Administrative Discretion in the Context of Automation and Artificial Intelligence

Ian Bernaziuk Doctor of Law, Professor; Judge of the Administrative Cassation Court within the Supreme Court (Ukraine); Scientific Expert, CEPEJ-GT-CYBERJUST (Council of Europe)

ORCID: 0000-0002-2353-4836

ADMINISTRATIVE DISCRETION IN THE CONTEXT OF AUTOMATION AND ARTIFICIAL

INTELLIGENCE

A Legal and Analytical Study Abstract This article examines the impact of artificial intelligence and automated systems on the exercise of discretionary powers by public administrative authorities. It distinguishes bound powers, administrative discretion and judicial discretion, and proposes three dimensions for analysing administrative choice: normative, case- specific and technologically mediated. Drawing on Ukrainian legislation, thirteen judicial decisions of the Supreme Court, and the case law of the European Court of Human Rights and the Court of Justice of the European Union, the study identifies the conditions under which automation may constrain administrative arbitrariness and the risks of unlawfully narrowing discretion through software architecture. It proposes

a seven-part test for judicial review of algorithmically assisted administrative decisions and an amendment to Article 62 of the Law of Ukraine “On Administrative Procedure” concerning the limits of fully automated adoption of administrative acts. The anti- corruption and economic effects of automation are treated as propositions requiring empirical verification.

Keywords: administrative discretion; judicial discretion; artificial intelligence; public administration; automated decision-making; algorithmic recommendation; human oversight; corruption risks; duty to give reasons; judicial review; legal certainty.

Introduction Administrative discretion is necessary where the law does not prescribe a single outcome. An authority must consider the facts, compare legally permissible measures, comply with the principle of proportionality and explain why it has chosen a particular course of action. At the same time, discretion creates scope for inconsistent administrative practice: similar applications may receive different responses because of differences in officials’ approaches, internal instructions or improper influence. A difference in outcomes does not, however, establish arbitrariness where the relevant factual circumstances differ.

Artificial intelligence (AI), including machine-learning models, may generate probabilistic assessments from data, whereas conventional automated systems generally apply predetermined rules. Both can affect the exercise of discretion by shaping the information available to officials, the decision-making options presented to them and the criteria by which those options are assessed. The legally decisive consideration is the technology’s actual effect on the scope of administrative choice, individual rights and the ability to verify the lawfulness of a decision [15; 16].

This article does not equate automated systems that implement predetermined rules with AI systems. Cases concerning conventional automation are examined to identify common legal issues of competence, responsibility and verifiability, not as evidence that AI was used in the administrative procedures at issue.

The article seeks to identify the conditions under which AI may help prevent administrative arbitrariness without allowing the rules embedded in AI systems and their errors to become a new source of influence over public decision-making. The research combines the analysis of legislation, judicial practice and comparative law.

Potential economic and anti-corruption benefits are treated as hypotheses to be measured, rather than as established outcomes of digitalisation in Ukraine [16].

The article builds on the author’s earlier research into automated systems in public administration and adjudication [27], the use of AI to promote consistency in judicial practice [28], judicial and administrative discretion [29], as well as training materials prepared for the National School of Judges of Ukraine between 2024 and 2026 [30; 31; 42]. Unlike the earlier review of case law [27], the present study focuses on the effect of software architecture on the practical scope of administrative discretion. Its approach draws on scholarship concerning digital discretion and the shift from discretion exercised by individual officials towards influence exercised by developers and administrators of information systems, commonly described as system-level

bureaucracy [14; 45; 46]. The original contribution lies in distinguishing the legal limits of discretion from its practical availability within technological systems and applying that distinction to judicial review and legislative reform in Ukraine. Thirteen publicly available judicial act s of the Supreme Court in which the author served as judge - rapporteur are examined: twelve judgments and one procedural ruling [4 – 13; 39 – 41].

This is a purposive study of judicial practice involving the author, not a representative survey of all approaches adopted by the Supreme Court. Cases directly involving automation are distinguished from decisions addre ssing general standards of discretion, evidence, reason giving and judicial review. Some cases concern legal relationships arising before the Law of Ukraine “On Administrative Procedure ” entered into force or areas excluded from its scope. The reasoning in those cases is used to examine general standards of legality, reason giving and judicial review, without extending the Law to the relationships concerned [2, Article 1(2)]. The legal

holdings belong to the respective judicial panels, not to the judge-rapporteur in a personal capacity.

1. Discretion: Lawful Choice and the Influence of Software Systems

Under Article 19(2) of the Constitution of Ukraine, state authorities, local self- government bodies and their officials must act only on the basis of, within the powers conferred by, and in the manner prescribed by the Constitution and laws of Ukraine [1, Article 19(2)]. Article 2(1)(7) of the Law of Ukraine “On Administrative Procedure” defines a discretionary power by reference to the authority’s legally conferred ability to choose a decision in accordance with the law and the purpose of the power. Article 6(3) specifies the conditions for the lawfulness of that choice: a legal basis; compliance with the limits, prescribed manner and purpose of exercising the power; observance of the principles of administrative procedure; and the absence of any unjustified

departure from the same authority’s previous decisions in identical or similar cases. Article 6(4) makes the conclusions on the application of legal provisions set out in judgments of the Supreme Court binding on administrative authorities when applying the relevant regulatory instrument; Article 8(4) requires an authority to give reasons for a change in its assessment and conclusions in comparable cases [2, Article 2(1)(7), Article 6(3)–(4), Article 8(4)].

