsupported assertion, translates material accurately or reveals a discrepancy in a draft may improve judicial work. Conversely, a judgment delivered after an excessive delay or at disproportionate expense, or one that fails to engage a material argument, cannot be defended by invoking the dignity of the judicial office alone. According to a November 2025 survey by the Razumkov Centre, 66.3 per cent of respondents in government-controlled territories of Ukraine expressed distrust of courts as a system and 22.5 per cent expressed trust [3]. The survey measures

attitudes; it does not establish the causes of those attitudes. Delay, cost, inconsistent application of law and perceived partiality must nevertheless be addressed as concrete institutional concerns rather than dismissed as public misunderstanding.

Civinini warns that accounts of backlog and judicial error can also be used to cast judges as dispensable [44, pp. 9–10]. The response is to address the defects experienced by parties while preserving an independent judge who decides the case and gives contestable reasons [8; 44, pp. 9– 10].

Even if AI were to outperform judges on speed, cost, consistency and error detection in specified tasks, the legal authority to decide a case would remain with a court established by law. Such authority presupposes a meaningful opportunity for the parties to be heard, independent assessment of the record and arguments within the court’s competence, a reasoned choice within the applicable law and the possibility of holding the deciding institution to account. The fact of judicial office and the act of signing cannot establish that these functions were performed. What matters is whether the judge made the dispositive legal judgment and can explain the reasons for it. I call this substantive judicial

authorship: the actual formation and accountable adoption of the outcome and its decisive reasons. The term concerns public judicial authority rather than copyright or the authorship of prose.

Human judgment need not be more accurate than machine output for this distinction to matter. Judicial authority is relational and institutional: a coercive public decision must be attributable to the legally constituted decision-maker who hears the parties, answers their material claims and remains accountable for the reasons given.

By reading constitutional and professional rules alongside European standards, and by applying them to distinct judicial tasks, the analysis identifies when substantive control over a court’s decision may shift from judge to system. Empirical evidence bears on the feasibility of oversight and the deficiencies of existing courts. The legal allocation of authority remains a separate question.

2. Why the existing court must be part of the comparison

The relevant comparison is between actual institutional practices and an alternative that is lawfully designed and independently tested. Research tasks, retrieval of legislation and precedent, anonymisation, transcription, classification, identification of inconsistent citations and preparation of working drafts can consume time without themselves constituting the final judicial choice.

Automating some of this work could reduce the expense and duration of proceedings and improve

access to pertinent material. My earlier research considers how AI-assisted identification and synthesis of case law might support the Supreme Court's responsibility for consistency while leaving final decisions with judges [4]. In an exploratory Lithuanian survey, Barysė and Sarel found stronger support for automating information acquisition than for automation at later stages of adjudication [5].

In focus groups with 12 UK judges, including five members of the UK Supreme Court, participants treated appropriate AI assistance as task- and court-dependent, while repeatedly raising concerns about verification and de-skilling [43].

Consistency is not simply the frequency with which cases receive identical outcomes. Like cases should be treated alike for legally relevant reasons, and unlike cases should be distinguished. A system that finds similar judgments can assist that inquiry. A system that rewards conformity with its own recommended result could obscure a material difference or prevent a justified development of the law. Access also requires more than fast disposal: parties must be able to understand and contest the reasons affecting their rights. A demonstrably faster process that does not allow such contestation may produce efficiency without adjudicative legitimacy.

Nor is human decision-making a neutral benchmark. Human judges can overlook a precedent, differ in the treatment of similar cases, misunderstand technical evidence or fail to identify an internal contradiction. The constitutional case for a human judge should not depend on the claim that these events never occur. At the same time, no conclusion that AI already improves real judicial outcomes follows merely from laboratory benchmarks or vendor demonstrations. A 2026 review focused on criminal-justice risk assessment finds that observed effects of AI decision aids on pretrial and sentencing outcomes are currently modest or nonexistent and that available field evidence does not establish a general pattern of judicial over-reliance; the settings studied and the evidence base remain

too limited for broader generalisation [6]. An empirical study of judicial use of large language models in Shenzhen likewise describes a particular institutional workflow and cannot establish a universal model of judicial adoption [7].

If an independently evaluated system becomes substantially better on a defined supporting task, it may be professionally appropriate to consider its use, subject to lawful procurement, security and procedural rules. Refusing to use a reliable means of locating a decisive authority solely to preserve an old professional routine would not protect the parties. But the move from improving the judge's information to determining the court's choice requires a different justification. Measured performance on a set of cases does not itself confer public authority or resolve who must give reasons to the person bound by a judgment.

Professional practice already raises a further comparison: the rate of AI adoption by legal actors and the rate at which courts govern it. UNESCO's survey recorded 563 responses from judges, prosecutors, lawyers, legal administrators and researchers in 96 countries; 44 per cent of respondents reported using AI chatbots for professional activities, whereas only 9 per cent reported institutional guidance or training [32]. In this surveyed group, practice outpaced reported institutional preparation.

If lawyers submit machine-assisted work while the judge lacks training to identify inaccurate authorities or fabricated material, the procedural burden is transferred to a less prepared part of the system.

Ukrainian cassation practice already illustrates this verification problem. In two rulings delivered in January and February 2026, the Administrative Cassation Court of the Supreme Court examined procedural submissions relying on non-existent Supreme Court decisions or legal conclusions. The Court indicated that such errors may result from the use of generative AI without adequate professional verification, while responsibility for the submitted material remained with the party that filed it [53–54]. These rulings concern responsibility for AI-assisted submissions rather than AI use by judges themselves; they illustrate the verification problem but do not apply the authorship test developed below. The spread of ‘shadow AI’ remains a governance concern [26, slide 23].

In a descriptive study of US federal civil cases, Shah and Levy report that the share of filings by self- represented litigants rose from approximately 11 to 16.8 per cent by fiscal year 2025, while docket activity increased [47, pp. 1–3, 20–22]. The study does not establish that generative AI caused the

increase. It nonetheless raises a capacity question: easier access to filing does not by itself ensure that courts can decide more cases within a reasonable time.

3. The governing legal distinction

Article 6(1) of the European Convention on Human Rights protects a hearing before an independent and impartial tribunal established by law, within a reasonable time. Its terms make both judicial independence and timeliness legally significant [8]. Article 124 of the Constitution of Ukraine assigns the administration of justice exclusively to courts and prohibits the delegation of their functions; Articles 126 and 129 provide additional guarantees of judicial independence and basic principles of proceedings [9]. The constitutional rule concerns the exercise of public power. A machine cannot be made a member of the lawfully constituted panel merely because its output is placed before that panel.

Judicial status and the lawfulness of a tribunal's composition have an independent legal significance. In Guðmundur Andri Ástráðsson v Iceland, the Grand Chamber examined whether a serious defect in the appointment of a judge denied the applicant a tribunal established by law [37]. The case concerns the appointment of a human judge, not machine participation, and supplies no direct holding on AI-assisted reasoning. It shows why the lawfulness of the decision-maker cannot be reduced to the quality of the resulting answer. That premise informs the analysis of judicial authorship without resolving the distinct question whether a judge exercised the assigned authority in a particular AI- assisted workflow.

