>>AΙ, State & Public Administration

by Christos Koutsospyros

Automated decision-making in public administration is gaining increasing ground, radically transforming bureaucratic procedures and the state apparatus itself. The processes most suitable for automation are routine administrative tasks, such as the recording, registration, preparation, classification, and verification of the accuracy of information received by public authorities.

The use of Artificial Intelligence (AI) systems and algorithms promises to simplify bureaucratic procedures, allowing public servants more time to focus on tasks that require human judgment, creativity, and discretion.

AI applications are already being deployed in areas such as the provision of social services, fraud prevention, taxation, risk assessment and evaluation, personnel evaluation, strategic planning, and target-setting (OECD, 2025). Organizational effectiveness and efficiency in public services, as well as enhanced productivity of public servants, are the main arguments in favor of automated decision-making, as the ability to process large volumes of data (Big Data) allows public administration to rapidly extract critical information, thereby enabling better data-driven decisions.

Without overlooking the potential benefits of using such systems,¹ a growing body of literature documents and highlights the risks and concerns arising from the use of algorithms and automated decision-making, particularly concerning the fair and equal treatment of citizens (O’Neil, 2016; Eubanks, 2018),² while at the same time emphasizing that these technologies effectively function as tools for further bureaucratization, rationalization, and standardization of social life (Pasquale, 2015).

The introduction of AI into the state constitutes a fundamental epistemological shift, as has been noted, marking a transition from variables to patterns, from rules to correlations, and from surveys to sensors (Törnberg, 2025). This shift has the potential to transform public administration and core concepts such as the “Rule of Law,” “Citizenship,” and “Democracy” (Hildebrandt, 2015).

Risks

• Algorithmic bias and unequal treatment of citizens:
  Machine learning algorithms are highly sensitive to the quality of training data, which makes them vulnerable to learning spurious correlations and to reinforcing and reproducing social stereotypes, distortions, and discriminations based on race, gender, age, etc., embedded in historical data. This may jeopardize the rights of citizens—particularly those of vulnerable social groups—when decisions concern access to public services and economic resources.³
• Lack of transparency and explainability of administrative decisions:
  AI systems based on deep learning⁴  typically function as “black boxes”, rendering their internal operations and result-generation processes opaque and inaccessible even to the designers of these systems. As a result, it becomes extremely difficult to explain and justify—particularly adverse—administrative decisions taken with the assistance of AI, thereby undermining the principle of transparency in administrative action (Burrell, 2016; Kaminski, 2019).
• Inability to effectively exercise the right to prior hearing and to challenge automated decisions:
  When the internal functioning of the algorithm, the data taken into account, and the stages of decision formation are not understandable or explainable, citizens are deprived of the ability to exercise their right to be heard and to seek legal remedies against automated decisions. Consequently, decisions cannot be adequately reasoned to ensure their comprehensibility to their recipients and their effective review by the courts (Prevedourou, 2024).
• Undermining of administrative discretion:
  Algorithmic decision-making undermines the administration’s capacity to exercise discretion when choosing between multiple equally lawful options οr interpreting vague legal concepts such as “public interest,” “necessary measures,” or “serious grounds,” or adapting general and abstract rules to the specific circumstances of individual cases. This is because algorithmic decision-making relies on statistical correlations that produce a specific outcome and a predefined option for the human user of the system (Mitrou, 2023).
• Difficulty in identifying the precise role of an AI system in the decision-making chain:
  The technical “black box” of AI systems is reinforced by bureaucratic structures and internal administrative procedures. Complex legislation, internal regulations, and organizational policies may create obstacles, making it difficult to determine whether AI outputs are intended to function as final decisions producing legal effects for citizens or merely as decision-support tools for human decision- makers.⁵
• Degradation of work quality:
  Concerns have been raised regarding the negative impact of AI and algorithmic tools on the quality of work, as AI may render working conditions less satisfying and more stressful through work intensification, the expansion of new forms of workplace surveillance, and the promotion of an ultra-efficient yet simultaneously exhausting form of “digital Taylorism,” in which work is subject to increased monitoring and algorithmic micromanagement (Baiocco et al., 2022; Singh, 2025).
• “Automation Bias” and Human Oversight: People usually perceive AI systems and their decisions as neutral and unbiased, which creates the risk that public officials, by placing excessive trust in algorithmic recommendations, may end up merely ratifying system outputs, despite the technical limitations, without thoroughly reviewing the results and exercising meaningful human oversight—which in many cases may even be impossible for the average user (Koulu, 2020; Mitrou et al., 2022).
• Lack of Administrative Accountability: Public administration is expected to rely on the private sector—at least in the initial stages—for the design, implementation, and procurement of AI systems. In practice, this makes it particularly challenging to determine which data were used to train the algorithms, how and by whom they were collected, and ultimately who is the official issuer of an administrative decision, and therefore who bears legal responsibility: the state or the private developer who designed and programmed the algorithm on behalf of the administration. As a result, the content of decisions may no longer be shaped by public officials, who lack the means to modify the system, but by the developers and engineers who define its operational parameters (Prevedourou, 2024; Langer, 2024; Cobe, 2021).