Recommendation No. R (80) 2 of the Committee of Ministers of the Council of Europe, Concerning the Exercise of Discretionary Powers by Administrative Authorities, of 11 March 1980, sets out basic requirements for the exercise of administrative discretion: using powers for their intended purpose, objectivity, impartiality, equality, proportionality, deciding within a reasonable time, and consistency of administrative practice (Appendix, Principles 1–6). Principles 7–11 concern the accessibility of administrative guidelines, the duty to give reasons for departures from them, and review of the exercise of discretion, including by a court or another independent body [38]. These safeguards remain relevant regardless of whether software tools are used.

A provision allowing a choice between several lawful options differs from one prescribing a single lawful outcome. Gathering evidence, checking a document or performing a complex calculation does not, by itself, involve discretion. The distinction is particularly important where an algorithm’s technical complexity creates the impression that an authority has acquired an additional freedom of choice.

In its judgment of 15 September 2026 in case No. 420/21964/23, the Supreme Court distinguished the identification of a technical error in a state register from the exercise of administrative discretion. Once such an error has been established from the registration file, the state registrar must correct it, because the law provides no other lawful course of action [12, paras. 106–115]. At the same time, a court cannot perform, in place of the registrar, an initial examination of documents that the

administrative authority itself failed to undertake. The Supreme Court also held that inadequate arrangements for considering an application should not deprive the applicant of effective protection of the right infringed [12, paras. 86–105].

A similar distinction was drawn in the judgment of 14 September 2026 in case No. 560/9501/24: where a statutory ground for refusing planning conditions and restrictions exists, the authority has no discretion to issue a favourable decision [40, paras. 46–47]. Accordingly, automating the verification of statutory requirements must not create scope for their arbitrary disregard.

Three interrelated dimensions of administrative discretion should be distinguished. The normative dimension comprises the purpose, limits and permissible outcomes prescribed by law. The case-specific dimension encompasses fact-finding, the assessment of interests and the selection of a lawful measure. The technologically mediated dimension concerns the influence of software architecture on the availability of information, assessment criteria and the ability to select or reject an option proposed by the system. The third dimension does not create a new public power; it describes the practical conditions under which legally conferred discretion is exercised [16].

If a system permits only two sanctions to be selected when the law provides three lawful alternatives, discretion is restricted in practice without any amendment to the law. The same may occur where an algorithm conceals a material circumstance or substitutes an indicator of institutional convenience for the statutory purpose. Legal scrutiny must therefore encompass not only the final administrative act but also the criteria according to which the system generated the options available to the authority [15; 37].

The practical importance of this distinction is apparent when determining the amount of a sanction. In its judgment of 3 June 2024 in case No. 380/7521/23, the Supreme Court held that the National Energy and Utilities Regulatory Commission (NEURC)

must give reasons for the amount of a fine and, where multiple infringements are involved, explain how that amount was calculated [39, para. 60]. A court reviews the justification for and proportionality of the sanction without itself determining a new amount in place of the regulator [39, paras. 53, 65–67]. Consequently, a software recommendation concerning the amount of a sanction does not relieve the administrative authority of its duty to explain its choice.

There is an opposite risk. Where the law requires a single outcome, intervention by an official under the guise of “human oversight” may itself become arbitrary. The appropriate form of oversight must depend on the nature of the power exercised, rather than on a uniform requirement that every system-generated result receive manual approval.

2. Can Algorithms Help Prevent Arbitrariness and Corruption?

Large volumes of routine administrative cases create opportunities for automated checks on consistency of administrative practice. A system may flag a decision that differs from previous decisions in similar circumstances, identify a relevant Supreme Court judgment and indicate that no reasons have been given for the departure. This function is directly connected to Article 6(3) and (4) and Article 8(4) of the Law “On Administrative Procedure” [2]. The automated detection of inconsistencies was considered in the author’s earlier study of consistency in judicial practice [28]; in public administration it also follows from the statutory duty to apply the law consistently.

Such checks must not freeze previous practice. New circumstances, legislative amendments or the need to correct an erroneous approach may justify a different decision. Technology is useful where it requires the difference to be explained, without preventing it.

The operation of electronic systems in the construction sector provides an illustration. In case No. 420/11131/22, the Supreme Court examined automated data checks established by law and the conditions for recording information in the Construction Activities Register [4, paras. 47–60]. In case No. 460/22384/23, the Court distinguished the system’s automated checks from the registrar’s limited authority to interfere with the results produced under those checks [5, para. 68]. Neither case establishes a general preference for algorithmic outputs over human decisions. The point is that neither an official nor a software tool may depart from the legally prescribed procedure for exercising a public function.

These cases support a testable hypothesis concerning the anti-corruption effects of automation: in procedures governed by unambiguous legal requirements, automation may reduce opportunities to demand unnecessary documents, delay consideration or invoke exceptions not provided for by law. Corruption risks may, however, shift to those who draft technical specifications, determine assessment rules or select datasets. Improper influence during system procurement or upgrades may sometimes be harder to detect than influence exerted in an individual administrative proceeding.

The effects of automation should be assessed procedure by procedure by comparing the frequency of unjustified refusals, the number of complaints and revised decisions, processing times, contacts between applicants and officials, and total budgetary costs.