European institutional texts express the consequence with increasing specificity. CCJE Opinion No 26 (2023) states that judicial decision-making must, “explicitly and implicitly ”, be carried out by judges, and warns that dependence on technology and its designers may impair both individual and institutional independence [10, paras 58 – 61, 92(iii) – (iv)]. CEPEJ's 20 25 guidelines provide that judicial decision making, legal reasoning and evidentiary assessment should not be delegated to generative AI, and confine such systems to auxiliary or preparatory tasks [11, para 55]. These Council of Europe texts are guidance, not an amendment to domestic procedural legislation or a treaty based categorical prohibition on every use of AI.

The EU Artificial Intelligence Act takes a different regulatory form. Annex III, point 8(a), identifies specified AI systems intended to assist judicial authorities in researching and interpreting facts and law and applying law to concrete facts as high-risk systems. Article 14 sets out requirements for effective human oversight, including measures addressing over-reliance on system output; the recitals identify the particular stakes of justice and the continuing role of human decision-making [12, recital 61, art 14, Annex III, point 8(a)]. Under the amended application timetable, Chapter III, Sections 1– 3, except Article 6(5), apply to high-risk systems classified pursuant to Article 6(2) and Annex III

from 2 December 2027 [39, art 1(40)(b)]. Risk classification and compliance duties do not amount to a licence to transfer adjudication to AI. The Regulation also does not replace Ukraine's constitutional allocation of judicial power.

Ukraine's Code of Judicial Ethics permits judicial use of AI where it does not affect judicial independence or impartiality, does not concern the assessment of evidence or the decision-making process, and does not violate the law [13, art 16]. The official Commentary to the Code of Judicial Ethics, approved by the Council of Judges of Ukraine on 2 March 2026, further explains Article 16 as protecting the judge's exclusive responsibility for evidentiary assessment, legal interpretation and decision-making, and rejects the delegation of these functions to AI [62, pp. 184–193]. The Commentary provides an authoritative national ethical reference for distinguishing technological assistance from the substitution of judicial judgment. The institutional context illustrates why these

terms must be interpreted by function and by user. The Supreme Court's 2025 rules for its staff restrict processing protected information in public systems and prohibit automated generation of drafts of procedural acts within proceedings [14]. The State Judicial Administration's 2025 concept for the Unified Judicial Information and Communication System, by contrast, envisages a range of AI functions, including search, legal citation checks and the generation of draft procedural documents [15, section 6.3]. A concept of prospective system functions is not itself authorisation for a judge or member of staff to perform every listed task. These instruments govern different actors and stages;

their interaction nevertheless requires an explicit rule before a generation function is deployed in judicial work. Ukraine thus already combines task-specific ethical rules, institutional safeguards and emerging case law on responsibility for AI-assisted legal work [13–14; 53–54], while further alignment remains necessary before more consequential generation functions are deployed.

The Council of Europe's Framework Convention on AI also requires attention to judicial independence, transparency, accountability, effective remedies and risk management within its field of application. It is a framework for public-law obligations rather than an operational allocation of particular drafting tasks [16, arts 5, 8–9, 14–16]. The appropriate doctrinal task is therefore to specify what counts as the judge's decision in a workflow where many preparatory operations may be automated.

The CJEU's 2023 AI strategy established an AI Management Board, while its Artificial Intelligence Ethics Charter was adopted on 26 January 2026 [33, pp. 1–2]. Ćapeta reported that version 1.0 of an internal AI system, Curia AI Brain, had launched on 15 September 2026; the presentation lists document summaries, text correction and a case-law research interface among its functions [30, slides 6–7]. The Charter states that AI systems “shall not be entrusted with decision-making responsibilities” and places final responsibility for an AI-assisted product on an identifiable person [33, p. 3]. Its requirements for traceability, confidentiality, quality control and verification show how a court can

govern assistance while reserving adjudication to its judges [33, pp. 2–4]. At the Summit, Wiggo Storhaug Larssen, Vice-President of the CCJE, described seven draft principles for the proposed CCJE Opinion No 29: human-rights compatibility, non-delegation, judicial digital sovereignty, proportionality, strategic use, transparency and systematic training. The draft also distinguishes prohibited uses, including automated decision-making, from high-risk uses such as tools capable of factual evaluation or assessment of evidence [29, slides 6–7]. The draft remains subject to plenary approval scheduled for late November 2026 [29, slides 4–8]. The adopted Opinion No 26 and CEPEJ Guidelines remain the relevant adopted Council of Europe guidance [10–11].

4. Substantive judicial authorship

The proposed criterion has three cumulative dimensions. First, substantive determination: the lawfully assigned judge or panel must itself resolve the factual and legal issues within its procedural competence, determine the operative outcome and adopt the decisive reasons. A judge may consider a system's suggestions or a draft from judicial staff, but must reach a position that can be explained from the record and law without treating the machine's recommendation as presumptively correct.

Second, procedural freedom: the judge must have a practicable opportunity to inspect sources, identify contrary arguments and reject or disregard the output without technical obstruction or institutional penalty. Third, accountable attribution: the judge must be able to defend the decisive reasons as the court's own, and the record must allow proper review of a material algorithmic contribution where the applicable rules require its disclosure. This third condition also reflects the established Article 6 requirement that judicial decisions adequately state their reasons and respond specifically to arguments decisive for the outcome of the proceedings [49, para 84]. Human oversight establishes participation; substantive judicial authorship requires the judge's actual exercise of

authority over the outcome and its decisive reasons. The test concerns control over the outcome and its reasons, not who first typed a sentence. Drafting by another person does not necessarily displace the judge; neither does drafting by a system necessarily do so. Conversely, a judge can personally rewrite an AI-generated paragraph while retaining its unexamined legal premise. The location of authorship is established by the actual choice among material interpretations and contested conclusions, the capacity to explain that choice, and the conditions under which alternatives were considered. The relevant line runs between assistance and the transfer of the dispositive choice. This distinction does not authorise uses of AI that are otherwise

prohibited by applicable law or judicial ethics rules; it determines whether an otherwise lawful workflow preserves the judge's adjudicative authority.

An illustration shows the difference. A system retrieves five relevant judgments and flags a missing argument. The judge checks the sources, hears the parties on the new issue where procedure requires it, determines their relevance and independently explains the result. The system has assisted adjudication. In a second workflow, a system ranks evidence by credibility, supplies a disposition and prepares the decisive reasons. The judge checks citations and signs within a prescribed time, with no genuine capacity to reopen the contested factual or legal issue. The operative judgment has passed to the system. A human signature cannot cure that transfer of authority. Between these cases lies a question of degree, assessed through the assigned task, the information shown to the judge, the time

and competence available for verification, and the power to depart. Judgment T-323/24 of the Constitutional Court of Colombia illustrates why the sequence of work matters. The Court found that the judge had identified the applicable law, assessed the case and reached his conclusion before consulting ChatGPT. The system had therefore not replaced the judge as decision-maker. The Court found shortcomings in transparency and responsibility but no breach of privacy. It nevertheless considered the decision valid [42, paras 374–376].