Critical Observations

Beyond the aforementioned risks, it is also necessary to present some critical observations to highlight the socio-political dimension of introducing AI systems in public administration. These can be summarized as follows:

• The integration of AI into public administration is neither linear nor predetermined (Bijker, 2012). It is defined by diverse approaches, debates, and visions among different actors regarding what constitutes the “state,” “digital transformation,” “good governance,” and “sound administration.”
• Political choices regarding the adoption of new technologies tend to reinforce existing social structures and inequalities, as the technical design of these systems incorporates and reflects prevailing social biases (Winner, 1980) through the historical data used to train algorithms.
• There is a need to focus on the micro-level to understand how new technologies and the narratives surrounding them are employed by state actors both in everyday practices and in contests over power and resources.
• The impact of AI is not merely technological but involves the entanglement and interaction of human actors, organizational processes, rules, and infrastructures, resulting in the transformation of classical concepts such as “citizen,” “rule of law,” “discretion,” and “public sector.”
• The integration of AI into state functions tends to further shift governance from deliberative toward more technocratic models, where decisions rely on data and algorithmic predictions, simplifying the complex and multidimensional nature of social phenomena and reducing alternative approaches to understanding public policies.⁶
• AI is often introduced as a “technological fix” (Katzenbach, 2021) to complex bureaucratic processes, ensuring social consent and funding resources, while risks are perceived as manageable through preventive measures and the creation of “explainable” and human-supervised systems that can be evaluated using appropriate standards and methodologies. This fosters the illusion of full and absolute control over these systems, despite the inherent limitations of transparency and oversight (Ananny, 2016; Yampolskiy, 2024).
• The notion of creating “trustworthy AI” or “ethical AI” for the public sector, frequently mentioned in policy texts, can function as an empty signifier for forming political alliances among powerful actors with shared interests, obscuring power asymmetries both among the developers of these systems and between developers and users.
• The dominant discourse on AI tends to obscure the human labor required to operate and maintain these systems. Although AI systems are presented as autonomous and efficient, they rely on a multitude of invisible tasks performed by various actors (including public officials), such as data preparation and cleaning, algorithm supervision, identifying exceptions to general rules, and error detection. Hiding these tasks reinforces the illusion of AI’s neutrality and objectivity, whereas humans remain critical in shaping, interpreting, and applying the rules. Consequently, the role of public officials is undervalued, and the social and political relations that enable AI to function in public administration are obscured (Crivellari et al., 2024).  

Policy proposals

The following section seeks to outline several policy proposals aimed at mitigating the risks identified above, taking into account the critical observations highlighted in the previous section.

At a theoretical and research level, the following are proposed:

• Conducting empirical research on how citizens experience and perceive administrative procedures as users in AI-mediated environments.
• Conducting empirical research examining the interaction between public officials and algorithmic systems, as well as the implications of such interaction for decision-making quality, to assess how judgment, discretion, and oversight are exercised by officials over fully automated and decision-support systems (Mindel, 2024; Bainbridge, 1983).
• Conducting interdisciplinary research aimed at analysing requirements for algorithmic transparency, the conditions for producing human-understandable explanations, the development of AI models that ensure effective human oversight and intervention, as well as methodologies for auditing the internal functioning of algorithms.