Such comparisons require baseline indicators collected before a system is introduced and must account for other institutional changes. Reduced processing times or staff numbers alone do not establish a reduction in corruption risks [16].

Reducing routine work may free staff to handle complex cases, but automation generates other costs, including data maintenance, information security, complaint

handling and error correction. The efficient use of public funds means reducing the full cost of delivering a legally compliant administrative service, rather than mechanically cutting posts. In view of these costs, automation may be economically justified for some procedures but disproportionate for others [16].

Finally, automation does not guarantee predictability. Changes to models, risk thresholds or training data may cause similar cases to produce different results without any corresponding change in the law. Legal certainty requires system-version records, change control and an ability to explain differences between individual outcomes [15; 37].

3. Automated Procedures, Responsibility and Judicial Practice

Article 62(1) of the Law “On Administrative Procedure” provides that, in the cases and according to the procedure prescribed by law, an administrative authority conducts administrative proceedings wholly or partly in an automated manner, using software without human intervention. Article 62(2) requires the authority to take steps to enable such handling of cases, while Article 62(3) makes it responsible for administrative acts adopted in an automated manner [2, Article 62]. A software system does not thereby become an independent holder of public authority.

Article 62 does not distinguish between bound and discretionary powers. The legal basis for automation requires separate examination: Article 62(1) permits automated proceedings only in cases and according to a procedure prescribed by law.

Subordinate legislation may specify technical rules for the operation of an information system, but may not independently alter the statutory conditions governing the exercise of a power or the limits of administrative choice [2, Article 62(1); 4, paras.

45–55; 7, paras. 51–62]. By comparison, § 35a of the German Verwaltungsverfahrensgesetz (VwVfG) permits a fully automated administrative act only where there is a legal basis and neither discretion (Ermessen) nor a margin of administrative assessment (Beurteilungsspielraum) is involved [19]. The German rule should not be transplanted mechanically into another legal system, but it illustrates the fundamental difference between automating a result prescribed by law and automating the exercise of administrative choice.

France has adopted a different approach. In Decision No. 2018-765 DC of 12 June 2018, the French Constitutional Council accepted, subject to specified safeguards, the possibility of an individual administrative decision based solely on an algorithm. The applicable regime requires a legal basis, predetermined rules and criteria, notification of the individual that an algorithm is being used, access to its principal characteristics, and the possibility of administrative and judicial challenge. A decision based solely on an algorithm is prohibited, in particular, where the relevant characteristics cannot be disclosed because of legally protected secrecy or where certain categories of sensitive

data are processed. On an administrative appeal, the authority may not rely exclusively on the algorithm; algorithms that autonomously modify the rules applied without control by the responsible authority likewise cannot form the sole basis of a decision [47, paras. 65–72].

Cases concerning the Unified State Electronic Database on Education (USEDE) show how an error in an algorithmic rule can affect individual rights. In its judgment of 26 November 2025 in case No. 160/29658/24, the Supreme Court examined the allocation of functions between the Ministry of Education and Science of Ukraine, as the entity responsible for the database, its technical administrator and an educational institution. The automatic generation of a certificate under the programmed algorithm left the institution with no technical means of altering the system’s assessment of the sequence in which education had been pursued [7, paras. 58–62].

In its judgment of 20 May 2026 in case No. 280/9417/24, the Supreme Court reviewed the lawfulness of a USEDE algorithm that treated repeated enrolment at a particular level as a breach of the required sequence of education without taking account of whether the student had previously obtained the relevant qualification. The Court held that the Ministry of Education and Science of Ukraine, as the entity responsible for USEDE, could not shift responsibility for an unlawful outcome to the algorithm or to an educational institution lacking authority to change its parameters [8, paras. 81– 99]. The judgment concerns a rule-based automated procedure, without establishing that an AI system was used, but directly supports the need to review the lawfulness

of an algorithm producing legally significant results. Another source of error is the underlying data. The judgment of 24 April 2025 in case No. 280/3642/23 concerned the automated calculation of administrative-economic sanctions on the basis of state-register data. The Court examined the statutory rules for generating information, the relevant cut-off date and the obligations of the entities supplying the information [6, paras. 58–76]. It also considered the use of automatically processed data as evidence and the need to assess the reliability, admissibility and sufficiency of such evidence [6, paras. 82–85]. In this context, an employer’s compliance with its obligation to submit and correct the original data in a

timely manner is material [6, paras. 75–77]. Article 53(2) of the Law “On Administrative Procedure” recognises data from national electronic information resources and the results of automated data processing or verification as means of evidence. Their automated origin does not, however, remove the need to assess their reliability. The source of a defect—an algorithmic rule, the underlying data or the legal use made of the result—is material to determining responsibility [2, Article 53(2); 6, paras. 75–85; 27, section 2.5].

The distinction also applies to discretionary decisions. Even an accurate risk assessment does not replace a legally required choice between several permissible

measures. Conversely, a defective algorithm does not become lawful merely because an official has approved its output [2; 16].