Paar's account of human oversight identifies declining vigilance and skill, categorises assistance by judicial task, and argues for a clearer division of labour rather than deeply interwoven human–AI work [17, pp. 1393–1403]. Green locates limits in policies that rely on an individual human overseer without institutional justification [18]. Popotas emphasises judicial independence and the danger that a judge may become an auditor of a system's result [19]. Those arguments describe risks and preferred allocations of work. The test developed here supplies a distinct legal inquiry for an actual workflow: did the assigned judge determine the result and its decisive reasons; could the judge reject a

recommendation without penalty; and can the decision be attributed to an accountable court? Even where a division of labour appears clear on paper, an outcome-ranking interface or a requirement to follow its recommendation can defeat the second condition. Conversely, a limited informational contribution may satisfy all three. The test evaluates whether a court has retained its legally assigned function after the technology has been introduced.

The proposed test also differs from approaches based on meaningful human oversight or judicial cognitive legitimacy [61]. These approaches examine whether judges can understand, verify and challenge algorithmic contributions. Substantive judicial authorship addresses a further legal question: whether the judge actually exercised the authority to determine the outcome and its decisive reasons. The distinction matters when assessing the procedural consequences of AI involvement.

Judicial authorship is not authorship of prose. Questions of authorship in AI-assisted scholarship raise a different issue of personal intellectual contribution and responsibility [20].

Practical guidance on AI-assisted judicial drafting already speaks of judges retaining ‘authorship, control, and responsibility’ [51, p. 1]. The concept developed here is different in kind: it is a public- law concept concerning who actually exercises the adjudicative authority to determine the dispositive judgment and its decisive reasons.

The Shenzhen study illustrates the difference between drafting reasoning and forming the judgment: judges first decide disputed issues, an LLM then generates reasoning based on those decisions, and judges review and modify that reasoning before the final judgment. Liu and Li warn that such a workflow may reinforce an initial judicial judgment rather than expose it to genuine reconsideration, although their experimental evidence is mixed rather than conclusive [7].

My recent comparative review of AI use in constitutional adjudication proposes a six-level classification of technological functions; the present article addresses the separate legal question of when their practical operation may displace the judge's adjudicative authority [63].

5. The point at which a recommendation acquires authority

An apparently optional recommendation can become practically compulsory. This occurs when a system's result is presented as a benchmark for correctness, departures trigger an exceptional approval process, or a judge's deviation rate becomes an adverse performance measure. Reliance on the system's recommendation becomes the default; departure requires special justification. At that point,

formal freedom to disagree may coexist with practical pressure to conform. CEPEJ's 2025 Guidelines state that AI-generated recommendations, suggestions and outputs shall never be binding on judges, and that judges should not be required to give a specific justification for departing from them [11, para 58]. CEPEJ identifies a corresponding risk even in decision-support systems when judges are encouraged not to question an AI result [46, risk 6]. That rule has implications for system design, case-management targets and evaluation of judges.

Paragraphs 43 and 57–58 of the same Guidelines address two forms of technological integration that require careful distinction. Paragraph 43 concerns such integration whose operation is predetermined by procedural law, without a separate act by each judge to activate the system. Paragraphs 57–58, however, preserve the judge's right to opt out of AI suggestions and reject any obligation to use recommendations [11, paras 43, 57–58]. These provisions regulate different decisions: a court may supply a uniformly deployed technical facility, while the judge must remain free from a default substantive recommendation. The second condition of substantive judicial authorship identifies the boundary. Automatic transcription, routing or citation retrieval may form part of a legally prescribed

workflow if their errors remain correctable. A legal analysis, credibility score or proposed disposition cannot acquire a binding default through the same architecture. Where the interface displays such recommendations automatically, judges must be able to avoid, set aside and depart from them without adverse evaluation. The legal question concerns the recommendation's practical influence on the dispositive choice, not whether a user pressed an activation button [11, paras 43–44, 55, 57–59; 35, para 13].

CEPEJ's Assessment Tool separately identifies forced use of AI as a risk where a system is integrated into judicial work and opting out is unavailable or difficult [46, risk 14]. The right to reject an analytical recommendation must therefore be effective in the court's ordinary workflow, including when the judge's departures are evaluated. Formal permission to intervene is insufficient if the judge lacks the cognitive and professional capacity for effective review [25, slide 18].

The Court of Justice's judgment in SCHUFA Holding (Scoring) supplies a functional analogy. A credit score may itself amount to automated individual decision-making under Article 22 of the GDPR where a bank relies strongly on it [31, paras 46–50, 60–62]. The Court rejected an interpretation that treated the score as merely preparatory when that characterisation would permit the protection of Article 22 to be circumvented [31, para 61]. Likewise, CEPEJ's reference to ‘auxiliary or preparatory tasks’ cannot by itself establish that a judge retained the dispositive choice [11, para 55]. Article 22 does not govern judicial authorship; the relevant question here is whether the judge formed the outcome and could reject the system's proposed conclusion.

A judge’s duty to give reasons for a judgment does not entail a duty to justify every departure from an AI recommendation. A judge who departs from binding precedent must comply with the applicable law on precedent and reasoning; an AI suggestion does not acquire the status of precedent by identifying patterns in prior decisions. The ordinary duty is to justify the decision by law, evidence and responses to the parties' material submissions. Recording an individual judge's rate of agreement with the model as an efficiency score would risk importing a separate, unlegislated criterion into judicial evaluation [10, paras 58–61; 11, paras 57–60].

The influence of a recommendation must be established through evidence about its use. In three studies of public-sector decision-making, Alon-Barkat and Busuioc found selective adherence under particular conditions rather than general automation bias [21]. The interface, work incentives, decision trail and rate of correction supply better evidence than a presumption that every judge follows a model or that every judge resists it.

The CJEU Charter identifies over-reliance as a risk to the development of case law, attention to other legal systems, the standing of different legal cultures and the quality of decisions [33, p. 1]. A system optimised to reproduce the existing distribution of outcomes can impede a court from giving a justified new interpretation or recognising that earlier decisions failed to address an affected group.

A model trained for a different legal system may carry assumptions from one legal culture into another, a risk expressly identified by CEPEJ [46, risk 1]. A justified departure may therefore improve legal consistency. The judge's reasons must be assessed in law and on the case record, rather than

against a compliance statistic. CEPEJ's Assessment Tool identifies judge profiling and forum shopping as risks to judicial independence, including attempts to identify a favourable forum or tailor submissions to a judge's profile [46, risk 3]. Drăgușin likewise highlights profiling and automated scoring in the reuse of judicial data [27, slides 6–7]. The CEPEJ curriculum addresses training court leaders in performance management; such training must distinguish lawful monitoring of case processing from pressure to concur with model recommendations [35, para 38].