At a more technical and practical level, the following are proposed:

• Requirement that public authorities provide a reasoned justification in each case for the automation of a process through the deployment of AI, specifying the expected benefits and risks, explaining why AI constitutes the most appropriate solution, and identifying which type of AI system is technically best suited to the intended purpose (e.g. supervised, unsupervised, or reinforcement learning, etc.).
• Εstablishment of structured and comparative assessments measuring AI performance against traditional methods—such as processing time for applications and impact on citizen service delivery—to assess the comparative advantages of introducing such systems, promote resource efficiency, and avoid unnecessary environmental burdens, given that the training and operation of AI systems are highly energy-intensive.
• Public funding of research aimed at developing open models and algorithms with transparent architectures, as opposed to “black box” systems (Rudin, 2019; Casper et al., 2023), as well as enabling synergies between academic institutions, research centres, and public authorities for the development of algorithms and models that respond to the needs of public administration and serve the public interest and social needs. It should be recognised that technology requires safety-by-design architectures and state capacity to develop them.
• Establishment of coherent and uniform definitions for the categorisation and classification of AI systems used by public administration, based on their purpose, technological type, and context of use, in order to ensure a shared frame of reference that enables effective communication among public bodies.
• Establishment of robust oversight mechanisms for public procurement and contracting of AI systems from the private sector, as public authorities—particularly in the early stages—are expected to enter into contracts with private AI vendors without sufficient scrutiny, thereby increasing the risk of vendor lock-in. Procurement contracts should therefore prioritise transparency and the public interest, rather than the promises and priorities of private technology companies.⁷
• Strict compliance with the institutional framework to prevent situations of “conflicts of interest” and “revolving doors,”⁸ given the initial dependence of the public sector on private suppliers, as noted above, so as to avert the risk of “public policy capture” by powerful actors through algorithms and AI systems.
• Drafting of a Code of Ethics for the ethical use of AI in the public sector, along with the delivery of training seminars to inform public officials and citizens about AI-related issues, in cooperation with experts from public administration and civil society.
• Optimisation and “cleaning” of public-sector databases, as a fundamental prerequisite and stage of the auditing process, in order to ensure that data are free from biases that may reinforce existing inequalities during the training of algorithms.
• Establishment of a public authority for the “Assessment and Oversight of Algorithms and AI Systems”, with the competence for auditing high-risk algorithmic systems, aiming to ensure compliance with the principles of transparency, accountability, and respect for human rights. In particular, the authority would assess the quality, completeness, and accuracy of the data used to feed AI systems, while also conducting bias and discrimination audits to identify potential biases embedded in historical training data. It would also evaluate the transparency of algorithmic model architectures and decision-making criteria. The authority would be equipped with adequate technical and material resources to fulfil its mandate and would be staffed by interdisciplinary personnel (including legal experts, computer scientists, historians, sociologists, etc.), while also cooperating with scientific personnel from research institutions. Audits would be conducted prior to systems being deployed in production and periodically thereafter. The authority would have the power to impose administrative fines, issue recommendations and guidelines, and would be required to publish annual audit reports setting out its findings.
• Creation of a dedicated unit within authorities with mediatory and advisory functions,⁹ appropriately resourced and staffed with specialised personnel, tasked with receiving and investigating citizens’ reports concerning violations of their rights by public authorities resulting from adverse individual administrative acts issued through AI systems, where such acts cannot be adequately explained or justified to their addressees due to the lack of transparency regarding the data used by the system to produce specific outcomes.
• Involvement of third-party independent actors in the auditing of algorithmic systems, aiming at providing the most comprehensive and objective assessment possible for identifying socio-technical risks. Independent external audits can reveal problems and biases that may remain invisible or be obscured in internal audits, as they aim not only at procedural and legal compliance but also at the substantive reduction of potential harm to users by enabling appropriate corrective interventions. A regulatory framework that allows researchers and civil society organisations access to data and models would further enhance transparency, accountability, and trust in these systems.¹⁰
• Establishment of minimum content requirements for the reasoning of administrative acts issued with the use of algorithms, so as to ensure their intelligibility to their addressees. Such decisions could include information on the extent of algorithmic involvement in decision-making, the data processed and their source, processing parameters, the operations performed by the system, as well as the factual elements (variables) taken into account. This would enable, on the one hand, the affected individual to ascertain whether the legally prescribed conditions for the examination of their case were respected, and, on the other hand, the judge to exercise effective judicial review (Prevedourou, 2024).