4. European Standards: The ECtHR, the CJEU and AI Regulation

The European Court of Human Rights assesses the use of technology against the guarantees of the Convention, regardless of whether the system technically qualifies as AI. In Glukhin v. Russia (application No. 11519/20, judgment of 4 July 2023), facial- recognition technology used to identify and arrest a participant in a peaceful protest was examined under Articles 8 and 10 of the Convention. The Court emphasised the need for detailed rules governing the use of such technologies and effective safeguards against abuse; in the circumstances of the case, the interference was found incompatible with Convention guarantees [20, §§ 82–91]. Although the case did not concern administrative discretion in the delivery of a public service, it demonstrates

that technical effectiveness cannot justify disproportionate interference by public authorities. In Ramos Nunes de Carvalho e Sá v. Portugal (Grand Chamber, applications Nos.

55391/13, 57728/13 and 74041/13, judgment of 6 November 2018), the European Court of Human Rights distinguished the functions of the specialist body deciding a disciplinary case from those of the court reviewing its decision. The scope of judicial review depends, among other things, on the nature of the issues in dispute and the significance of fact-finding; respect for an administrative body’s competence does not relieve a court of the duty to examine the decisive arguments with sufficient thoroughness [21, §§ 180–184, 200–205]. The judgment provides guidance on the judicial discretion involved in determining the scope of review of an algorithmically assisted decision, not a basis for automatically reassessing every technical indicator.

The Court of Justice of the European Union has addressed algorithmic processes more directly. In Ligue des droits humains, Case C-817/19, judgment of 21 June 2022, the Court, interpreting the Passenger Name Record (PNR) Directive, emphasised the need for predetermined, non-discriminatory criteria and individual review of the results.

Under that specific legal regime, the use of self-learning systems that alter the assessment process and the weighting of criteria without human intervention is not permitted [22, paras. 194–200]. This holding is not a general prohibition on machine learning in the public sector.

In SCHUFA Holding (Scoring), Case C-634/21, judgment of 7 December 2023, the CJEU held that the automated calculation of a credit score may amount to automated individual decision-making within the meaning of Article 22 GDPR where a third party relies heavily on that score in deciding whether to establish, implement or terminate a contractual relationship [23, para. 73]. What matters legally, therefore, is the

assessment’s actual influence on the final decision, not merely its description as a recommendation. In Dun & Bradstreet Austria, Case C-203/22, judgment of 27 February 2025, the CJEU clarified an individual’s right to receive meaningful information about the logic involved in automated processing, enabling the person to check the accuracy of the assessment and effectively protect their rights [24, paras. 55–66]. Both judgments primarily concern personal-data protection and cannot simply be treated as provisions of Ukrainian administrative procedure. They do, however, illuminate why nominal human involvement and a general reference to an automated calculation may be insufficient.

In Technische Universität München v Hauptzollamt München-Mitte (Case C-269/90, judgment of 21 November 1991), the CJEU linked broad administrative powers of appraisal to the particular importance of procedural safeguards: careful and impartial examination of the relevant circumstances, the right to be heard and adequate reasons for the decision [43, para. 14]. This approach has also been applied in the case law of the Administrative Cassation Court within the Supreme Court [10, para. 70].

The importance of human review of algorithmic outputs in CJEU case law is examined in Julien Sterck’s article, Artificial Intelligence before the Court of Justice of the European Union: Asserting the Importance of a Human-Centric Dimension [36].

Regulation (EU) 2024/1689 (the Artificial Intelligence Act) lays down requirements for human oversight of high-risk AI systems (Article 14), fundamental-rights impact assessments in specified circumstances (Article 27), and explanations of certain decisions (Article 86). As amended by Regulation (EU) 2026/1744, its principal requirements for high-risk systems covered by Annex III apply from 2 December 2027, while those for the relevant product-related systems covered by Annex I apply from 2 August 2028 [25, Article 113]. These dates must not be confused with the entry into force of the Regulation as a whole or with provisions that are already applicable.

The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225) establishes a framework for risk management, transparency and access to remedies [26]. Recommendation CM/Rec(2020)1 [15] and the Handbook on Human Rights and Artificial Intelligence prepared by the Council of Europe Steering Committee for Human Rights in April 2026 [18] are also useful comparative and methodological reference points for Ukraine.

They do not replace applicable national law, and due regard must be paid to the legal status of each instrument.

5. Meaningful Human Oversight and Its Limits

An official does not exercise discretion independently if unable to see the relevant data, verify the grounds for a recommendation or choose a lawful option different

from the one proposed by the system. Mere approval of a machine-generated recommendation does not establish meaningful oversight. Yet the proportion of recommendations approved does not by itself demonstrate abuse: the system may be functioning correctly. Assessment requires information about the complexity of cases, the legal conditions governing departures from recommendations, examples of errors corrected and actual practices in giving reasons for decisions [16; 35].

The risks extend beyond inaccurate predictions. In The Mechanisms of AI Harm: Lessons Learned from AI Incidents, Mia Hoffmann considers different mechanisms through which AI systems may cause harm and ways of preventing it [17]. A practical conclusion follows: a model’s accuracy does not resolve the problem of an authority using it for a purpose other than that prescribed by law, or of failures accumulating because of defective system governance.

The verifiability of automated procedures requires a clear allocation of responsibility for criteria, data and system operation, records of changes, and procedures for correcting errors. The March 2025 guidance, Advice on the Responsible Use of Artificial Intelligence by Public Servants, issued by Ukraine’s Ministry of Digital Transformation, National Agency of Ukraine on Civil Service and Higher School of Public Governance, addresses practical issues concerning responsible AI use, data protection and risk management [44]. Comparative guidance published by the Australasian Institute of Judicial Administration on AI in judicial work [35] may be used with due regard to the differences between judicial and administrative functions.