6. Oversight must be institutionally possible

An instruction to “verify the output” has little meaning if the judge lacks access to primary materials, relevant technical information, adequate time or a real choice of workflow. Oversight should be designed before deployment. At minimum, the judiciary should determine the permitted purposes of a system, which information it may process, the quality and legal status of its sources, how generated material is identified, who may modify the model and its prompts, and how consequential errors can be detected, corrected and contested. An audit log must serve effective review while respecting deliberative confidentiality and data-protection obligations. Judicial governance must be able to suspend a function that changes materially after an update [10, para 92(iv); 11, paras 51–61].

CEPEJ's Assessment Tool is intended for use before the acquisition or development of a judicial AI system and when the system is updated [46, p. 3]. It identifies loss of access to a judge and diminished judicial accountability through non-explainable AI as distinct risks [46, risks 6, 12]. Procurement and change control should therefore examine who will determine contested results, whether judges can depart from system output in practice and how the court will account for decisive reasons.

This is a matter of independence as well as information security. Where an executive department or supplier alone controls relevant data, model updates, access rights and performance indicators, it can shape the court's range of practical options without issuing a formal instruction. The judiciary should therefore participate decisively in specification, evaluation and change control; it need not build every system itself. My earlier account of sovereign AI considers control over data, infrastructure and the ability to change providers as elements of institutional resilience [22]. In Italy, for example, the CEPEJ curriculum reports that judges receive Microsoft Office, including Copilot, in a software

package purchased by the Ministry of Justice [35, n. 8]. Such an arrangement is not in itself problematic, but it makes governance of permitted uses, updates and court data a concrete institutional question.

In Rowicz, Advocate General Spielmann considered publication of the allocation algorithm and allocation reports sufficient without publication of source code, and found no specific evidence of improper executive interference in the allocations at issue [38, paras 65–69]. His assessment concerns the assignment of cases under predefined rules. A tool supplying legal reasons or a proposed disposition requires scrutiny of its sources, changes and influence on the judge's choice. Courts must combine judicial control with external technical scrutiny appropriate to fundamental-rights risks.

Banks argues that individual judges cannot carry the entire oversight burden for high-risk judicial decision support [23].

Presentation materials from Francisco Alves de Matos, Global Compliance Director, Legora, distinguish AI-assisted preparation from responsibility for the decision and identify data governance, auditability, retention, access control and non-training guarantees as conditions for serious institutional deployment [50, slides 3, 8, 14]. These considerations do not determine the legal allocation of judicial authority, but they illustrate how that allocation must be supported by technical and contractual design.

Drăgușin frames digital judicial independence around infrastructure, the reuse of judicial data and AI, directing attention to the institutional powers necessary to preserve such independence [27, slides 2– 8]. A court must be able to specify a system's purposes, inspect material changes, govern access to its records and respond to an incident even if another institution owns the servers. A judicial data policy should govern publication, anonymisation, reuse and protection against re-identification or profiling.

The same safeguards matter when court records form part of a model's data, and when its outputs influence a judgment.

UNESCO's published Guidelines for the Use of AI in Courts and Tribunals set out fifteen principles, including human oversight, transparency and accountability [34]. They can structure a court's assessment of purpose, training data, affected rights and avenues for contesting errors. The decisive question for any readiness assessment is whether a court has the legal authority, records, skills and governance to use the proposed tool while retaining responsibility for adjudication.

The ELI–UNESCO project presented by Gianluca Grasso at the Riga Summit addresses the institutional implementation of AI through governance frameworks, ethical impact assessments and readiness methodologies [28, slides 6, 9–10]. These tools may help courts assess whether effective judicial oversight is possible before deploying an AI system.

Professional competence is a condition of genuine freedom to depart from AI output: judges must have the knowledge and access to primary materials necessary to assess a recommendation independently. Training must cover effective uses and situations requiring a judge to decline a tool.

The professional standard turns on the task, the approved environment and the judge's actual capacity to check a result.

The CEPEJ Model Technology Curriculum for Judges, based on the work of Lorenza Calcagno and Maria Giuliana Civinini, covers basic digital knowledge, security, advanced AI and judicial leadership. It also envisages specialist trainers [35, executive summary, n. 1, paras 41–42]. Civinini's Riga presentation identifies the “trainer's paradox”: judicial educators are expected to teach responsible AI use without necessarily having sustained experience with deployed judicial AI systems [25, slides 16–17]. A practical training exercise would require a judge to compare a generated summary with the complete record, identify a missing adverse authority and decide a case after a model recommendation changes. The assessment should concern the legal reasons given, not

agreement with the model. UNESCO's AI4JUST programme, presented at the Summit by Kamel El Hilali, separates training for court leaders, court staff building AI solutions, and judges, prosecutors and trainers [26, slides 20– 28]. A court president’s choice of performance indicators can nullify individual training if judges are rewarded for following a model. The institution should combine legally authorised tools, access to primary sources, task-specific instruction and a procedure for correcting errors or suspending unsafe use [33, pp. 3–4; 35, paras 8–9].

7. Applying the criterion and protecting the parties

The test applies differently to three kinds of assistance. Technical or informational support includes transcription, translation with verification, retrieval, formatting, anonymisation and citation checks.

These operations can generally be separated from the judicial conclusion, though errors can still prejudice a party. Analytical assistance includes summaries of the record, identification of potentially comparable cases and alternative arguments. It requires access to primary material, sufficient time for evaluation and, where content affects adjudication, a fair opportunity for the parties to address it.

Proposed dispositions or decisive reasons present the greatest risk because their influence falls on the judgment itself [5; 11, paras 55–63].

Practical guidance on AI-assisted judicial drafting reflects a similar task-based distinction. Malý treats the judge’s responsibility as non-delegable and recommends limited uses such as editing, rephrasing, structural assistance and language refinement rather than delegating original legal reasoning to the system [51]. That practical allocation is compatible with, but does not itself answer, the legal question addressed here: whether the judge retained substantive authorship of the outcome and its decisive reasons.

In Ukraine, the legal inquiry has two consecutive stages. First, Article 16 of the Code of Judicial Ethics determines whether the judge may use AI for the proposed task: its use is permissible only where it does not affect judicial independence or impartiality, does not concern the assessment of evidence or the decision-making process, and does not violate the law [13, art 16]. Compliance with the substantive-authorship test cannot make permissible a use that Article 16 excludes. If an ostensibly technical tool in fact concerns the assessment of evidence or the decision-making process, its use must first be reconsidered under Article 16. Second, for a use permitted under Article 16, the

substantive-authorship test assesses whether the actual workflow preserved the judicial function. The test also identifies the limit that any future procedural framework would need to preserve: when a recommendation influences the outcome, the judge's ability to verify, decline and explain must remain effective.

The same technology can cross categories through design. A tool that returns case citations may merely support research; one that orders them as an unexplained probability of success and makes that probability the default disposition begins to influence the normative choice. A case-management system may improve allocation; if its categorisation effectively determines whether a person is heard or which issues the court considers, the implications for access and adversarial procedure are much greater. Independent performance evaluation must therefore examine the complete workflow and differential effects, not only the isolated accuracy of generated text.