In conclusion, and based on the foregoing, it is evident that the decision regarding whether, in which areas, and how AI will be used in public administration and the state is not merely a technical matter. Rather, it constitutes a political and strategic choice that requires consultation and democratic dialogue, primarily with the users and recipients of the services being reshaped, in order to ensure that technological choices serve genuine social needs and the public interest. Instead of fully replacing or automating human decision-making, AI should enhance and support human users, if public services are to remain human-centered, accountable, transparent in their operations, and effective

REFERENCES

Ananny, 2016 M., & Crawford, K. (2016). Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media & Society, 20(3), 973-989. https://doi.org/10.1177/1461444816676645 .  

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8

Baiocco, S., Fernandez Macias, E., Rani, U. and Pesole, A., The Algorithmic Management of work and its implications in different contexts, European Commission, 2022, JRC129749.

Bijker, W. E., Hughes, T. P., & Pinch, T. J. (Eds.). (2012). The social construction of technological systems: New directions in the sociology and history of technology (Anniversary ed.). MIT Press.

Burrell, J. (2016). How the machine ‘thinks’: Understanding opacity in machine learning algorithms. Big Data & Society, 3(1). https://doi.org/10.1177/2053951715622512

Casper, S., Ezell, C., Siegmann, C., Kolt, N., Curtis, T. L., Bucknall, B., Haupt, A., Wei, K., Scheurer, J., Hobbhahn, M., et al. (2023). Black-box access is insufficient for rigorous AI audits. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’23). ACM. https://doi.org/10.1145/3630106.3659037

Cobbe, J., Seng Ah Lee, M., & Singh, J. (2021). Reviewable automated decision-making: A framework for accountable algorithmic systems. arXiv. https://arxiv.org/abs/2102.04201

Crivellari, I., Oskarsen, J.S., Grøder, C.H. (2024). Unveiling the Invisible: Caseworkers’ Data Work in a Public Welfare Organization. In: Janssen, M., et al. Electronic Government. EGOV 2024. Lecture Notes in Computer Science, vol 14841. Springer, Cham. https://doi.org/10.1007/978-3-031-70274-7_11

Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press.

Hartmann, D., de Pereira, J.R.L., Streitbörger, C. et al. Addressing the regulatory gap: moving towards an EU AI audit ecosystem beyond the AI Act by including civil society. AI Ethics 5, 3617–3638 (2025).  https://doi.org/10.1007/s43681-024-00595-3

Hildebrandt, M. (2015). Smart technologies and the end(s) of law. Cheltenham, UK: Edward Elgar Publishing. https://doi.org/10.4337/9781849808774

OECD (2025), Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions, OECD Publishing, Paris, https://doi.org/10.1787/795de142-en.

Kaminski, M. E. (2019). The right to explanation, explained. Berkeley Technology Law Journal. https://doi.org/10.15779/Z38TD9N83H

Katzenbach, C. (2021). “AI will fix this” – The technical, discursive, and political turn to AI in governing communication. Big Data & Society, 8(2). https://doi.org/10.1177/20539517211046182

Koulu, R. (2020). Proceduralizing control and discretion: Human oversight in artificial intelligence policy. Maastricht Journal of European and Comparative Law, 27(6), 720–735. https://doi.org/10.1177/1023263X20978649

Langer, C. (2024). Decision-making power and responsibility in an automated administration. Discover Artificial Intelligence, 4, 59. https://doi.org/10.1007/s44163-024-00152-1

Mindell, D. A. (2024). Τα ρομπότ μας, οι εαυτοί μας: Η ρομποτική και οι μύθοι της αυτονομίας. Angelus Novus.