The use of generative AI in preparing legal texts calls for particular scrutiny. In its procedural ruling of 15 January 2026 in case No. 240/14153/24, the Supreme Court found that a cassation appeal cited judicial decisions that did not exist in the form stated. The Court noted that such errors may arise through the use of AI and emphasised the need for professional human verification (human-in-the-loop), including checking generated information against official sources [13]. The Court made this observation when examining a cassation appeal lodged by a public authority, the Main Department of the Pension Fund of Ukraine in Zhytomyr Region.

It did not establish AI use as a proven fact, but identified features of the submissions consistent with the possible generation of fictitious case-law references by an AI system. By analogy, it is relevant to administrative procedures: an official must verify the factual and legal basis of a draft administrative act prepared with AI assistance.

Unverified text cannot replace fact-finding and an adequately reasoned decision [2, Article 8; 13]. Proper human oversight does not confer an unconditional right on an official to alter an automatically generated result at will. Where powers are bound by law, such an alteration itself requires a legal basis [4; 5]. Where discretion is conferred, oversight

entails the ability to choose another lawful option and give reasons for doing so, rather than automatically approving the most probable answer.

6. The Duty to Give Reasons and the Limits of Judicial Discretion

The lawfulness of an administrative act requires adequate justification in law and fact. In its judgment of 11 September 2023 in case No. 420/14943/21, the Supreme Court identified criteria for the adequacy of reasons given by a public authority, including whether the reasons correspond to the established facts and the applicable legal provisions [9, paras. 45–48, 55–57]. The judgment concerns general requirements for reasoned administrative decisions, which are equally relevant to acts prepared with the assistance of algorithmic systems.

The judgment of 7 July 2026 in case No. 460/14758/24 is similarly concerned with general standards. The Court observed that the determination of legally significant indicators must rest on appropriate, reliable and verified information, rather than the merely formal application of prescribed thresholds or figures [10, paras. 62–71]. In its judgment of 26 August 2026 in case No. 160/34163/25, the Court examined the evidential value of an automatically generated notification that a document had been delivered through an electronic account. Where receipt was disputed on substantiated grounds, proper identification of the addressee, rather than the mere presence of a machine-generated record, was material [11, paras. 85–102]. In both

cases, verification of the reliability and relevance of the information used to justify a legally significant outcome was decisive.

Under Article 2(2) of the Code of Administrative Proceedings of Ukraine, a court examines whether a public power was used for its proper purpose, whether all material circumstances were considered, and whether equality, proportionality, the individual’s right to participate in the procedure and other prescribed requirements were observed. Article 245(4) distinguishes between ordering an authority to adopt a decision in the person’s favour where all statutory conditions have been met and no discretion remains, and requiring it to decide the matter afresh in accordance with the court’s legal assessment where the law preserves administrative choice [3, Article 2(2), Article 245(4)]. Accordingly, when exercising judicial discretion, a court reviews

the lawfulness of an authority’s exercise of discretion without replacing the authority’s choice between legally permissible administrative decisions with its own.

In its judgment of 13 October 2022 in case No. 380/13558/21, the Supreme Court distinguished powers that permit a choice between lawful alternatives from powers allowing only one lawful course of action [41, paras. 68–76]. As the competent authority had not decided the application on its merits, the appropriate remedy was an order to reconsider the application and issue a reasoned decision, rather than the court itself awarding the benefit [41, paras. 94–97]. The same distinction matters for

algorithmically assisted decisions: identifying a system error does not empower the court to exercise the administrative choice entrusted to the competent authority.

A seven-part test is proposed for judicial review of algorithmically assisted administrative decisions, comprising the following interrelated questions. First, is there a legal basis for using the system, and does the algorithm introduce restrictions on rights not provided for by law? Second, does the power involve discretion and, if not, does the system ensure the single lawful outcome? Third, are the data underlying the recommendation sufficient and reliable? Fourth, where discretion exists, has the authority retained a genuine legal, informational and technical ability to choose a different lawful option? Fifth, have the legally relevant facts, the criteria applied and any departure from previous practice been explained? Sixth, does the measure

chosen comply with the principle of proportionality? Seventh, can the system version, material data and sequence by which the result was produced be examined to the extent necessary to challenge the decision? [2; 3; 12; 15].

The first two questions concern the normative dimension of discretion; the third, fifth and sixth concern its case-specific dimension; and the fourth and seventh concern its technologically mediated dimension. A negative answer does not automatically require an administrative act to be set aside. The court must establish the legal significance of the defect, its effect on the individual’s rights and the applicable remedy. In disputes challenging the lawfulness of a public authority’s decision, Article 77(2) of the Code of Administrative Proceedings of Ukraine governs the burden of proving its lawfulness [3, Article 77(2)]. Nevertheless, the defendant authority’s burden of proof does not mean that any speculation about a possible technical error

is sufficient to set the decision aside. The specific defect alleged, the available evidence and the possible effect on the outcome are material. Such matters must be assessed under the rules of evidence laid down in the Code [3, Articles 72–77].