Agentic systems sharpen this problem [60]. A system that plans and executes a sequence of steps (searching, selecting authorities, summarising the record, drafting and revising) may present the judge only with the end product of a chain of intermediate choices concerning which sources were consulted, which arguments were omitted and which issues were framed as decisive. Classification by individual task then loses part of its protective value, since each step may appear technical while their combination structures the decision. For the substantive-authorship test, the relevant unit of assessment becomes the delegated workflow as a whole. Where an agent exercises material discretion in selecting or ordering the legal issues, evidence or authorities presented to the judge, its contribution

should be treated as analytical rather than merely technical assistance. The workflow should therefore preserve enough traceability for the judge, and where procedurally material for the parties or a reviewing court, to understand the intermediate choices that shaped the result without requiring disclosure of protected judicial deliberations.

Chernohorenko’s proposal for a distinct classification code for war-damage claims illustrates the boundary: automated routing may assist case management, while assessment of the claim and the court’s order remain with the assigned judge [48, p. 11; 13, art 16].

The Court of Justice has addressed safeguards governing the allocation and withdrawal of cases. In Sinalov, it observed that random allocation does not in itself appear likely to expose a court's formations to undue influence, while leaving room to examine possible manipulation in a particular case [40, para 51]. In D.K., it held that withdrawing cases from a judge under rules lacking objective criteria and reasons threatens judicial independence [41, paras 77, 83]. In his Opinion of 19 March 2026 in Rowicz, Advocate General Spielmann examined a computerised case-allocation system against these principles [38, paras 56–70]. He also questioned whether that system, based on predefined rules, falls within the AI Act's definition of an AI system [38, para 33]. Case allocation

and software proposing a judgment perform different functions. The law must protect both the tribunal's lawful composition and the independent judgment of the assigned judge [37–38; 40–41].

There should be a proportionate, contestable record of significant AI assistance. The object of disclosure should be the material external influence on adjudication, not the reconstruction of a judge’s private deliberative process. The parties need to know enough to challenge a materially influential generated assertion, discover an erroneous source and obtain meaningful appellate review.

Routine spelling correction need not become a separate legal issue. A generated analysis that supplies a disputed factual inference or decisive legal ground does. Disclosure rules must be designed with the secrecy of judicial deliberations, protected information and the parties' right to a fair hearing in mind; indiscriminate publication of all judicial queries is not a workable substitute for reasoned disclosure [8; 11, para 61; 16, arts 8–9, 14].

Material AI output may also enter the record through a court-appointed expert. Civinini argues that an expert relying on AI to prepare an evaluative opinion should identify the tool's role and provide information allowing the judge and parties to assess the conclusion [45, pp. 287, 296–297]. The judge must then assess the expert opinion under the applicable rules of evidence; the expert's use of AI cannot determine a disputed fact for the court.

The distinction between human adjudication and automated public administration must also remain clear. Ukrainian administrative law permits certain administrative acts to be produced automatically under statutory conditions. My recent review of Ukrainian administrative case law identifies responsibility for the governing rules, source data and legal use of automated administrative acts [24].

Those administrative powers provide no authority to delegate adjudication. Effective judicial review may require the court to examine the rules, data and grounds from which the administrative result followed.

8. Institutional and practical objections

The strongest accuracy-based objection goes further than a demand for efficient assistance: if an AI system could decide specified cases more accurately, consistently and cheaply than human judges, why should law insist on human adjudication at all [55; 56]? That objection exposes the premise requiring justification here. The claim of this article is narrower than a general philosophical right to human decision-making: under a legal order that assigns adjudicative authority to a lawfully constituted court, superior performance does not by itself reallocate that authority. Performance evidence may justify redesigning judicial work; it does not, without a lawful change in institutional authority, identify a new decision-maker.

A legal system could debate a change to that allocation through constitutional and democratic processes. It cannot accomplish the same result by requiring judges to approve machine-selected outcomes while describing them as independent decisions. The objection properly presses courts to minimise avoidable delay and justify every burden they impose on access to justice.

Judges have always relied on assistants, clerks, expert evidence and prior judgments. The comparison turns on whether the court can examine the underlying material, reject the contribution and make the dispositive choice. AI changes the scale, speed and apparent completeness of assistance, and may embed default decisions in a system that also measures conformity. Pairing a judge with AI is therefore no general solution. Paar argues that a clearer division of labour can protect judgment better than closely interwoven cooperation [17, pp. 1401–1403]. Where an informational task can be segregated, that allocation is persuasive. Where a summary omits an argument, a recommendation becomes the default, or a proposed disposition structures deliberation, the three-part test must still

ask whether the judge formed the operative decision. If independent judicial authorship cannot be sustained for the task, the institution must remove that use or seek a lawful change to the allocation of authority; it cannot rely on formal human approval.

Proof of substantive authorship does not require access to a judge’s private thoughts. It asks whether the court's reasons address the legally material dispute, whether the procedure offered meaningful access to sources and contrary positions, whether the judge could reject the recommendation without penalty, and whether governance records can support review of a material system contribution where necessary. Together, these features help assess whether the judge’s control was substantive rather than nominal. No single indicator is conclusive; the inquiry is functional and cumulative, directed to the actual conditions under which the decision was formed.

Where these features establish that a system, rather than the assigned judge, substantively determined the operative judgment, the procedural consequences depend on the applicable legal order. In Ukrainian administrative proceedings, this raises a question under Article 124 of the Constitution and the rules governing judicial composition and the lawful exercise of adjudicative authority, including Article 353(3)(1) of the Code of Administrative Justice of Ukraine [9; 57]. A failure by a lawfully constituted court to exercise independent judgment does not necessarily constitute a defect in the court's composition. Whether substantive substitution by AI could justify applying the procedural ground concerning unlawful judicial composition remains an open question under Ukrainian law.

This article advances that possibility as a scholarly hypothesis, not as established Ukrainian case law or an automatic ground for setting aside a judgment. The judgment in Ástráðsson illustrates the importance of a tribunal established by law but does not resolve this separate question [37].

9. Conclusion: the last judge

The preceding analysis suggests a practical sequence. Before using a system, a court must establish that its function is lawful, evaluate its performance on relevant tasks, give parties the means to contest material output, ensure that those responsible for verification are adequately trained and govern the tool throughout its use [11; 33–35]. The Riga materials describe both operational initiatives and proposals for training, governance and judicial use [25–30; 50–51]. The constitutional question arises at the point where a system's contribution can determine the result. Performance alone cannot give it judicial office, and a formal signature cannot substitute for a court's independent decision.

Courts can use AI to improve the speed, consistency, accessibility and quality of judicial work. They must also confront delay, cost, inconsistent case law and diminished public confidence. Those defects strengthen the case for reform, but they do not change who holds judicial power.

The gradual transfer of adjudicative authority to AI may occur without any formal change to the judicial institution. A judge may continue to hear the parties, sign judgments and bear legal responsibility while increasingly relying on system-generated conclusions. Default recommendations, institutional incentives and limited opportunities for independent review can make that reliance difficult to resist. The central safeguard is the judge's effective control over the outcome and its decisive reasons.

The parallel with Fukuyama concerns the possible survival of an institution after its essential function has changed. The last judge may still occupy the bench and sign judgments, yet have little substantive influence over decisions issued in the court's name. The legal responsibility would remain with the judge, while the exercise of judicial authority would have moved elsewhere.