7. Legislative Proposal and Evaluation of the Effects of Automation

Ukrainian legislation should be supplemented by a rule distinguishing automated performance of bound powers from the use of software tools in exercising administrative discretion. To that end, it is proposed to insert the following subsection (4) into Article 62 of the Law of Ukraine “On Administrative Procedure”: “4. An administrative act may not be adopted through a fully automated procedure where the law leaves the administrative authority free to choose between several lawful decisions. Software tools may be used to support such a choice provided that an authorised official or another competent decision-maker designated by law has access to the material circumstances of the case, the criteria and data used, can verify

the system’s output, reject an unjustified recommendation and independently adopt

a decision within the powers conferred by law. The reasons given for an administrative act may not be confined to a reference to the output of a software tool.” The proposed wording takes account of the more cautious approach in § 35a VwVfG, under which administrative choice is retained by the competent authority [19], as distinct from the French model allowing fully algorithmic decisions subject to safeguards [47, paras. 69–72]. Any such amendment must be reconciled with personal-data protection, the confidentiality of restricted information and the allocation of powers prescribed by law. A risk-based system of oversight and documentation of technology use has also been proposed in the author’s published

scholarly draft Law of Ukraine “On Artificial Intelligence” (2026) [32]. That draft is used here as a comparative legislative proposal, not as law in force. The preference for a stricter model rests on the need to preserve the individual administrative choice provided for by law: technical and procedural safeguards do not, by themselves, eliminate the risk of an algorithm effectively substituting its output for that choice.

In addition to the performance indicators considered in section 2, evaluation of automated procedures should examine disparities in outcomes between comparable groups of applicants, instances of error correction and the frequency with which algorithmic recommendations are rejected, as measures of system performance rather than of an individual official’s performance. Whether independent review is necessary, and how often it should be undertaken, depend on the nature of the procedure and the risks it poses to individual rights [18; 25; 26].

Transparency does not require disclosure of personal data, software vulnerabilities or legally protected information. Information sufficient to enable a challenge may be made available to the individual and the court under different procedural safeguards.

What matters is that trade secrecy or technical complexity must not make it impossible to review the lawfulness of an individual administrative act [24; 37].

Conclusions

1. Administrative discretion is defined by law. Software architecture does not alter its

legal limits but may in practice restrict the decisions available to an official, the criteria used to assess them and the ability to reject an algorithmic recommendation. Legal scrutiny must therefore encompass not only the final administrative act but also the technological conditions under which its outcome was generated.

2. Automation of bound powers can reduce opportunities for arbitrary intervention

by officials. Where discretion exists, the use of AI requires the preservation of a genuine opportunity for an individualised, lawful choice. Reductions in corruption risks, processing times and public expenditure cannot be presumed and require empirical verification.

3. The case law of the Administrative Cassation Court within the Supreme Court

examined in this study confirms the importance of the lawfulness of algorithmic rules, the reliability of underlying data, the proper allocation of competence and the verifiability of automated outcomes. The use of technology does not relieve an administrative authority of its legal obligations.

4. The article proposes a seven-part test for judicial review of algorithmically assisted

administrative decisions, covering the legal basis, the nature of the power exercised, data reliability, the genuine availability of administrative choice, adequacy of reasons, proportionality and technological verifiability, without substituting the court’s decision for a choice entrusted to the administrative authority.

5. To safeguard the lawfulness of automated administrative decision-making, the

article proposes an amendment to Article 62 of the Law of Ukraine “On Administrative Procedure” prohibiting the fully automated exercise of discretionary choice and providing safeguards for meaningful human oversight.

Declaration on the Use of Artificial Intelligence ChatGPT (OpenAI) and Claude (Anthropic) were used in the preparation of this article to process research materials, produce working drafts and individual formulations, critically examine arguments, identify inconsistencies, edit language and carry out preliminary checks of source details. This description of their functions is consistent with the approach to transparent disclosure of tasks delegated to generative AI proposed in the Generative AI Delegation Taxonomy (GAIDeT) [33] and discussed in the author’s earlier research [34]. AI-generated outputs are not relied upon as independent primary sources. The author bears responsibility for the article’s content,

scholarly conclusions, accuracy of references and decision to publish. Status of the Research This publication represents the author’s own scholarly conclusions and does not express the official position of the Supreme Court, the Council of Europe or CEPEJ. The discussion of judicial practice is based on publicly available judicial decisions. The research does not disclose judicial deliberations or other restricted information.

Legislation, judicial practice and scholarly materials have been considered as at 9 November 2026. References