Acknowledgements This article grew out of the discussions at the EJTN AI Summit “The Impact of AI on the Judiciary” (DIGI/2026/01), held in Riga on 24–25 September 2026 and organised by the European Judicial Training Network, with the Judicial Academy of the Republic of Latvia as hosting member. The discussions provided the immediate inspiration for the question at the centre of this article: under what conditions does a judge retain substantive authority over a decision prepared with AI assistance?

The author is grateful to the organisers, speakers and participants for the presentations, workshop discussions and professional exchanges that helped sharpen the questions examined in this article.

EJTN confirmed that the Summit presentations may be cited in academic work with attribution to the speakers and EJTN. The views and conclusions expressed remain solely those of the author.

Declarations Authorial position and potential interests. The author is a judge of the Administrative Cassation Court of the Supreme Court (Ukraine) and serves as a scientific expert of the CEPEJ Working Group on Cyberjustice and Artificial Intelligence. This article discusses CEPEJ instruments. The analysis expresses the author's scholarly views and does not represent a position of the Supreme Court or the Council of Europe. The interpretive argument concerning Article 353 of the Code of Administrative Justice of Ukraine is advanced solely as a scholarly hypothesis and does not express any predetermined position concerning pending or future judicial proceedings.

The author served as reporting judge in the two Ukrainian cases discussed in section 2 [53–54]. The author participated in the working group that prepared the Commentary to the Code of Judicial Ethics cited in this article.

The analysis relies exclusively on published judicial decisions and other publicly available materials and does not draw on judicial deliberations or non-public case information.

Use of generative AI Generative AI tools, including OpenAI's ChatGPT and Anthropic's Claude, were used under the author's direction for linguistic and structural editing, consistency checks, identification of possible repetition and internal inconsistencies, and alternative formulations of selected passages. These uses were assistive in nature, consistent with the distinction reflected in the GAIDeT taxonomy [52]. All AI-assisted material was reviewed by the author. The selection and verification of sources, legal

analysis, the proposed test of substantive judicial authorship and conclusions remain the author's responsibility. References All URLs were last accessed on 14 October 2026.

1. FUKUYAMA, Francis. The End of History and the Last Man. New York : Free Press, 1992.

ISBN 0-02-910975-2. URL: https://search.worldcat.org/title/The-end-of-history-and-the-last- man/oclc/24378505

2. EUROPEAN JUDICIAL TRAINING NETWORK. AI Summit: The Impact of AI on the

Judiciary. Agenda. DIGI/2026/01. Riga, 24–25 September 2026. 9 p. URL: https://padlet.com/tieslietuakademija/ai-summit-riga-m7xekv7ncgdlasit

3. RAZUMKOV CENTRE. Assessment of the Situation in the Country, Trust in Social Institutions

and Assessment of Problems in Ukrainian Society (November 2025) [in Ukrainian]. Kyiv, 2025. URL: https://razumkov.org.ua/napriamky/sotsiologichni-doslidzhennia/otsinka-sytuatsii-v-kraini- dovira-do-sotsialnykh-instytutiv-otsinka-problem-shcho-isnuiut-v-ukrainskomu-suspilstvi-lystopad- 2025r

4. BERNAZIUK, Ian. Artificial Intelligence and Its Use to Ensure the Uniformity of Case Law as a

Component of Public Trust in the Court [in Ukrainian]. Word of the National School of Judges of Ukraine. 2024. No 4 (49). P. 16–35. DOI: 10.37566/2707-6849-2024-4(49)-2. URL: https://slovo.nsj.gov.ua/images/pdf/2024_4_49/bernaziuk.pdf

5. BARYSĖ, Dovilė; SAREL, Roee. Algorithms in the Court: Does It Matter Which Part of the

Judicial Decision-Making Is Automated? Artificial Intelligence and Law. 2024. Vol. 32. P. 117–146. DOI: 10.1007/s10506-022-09343-6. URL: https://doi.org/10.1007/s10506-022-09343-6

6. DYEVRE, Arthur; SHAHVAROUGHI, Ahmad. Man and Machine: Artificial Intelligence and

Judicial Decision Making. Preprint. 2026. URL: https://arxiv.org/pdf/2603.19042

7. LIU, John Zhuang; LI, Xueyao. How Do Judges Use Large Language Models? Evidence from

Shenzhen. Journal of Legal Analysis. 2024. Vol. 16. No 1. P. 235–262. DOI: 10.1093/jla/laae009. URL: https://hub.hku.hk/handle/10722/353494

8. COUNCIL OF EUROPE. Convention for the Protection of Human Rights and Fundamental

Freedoms. Rome, 4 November 1950. Art. 6. URL: https://www.echr.coe.int/documents/d/echr/convention_ENG

9. UKRAINE. Constitution of Ukraine. Adopted 28 June 1996. Arts 124, 126, 129 [in Ukrainian].

URL: https://zakon.rada.gov.ua/laws/show/254%D0%BA/96-%D0%B2%D1%80#Text

10. CONSULTATIVE COUNCIL OF EUROPEAN JUDGES. Opinion No 26 (2023): Moving

Forward: The Use of Assistive Technology in the Judiciary. Strasbourg, 2023. URL: https://rm.coe.int/ccje-opinion-no-26-2023-final/1680adade7

11. EUROPEAN COMMISSION FOR THE EFFICIENCY OF JUSTICE. Guidelines on the Use of

Generative Artificial Intelligence for Courts. CEPEJ(2025)18Final. Strasbourg, 19 December 2025. URL: https://rm.coe.int/cepej-2025-18final-en-draft-guidelines-on-the-use-of-generative-ai- for/48802a4ad1

12. EUROPEAN UNION. Regulation (EU) 2024/1689 of the European Parliament and of the

Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence. Official Journal of the European Union. 2024. Art. 14; recital 61; Annex III, point 8(a). URL: https://eur- lex.europa.eu/eli/reg/2024/1689/oj/eng

13. CONGRESS OF JUDGES OF UKRAINE. Code of Judicial Ethics. Approved 18 September

2024. Art. 16 [in Ukrainian]. URL: https://rsu.gov.ua/uploads/article/rishennjzizdprokodeksetykypdf-20e12e8e01.pdf

14. SUPREME COURT (UKRAINE). Regulation on the Use of Artificial Intelligence Technologies

by Employees of the Office of the Supreme Court. Administrative Order No 117 of 8 December 2025. Official English translation. URL: https://court.gov.ua/storage/portal/supreme/rizne/ADMINISTRATIVE_ORDER_HEAD_OFFICE_ SC.pdf

15. STATE JUDICIAL ADMINISTRATION OF UKRAINE. Concept of the Unified Judicial

Information and Communication System. Order No 178 of 30 April 2025. Section 6.3 [in Ukrainian]. URL: https://court.gov.ua/storage/portal/dsa/normatyvno- pravova%20baza/N_178_2025_dodatok.pdf