1. Constitution of Ukraine of 28 June 1996.

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

2. Law of Ukraine “On Administrative Procedure” of 17 February 2022, No. 2073-IX.

https://zakon.rada.gov.ua/laws/show/2073-20

3. Code of Administrative Proceedings of Ukraine of 6 July 2005, No. 2747-IV.

https://zakon.rada.gov.ua/laws/show/2747-15

4. Supreme Court judgment of 24 October 2023, case No. 420/11131/22, cassation

proceedings No. K/990/9882/23. https://reyestr.court.gov.ua/Review/114524417

5. Supreme Court judgment of 11 March 2025, case No. 460/22384/23, cassation

proceedings No. K/990/44317/24. https://reyestr.court.gov.ua/Review/125756659

6. Supreme Court judgment of 24 April 2025, case No. 280/3642/23, cassation

proceedings No. K/990/77/25. https://reyestr.court.gov.ua/Review/126854551

7. Supreme Court judgment of 26 November 2025, case No. 160/29658/24, cassation

proceedings No. K/990/28815/25. https://reyestr.court.gov.ua/Review/132102704

8. Supreme Court judgment of 20 May 2026, case No. 280/9417/24, cassation

proceedings No. K/990/20592/25. https://reyestr.court.gov.ua/Review/136682961

9. Supreme Court judgment of 11 September 2023, case No. 420/14943/21, cassation

proceedings No. K/990/8225/23. https://reyestr.court.gov.ua/Review/113405942

10. Supreme Court judgment of 7 July 2026, case No. 460/14758/24, cassation

proceedings No. K/990/20442/26. https://reyestr.court.gov.ua/Review/138037920

11. Supreme Court judgment of 26 August 2026, case No. 160/34163/25, cassation

proceedings No. K/990/25589/26. https://reyestr.court.gov.ua/Review/139254792

12. Supreme Court judgment of 15 September 2026, case No. 420/21964/23,

cassation proceedings No. K/990/15599/26. https://reyestr.court.gov.ua/Review/139771689

13. Supreme Court procedural ruling of 15 January 2026, case No. 240/14153/24,

cassation proceedings No. K/990/56043/25. https://reyestr.court.gov.ua/Review/133336040

14. Busch P. A., Henriksen H. Z. Digital discretion: A systematic literature review of ICT

and street-level discretion. Information Polity. 2018. Vol. 23, No. 1. P. 3–28. DOI: 10.3233/IP-170050. https://doi.org/10.3233/IP-170050

15. Committee of Ministers of the Council of Europe. Recommendation

CM/Rec(2020)1 to member States on the human rights impacts of algorithmic systems. 8 April 2020. https://rm.coe.int/09000016809e1154

16. Colombi Ciacchi A., Flórez Rojas M. L., Lane L., Nowak T. (eds.). JuLIA Handbook: AI

and Public Administration: The (legal) limits of algorithmic governance. JuLIA Project, 2025. https://www.julia-project.eu/sites/default/files/2025- 07/JuLIA_Handbook_AI_and_Public_Administration.pdf

17. Hoffmann M. The Mechanisms of AI Harm: Lessons Learned from AI Incidents.

Center for Security and Emerging Technology. October 2025. https://cset.georgetown.edu/publication/the-mechanisms-of-ai-harm-lessons- learned-from-ai-incidents

18. Council of Europe, Steering Committee for Human Rights (CDDH). Handbook on

Human Rights and Artificial Intelligence. April 2026. https://www.coe.int/en/web/cddh-handbook-on-ai-and-hr/home

19. Verwaltungsverfahrensgesetz (VwVfG), § 35a. Vollständig automatisierter Erlass

eines Verwaltungsaktes. Federal Republic of Germany. https://www.gesetze-im- internet.de/vwvfg/__35a.html

20. European Court of Human Rights. Glukhin v. Russia, application no. 11519/20,

judgment of 4 July 2023. https://hudoc.echr.coe.int/?i=001-225655

21. European Court of Human Rights [GC]. Ramos Nunes de Carvalho e Sá v. Portugal,

applications nos. 55391/13, 57728/13, 74041/13, judgment of 6 November 2018. https://hudoc.echr.coe.int/eng?i=001-187507

22. Court of Justice of the European Union. Ligue des droits humains, C-817/19,

judgment of 21 June 2022, ECLI:EU:C:2022:491. https://eur-lex.europa.eu/legal- content/EN/TXT/HTML/?uri=CELEX:62019CJ0817

23. Court of Justice of the European Union. SCHUFA Holding (Scoring), C-634/21,

judgment of 7 December 2023, ECLI:EU:C:2023:957. https://eur-lex.europa.eu/legal- content/EN/TXT/HTML/?uri=CELEX:62021CJ0634

24. Court of Justice of the European Union. Dun & Bradstreet Austria, C-203/22,

judgment of 27 February 2025, ECLI:EU:C:2025:117. https://eur-lex.europa.eu/legal- content/EN/TXT/?uri=CELEX:62022CJ0203

25. Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated version of 27

July 2026, as amended by Regulation (EU) 2026/1744. https://eur- lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727

26. Council of Europe Framework Convention on Artificial Intelligence and Human

Rights, Democracy and the Rule of Law (CETS No. 225), opened for signature on 5 September 2024. https://www.coe.int/en/web/artificial-intelligence/the-framework- convention-on-artificial-intelligence

27. 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 Administrative Cassation Court within the Supreme Court. Zenodo. 2026. Version 1.0. DOI: 10.5281/zenodo.22274766.

https://doi.org/10.5281/zenodo.22274766

28. Bernaziuk, Ian. Artificial Intelligence and Its Use to Promote Consistency of Judicial

Practice as a Component of Public Trust in the Courts. Slovo of the National School of Judges of Ukraine. 2024. No. 4 (49). Pp. 16–35. DOI: 10.37566/2707-6849-2024-4(49)-

2. https://slovo.nsj.gov.ua/images/pdf/2024_4_49/bernaziuk.pdf

29. Bernaziuk, Ian. Judicial Discretion and Administrative Discretion: Standards of

Proportionality and Reason-Giving in the Case Law of the Supreme Court. Constitutionalist. 2026. https://constitutionalist.com.ua/sudovyj-rozsud-i-dyskretsiia- administratyvnoho-orhanu-standarty-proportsijnosti-ta-motyvovanosti-u-praktytsi- verkhovnoho-sudu