16. COUNCIL OF EUROPE. Framework Convention on Artificial Intelligence and Human Rights,

Democracy and the Rule of Law. CETS No 225. Vilnius, 5 September 2024. URL: https://rm.coe.int/1680afae3c

17. PAAR, Elisabeth. Artificial Intelligence in Court Proceedings: Judge’s Little Helper or the

Beginning of AI’s Hostile Takeover? German Law Journal. 2025. Vol. 26. Special issue 7. P. 1383– 1403. DOI: 10.1017/glj.2026.10181. URL: https://doi.org/10.1017/glj.2026.10181

18. GREEN, Ben. The Flaws of Policies Requiring Human Oversight of Government Algorithms.

Computer Law & Security Review. 2022. Vol. 45. Article 105681. DOI: 10.1016/j.clsr.2022.105681. URL: https://doi.org/10.1016/j.clsr.2022.105681

19. POPOTAS, Costas. AI and Judicial Independence: A Critical Analysis. Afterword and guest

editorial, published without peer review. Access to Justice in Eastern Europe. 2026. Vol. 9. No 2. P. 529–552. DOI: 10.33327/AJEE-18-9.2-o000195. URL: https://ajee-journal.com/ai-and-judicial- independence-a-critical-analysis

20. BERNAZIUK, Ian. Artificial Intelligence in Scholarly Research: Personal Scientific

Contribution, Academic Integrity and Responsibility [in Ukrainian]. Zenodo. 2026. DOI: 10.5281/zenodo.22257136. URL: https://constitutionalist.com.ua/wp-content/uploads/2026/09/129.- Bernaziuk_AI_Scholarly_Research_Academic_Integrity_2026.pdf

21. ALON-BARKAT, Saar; BUSUIOC, Madalina. Human–AI Interactions in Public Sector

Decision Making: “Automation Bias” and “Selective Adherence” to Algorithmic Advice. Journal of Public Administration Research and Theory. 2023. Vol. 33. No 1. P. 153–169. URL: https://academic.oup.com/jpart/article/33/1/153/6524536

22. BERNAZIUK, Ian. Sovereign AI: From a Technological Idea to a Matter of State Resilience.

Constitutionalist. 2026. URL: https://constitutionalist.com.ua/sovereign-ai-from-a-technological- idea-to-a-matter-of-state-resilience/

23. BANKS, Isabella. Judges-in-the-Loop? Judicial Involvement in Human Oversight of High-Risk

Decision Support Systems under the EU AI Act. International Journal of Law and Information Technology. 2026. DOI: 10.1093/ijlit/eaag001. URL: https://academic.oup.com/ijlit/article/doi/10.1093/ijlit/eaag001/8475419

24. BERNAZIUK, Ian. Artificial Intelligence and Automated Systems in Public Administration and

Adjudication: Human Oversight, Reason-Giving, Verifiability and Responsibility in the Case Law of the Cassation Administrative Court within the Supreme Court. English Edition, Version 1.0.

Zenodo. 2026. DOI: 10.5281/zenodo.23205043. URL: https://doi.org/10.5281/zenodo.23205043

25. CIVININI, Maria Giuliana. The Impact of AI on the Judiciary: Partners’ Panel – AI Training

Resources for Justice Practitioners. Presentation, EJTN AI Summit, Riga, 24 September 2026. 28 slides. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

26. EL HILALI, Kamel. AI & the Rule of Law Programme: An Overview. Presentation, EJTN AI

Summit, Riga, 24 September 2026. 30 slides. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

27. DRĂGUȘIN, Claudiu (ENCJ; Superior Council of Magistracy, Romania). ENCJ Digital Justice

Forum 2024–2026: Digital Transformation and Judicial Independence. Presentation, Paris, June 2026; distributed at the EJTN AI Summit, Riga, 24 September 2026. 12 slides. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

28. GRASSO, Gianluca. The Use of AI in Judicial Decision Making: The ELI–UNESCO Project.

Presentation, EJTN AI Summit, Riga, 24 September 2026. 14 slides. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

29. LARSSEN, Wiggo Storhaug. The Use of Artificial Intelligence in the Judicial Process.

Presentation, EJTN AI Summit, Riga, 24 September 2026. 9 slides. Discusses draft CCJE Opinion No 29 awaiting plenary adoption. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

30. ĆAPETA, Tamara. AI at the CJEU. Presentation, EJTN AI Summit, Riga, 24 September 2026.

10 slides. Organiser's link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

31. COURT OF JUSTICE OF THE EUROPEAN UNION. Judgment of 7 December 2023, OQ v

Land Hessen, Case C-634/21, SCHUFA Holding (Scoring), ECLI:EU:C:2023:957. URL: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:62021CJ0634

32. UNESCO. UNESCO Global Judges’ Initiative: Survey on the Use of AI Systems by Judicial

Operators. Paris, 2024. URL: https://unesdoc.unesco.org/ark:/48223/pf0000389786

33. COURT OF JUSTICE OF THE EUROPEAN UNION. Artificial Intelligence Ethics Charter.

Adopted 26 January 2026. Luxembourg, 2026. 4 p. URL: https://curia.europa.eu/site/upload/docs/application/pdf/2026-03/ethics_charter_ai_-_en.pdf

34. GUTIÉRREZ, Juan David; UNESCO. Guidelines for the Use of AI Systems in Courts and

Tribunals. Paris : UNESCO, 2025. 44 p. DOI: 10.58338/LIEY8089. URL: https://unesdoc.unesco.org/ark:/48223/pf0000396582_eng

35. EUROPEAN COMMISSION FOR THE EFFICIENCY OF JUSTICE. Model Technology

Curriculum for Judges. CEPEJ(2026)8. Adopted 18–19 June 2026. Strasbourg : Council of Europe, 2026. URL: https://rm.coe.int/cepej-2025-8-en-model-technology-curriculum-for- judges/48802cbec4

36. FUKUYAMA, Francis. The End of History? The National Interest. 1989. No 16 (Summer). P.

3–18. URL: https://www.jstor.org/stable/24027184

37. EUROPEAN COURT OF HUMAN RIGHTS. Guðmundur Andri Ástráðsson v Iceland. Grand

Chamber judgment of 1 December 2020, application no 26374/18. URL: https://hudoc.echr.coe.int/eng?i=001-206582

38. SPIELMANN, Dean. Opinion of the Advocate General delivered on 19 March 2026, Case C-

159/25, Rowicz, ECLI:EU:C:2026:229. URL: https://juris.curia.europa.eu/juris/document/document.jsf?docid=310088&doclang=EN

39. EUROPEAN PARLIAMENT AND COUNCIL OF THE EUROPEAN UNION. Regulation

(EU) 2026/1744 of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). Official Journal of the European Union, OJ L, 2026/1744, 24 July. URL: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng

40. COURT OF JUSTICE OF THE EUROPEAN UNION. Judgment of 27 February 2025, Case C-

16/24, Sinalov, ECLI:EU:C:2025:116. URL: https://eur-lex.europa.eu/legal- content/EN/TXT/?uri=CELEX:62024CJ0016

41. COURT OF JUSTICE OF THE EUROPEAN UNION. Judgment of 6 March 2025, Joined

Cases C-647/21 and C-648/21, D.K. (Withdrawal of Cases from a Judge), ECLI:EU:C:2025:143. URL: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62021CJ0647

42. CONSTITUTIONAL COURT OF COLOMBIA. Judgment T-323/24 of 2 August 2024, Case T-

9.301.656 [in Spanish]. URL: https://www.corteconstitucional.gov.co/relatoria/2024/t-323-24.htm

43. SOLOVEY, Erin; FLANAGAN, Brian; CHEN, Daniel. Interacting with AI at Work: Perceptions

and Opportunities from the UK Judiciary. Proceedings of the 4th Annual Symposium on Human- Computer Interaction for Work (CHIWORK ’25). New York : ACM, 2025. 8 p. DOI: 10.1145/3729176.3729192. URL: https://users.wpi.edu/~esolovey/papers/solovey-chiwork25.pdf

44. CIVININI, Maria Giuliana. Introduzione. La giustizia di fronte alla sfida tecnologica.

Reimmaginare la giustizia [in Italian]. Questione Giustizia. 2026. No 3. P. 9–14. URL: https://www.questionegiustizia.it/data/rivista/articoli/1384/3_2026_qg_introduzione-civinini.pdf

45. CIVININI, Maria Giuliana. L’utilizzo dell’IA da parte dell’esperto giudiziario (con particolare

riferimento al consulente/perito medico) [in Italian]. Questione Giustizia. 2026. No 3. P. 287–301. URL: https://www.questionegiustizia.it/rivista/articolo/l-utilizzo-dell-ia-da-parte-dell-esperto- giudiziario-con-particolare-riferimento-al-consulente-perito-medico

46. EUROPEAN COMMISSION FOR THE EFFICIENCY OF JUSTICE. Assessment Tool for the

Operationalisation of the European Ethical Charter on the Use of Artificial Intelligence in Judicial Systems and Their Environment. CEPEJ(2023)16rev. Adopted 4–5 December 2023; revised 4–5 June 2025. Strasbourg : Council of Europe, 2025. 17 p. URL: https://rm.coe.int/cepej-2023-16rev- operationalisation-ai-ethical-charter-en/1680b65218

47. SHAH, Anand V.; LEVY, Joshua Y. Access to Justice in the Age of AI: Evidence from U.S.

Federal Courts. Working paper. March 2026. URL: https://avshah1.github.io/assets/pdf/papers/pro- se/Pro_Se_Automation.pdf

48. CHERNOHORENKO, Illia. AI in the Courts, by the Council of Europe Standard. Presentation,

19 August 2026. PDF, 12 p. URL: https://zenodo.org/records/22774412

49. EUROPEAN COURT OF HUMAN RIGHTS. Moreira Ferreira v Portugal (No 2) [GC].

Judgment of 11 July 2017, application no 19867/12. Para 84. URL: https://hudoc.echr.coe.int/eng?i=001-175646

50. ALVES DE MATOS, Francisco. What the Private Sector Already Learned – and What It Means

for the Judiciary. Responsible AI for Justice. Presentation materials. Global Compliance Director, Legora. EJTN AI Summit, Session IX: Stakeholders’ Panel (Part II), Riga, 25 September 2026. 18 slides. Organiser’s link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

51. MALÝ, Miloš. Using Artificial Intelligence (AI) as an Assistant in Drafting Judicial Decisions.

Training material dated 17 April 2026, adapted for a European context from a Czech judiciary methodology; distributed at the EJTN AI Summit, Riga, 24–25 September 2026. 11 p. Organiser’s link to expert presentations: https://padlet.com/tieslietuakademija/AI_SUMMIT_RIGA/wish/4055163006

52. SUCHIKOVA, Yana; TSYBULIAK, Nataliia; TEIXEIRA DA SILVA, Jaime A.;

NAZAROVETS, Serhii. GAIDeT (Generative AI Delegation Taxonomy): A Taxonomy for Humans to Delegate Tasks to Generative Artificial Intelligence in Scientific Research and Publishing.

Accountability in Research. 2026. Vol. 33. No 3. Article 2544331. DOI: 10.1080/08989621.2025.2544331. URL: https://doi.org/10.1080/08989621.2025.2544331

53. SUPREME COURT (UKRAINE). Administrative Cassation Court. Ruling of 15 January 2026,

case No 240/14153/24 [in Ukrainian]. URL: https://reyestr.court.gov.ua/Review/133336040

54. SUPREME COURT (UKRAINE). Administrative Cassation Court. Ruling of 3 February 2026,

case No 340/2365/25 [in Ukrainian]. URL: https://reyestr.court.gov.ua/Review/133764626

55. VOLOKH, Eugene. Chief Justice Robots. Duke Law Journal. 2019. Vol. 68. No 6. P. 1135–

1192. URL: https://scholarship.law.duke.edu/dlj/vol68/iss6/2

56. HUQ, Aziz Z. A Right to a Human Decision. Virginia Law Review. 2020. Vol. 106. No 3. P.

611–688. URL: https://virginialawreview.org/articles/right-human-decision/

57. UKRAINE. Code of Administrative Justice of Ukraine. No 2747-IV of 6 July 2005. Art. 353 [in

Ukrainian]. URL: https://zakon.rada.gov.ua/laws/show/2747-15#Text

58. CASAL, Simón (dir.). Justicia artificial [Artificial Justice]. Spain–Portugal, 2024. Feature film.

Catalogue record: Institute of Cinematography and Audiovisual Arts (ICAA), Ministry of Culture of Spain. URL: https://infoicaa.mcu.es/CatalogoICAA/es-es/Peliculas/Detalle?Pelicula=40722

59. UNITED STATES. Executive Order 14434 of 29 September 2026, Inaugurating the Era of

Super Intelligence. Federal Register. 2 October 2026. Vol. 91. No 190. P. 63129. URL: https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super- intelligence/

60. NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY. Agentic AI. U.S.

Department of Commerce, 2026. URL: https://www.nist.gov/agentic-ai

61. TINNA, Nur. Judicial Cognitive Legitimacy and the Constitutional Limits of AI-Assisted

Adjudication: Beyond Explainability, Human Oversight, and Algorithmic Authority. SSRN, 2026. DOI: 10.2139/ssrn.7115478. URL: https://doi.org/10.2139/ssrn.7115478

62. COUNCIL OF JUDGES OF UKRAINE. Commentary to the Code of Judicial Ethics [in

Ukrainian]. Approved by Decision No 14 of 2 March 2026. Kyiv, 2026. 268 p. Commentary on Article 16, pp. 184–193. URL: https://rsu.gov.ua/uploads/article/komentar2603061320pbkce- af35f4321b.pdf

63. BERNAZIUK, Ian. Artificial Intelligence in Constitutional Adjudication: Implementation

Practices, Limits of Use and Standards of Judicial Oversight. An Author's Analytical Review as of 2 October 2026. English Edition, Version 1.0. Zenodo, 2026. DOI: 10.5281/zenodo.23103501. URL: https://doi.org/10.5281/zenodo.23103501