30. Bernaziuk, Ian. Discretionary Powers in the Light of the New Law of Ukraine “On

Administrative Procedure”. Training presentation for the National School of Judges of Ukraine. 7 October 2024. https://supreme.court.gov.ua/userfiles/media/new_folder_for_uploads/supreme/20 24_prezent/99-%20administrative_procedure_discretion.pdf

31. Bernaziuk, Ian. Judicial Discretion and Administrative Discretion: Standards of

Proportionality and Reason-Giving in the Case Law of the Supreme Court and the ECtHR. Training presentation for the National School of Judges of Ukraine. 22 January 2026. https://constitutionalist.com.ua/wp-content/uploads/2026/01/164.- Judicial_Discretion_Administrative_Discretion_bernaziuk..pdf

32. Bernaziuk, Ian. Draft Law of Ukraine “On Artificial Intelligence” (Author’s Scholarly

Draft). Constitutionalist. 29 May 2026. https://constitutionalist.com.ua/proiekt- zakonu-ukrainy-pro-shtuchnyj-intelekt

33. Suchikova Y., Tsybuliak N., Teixeira da Silva J. A., Nazarovets S. 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. https://doi.org/10.1080/08989621.2025.2544331

34. Bernaziuk, Ian. Artificial Intelligence in Scholarly Research: Personal Intellectual

Contribution, Academic Integrity and Responsibility. Zenodo. 2026. DOI: 10.5281/zenodo.22257136. https://doi.org/10.5281/zenodo.22257136

35. Bell F., Bennett Moses L., Legg M., McNamara V. (with O’Reilly A.). AI Decision-

Making and the Courts: A Guide for Judges, Tribunal Members and Court Administrators. Australasian Institute of Judicial Administration, 2026 update.

https://aija.org.au/publications/ai-decision-making-and-the-courts-a-guide-for- judges-tribunal-members-and-court-administrators-2026-update

36. Sterck J. Artificial Intelligence before the Court of Justice of the European Union:

Asserting the Importance of a Human-Centric Dimension. Toyo Law Review. 2025. Vol.

69, no. 1. P. 225–252. DOI: 10.34428/0002001834. https://toyo.repo.nii.ac.jp/records/2001834

37. Council of Europe. The Administration and You: A Handbook. 3rd ed. Strasbourg:

Council of Europe Publishing, 2024. https://book.coe.int/en/international-law/11917- pdf-the-administration-and-you-a-handbook-3rd-edition.html

38. Committee of Ministers of the Council of Europe. Recommendation No. R (80) 2

concerning the exercise of discretionary powers by administrative authorities. Adopted on 11 March 1980 at the 316th meeting of the Ministers’ Deputies.

https://rm.coe.int/16804f22ae

39. Supreme Court judgment of 3 June 2024, case No. 380/7521/23, cassation

proceedings No. K/990/44355/23. https://reyestr.court.gov.ua/Review/119503759

40. Supreme Court judgment of 14 September 2026, case No. 560/9501/24, cassation

proceedings No. K/990/19285/26. https://reyestr.court.gov.ua/Review/139730748

41. Supreme Court judgment of 13 October 2022, case No. 380/13558/21, cassation

proceedings No. K/990/17977/22. https://reyestr.court.gov.ua/Review/106763738

42. Bernaziuk, Ian. Artificial Intelligence and Automated Decision-Making: Standards

of Transparency, Human Oversight and Judicial Review. Training presentation for the National School of Judges of Ukraine. 8 September 2026.

https://court.gov.ua/storage/portal/supreme/prezent2026/199_AI_Automated_Deci sion_Making_bernaziuk.pdf

43. Court of Justice of the European Union. Technische Universität München v

Hauptzollamt München-Mitte. Case C-269/90. Judgment of 21 November 1991. ECLI:EU:C:1991:438. https://eur-lex.europa.eu/legal- content/EN/TXT/HTML/?uri=CELEX:61990CJ0269

44. Ministry of Digital Transformation of Ukraine; National Agency of Ukraine on Civil

Service; Higher School of Public Governance. Advice on the Responsible Use of Artificial Intelligence by Public Servants. March 2025.

https://storage.thedigital.gov.ua/files/f/bf/a9595e0dcd238ab2b3602909107aabf9.p df

45. Bovens M., Zouridis S. From Street-Level to System-Level Bureaucracies: How

Information and Communication Technology Is Transforming Administrative Discretion and Constitutional Control. Public Administration Review. 2002. Vol. 62, No.

2. P. 174–184. DOI: 10.1111/0033-3352.00168. https://doi.org/10.1111/0033-

3352.00168

46. Northcott E. W. Automating street-level discretion: A systematic literature review

and research agenda. Policy Studies Journal. 2026. Vol. 54, No. 3. Article e70024. DOI: 10.1111/psj.70024. https://doi.org/10.1111/psj.70024

47. Conseil constitutionnel. Décision n° 2018-765 DC du 12 juin 2018. Loi relative à la

protection des données personnelles. ECLI:FR:CC:2018:2018.765.DC. https://www.legifrance.gouv.fr/juri/id/CONSTEXT000037148523 Suggested citation: Bernaziuk, Ian. Administrative Discretion in the Context of Automation and Artificial Intelligence. 2026. DOI: 10.5281/zenodo.24084506.

© 2026 Ian Bernaziuk.