McBride, K., van Noordt, C., Misuraca, G., &Hammerschmid, G. (2024). Chapter 4: Towards a systematic understanding on the challenges of public procurement of artificial intelligence in the public sector. In Research Handbook on Public Management and Artificial Intelligence. Cheltenham, UK: Edward Elgar Publishing. https://doi.org/10.4337/9781802207347.00013

Mitrou, L., Janssen, M., &Loukis, E. (2022). Human control and discretion in AI-driven decision-making in government. In Proceedings of the 14th International Conference on Theory and Practice of Electronic Governance (ICEGOV ’21) (pp. 10–16). Association for Computing Machinery. https://doi.org/10.1145/3494193.3494195

O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy.Crown.

PASQUALE, F. (2015). The black box society: The secret algorithms that control money and information. Harvard University Press. http://www.jstor.org/stable/j.ctt13x0hch

Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1, 206–215. https://doi.org/10.1038/s42256-019-0048-x

Singh, V. K.,& Gupta, R. (2025). From Taylorism to Algorithmic Management: How Technological Evolution Reshapes Foundational Management Principles.International Journal of Latest Technology in Engineering Management & Applied Science, 14(8), 1688–1695. https://doi.org/10.51583/IJLTEMAS.2025.1408000206

Törnberg, P., Söderström, O., Barella, J., Greyling, S., & Oldfield, S. (2025). Artificial intelligence and the state: Seeing like an artificial neural network. Big Data & Society, 12(2). https://doi.org/10.1177/20539517251338773     

Wachter, Sandra, Limitations and Loopholes in the EU AI Act and AI Liability Directives: What This Means for the European Union, the United States, and Beyond (July 01, 2024). Yale Journal of Law & Technology, Volume 26, Issue 3, Available at SSRN: https://ssrn.com/abstract=4924553 or http://dx.doi.org/10.2139/ssrn.4924553

Yampolskiy, R.V. (2024). AI: Unexplainable, Unpredictable, Uncontrollable (1st ed.). Chapman and Hall/CRC. https://doi.org/10.1201/9781003440260

Winner, L. (1980). Do artifacts have politics? Daedalus, 109(1), 121–136.

Μήτρου, Λ. (2023). Μπορεί ο αλγόριθμος. Αθήνα: Πανεπιστημιακές Εκδόσεις Κρήτης.

Πρεβεδούρου, Ε. (2024). Η ανθεκτικότητα των κλασικών διαδικαστικών εγγυήσεων: Αναβάθμιση του άρθρου 17 Κώδικα Διοικητικής Διαδικασίας (ΣτΕ 1206/2024). Prevedourou. https://www.prevedourou.gr/

Endnotes

1.  Focusing on the Greek context, examples of such applications—particularly for repetitive tasks with limited scope for human judgment—include the National Cadastre, where manual document processing has been replaced by an AI system that automates the reading and evaluation of contracts, reducing processing time from several hours to less than ten minutes. In the field of Civil Protection, AI is used for the early detection of wildfires and flood prevention. In the Justice sector, the use of natural language processing (NLP) algorithms supports the analysis of legal documents, enabling efficient search of case law and legislation, summarization and anonymization of documents, faster management and organization of court records, while other ongoing initiatives incorporate AI for speech-to-text conversion and an AI digital assistant providing guidance to citizens and lawyers regarding both electronic and non-electronic services of the Greek Justice System (OECD, 2025).

2.  The use of AI becomes particularly problematic when it assists—or, worse, replaces—human judgment in cases involving any form of “assessment” of individual characteristics and the exercise of broad administrative discretion. This occurs where various factors are weighed to adapt administrative action to the specific circumstances, such as the use of algorithms for the automatic issuance of Environmental Permit Decisions (AEPo), the selection of senior officials in public administration, the granting of asylum or subsidiary protection status, or the determination of optimal building conditions and restrictions in the development of urban plans (Prevedourou, 2024). Related concerns also arise in the evaluation of public servants (see: https://www.tanea.gr/2025/05/19/greece/i-ai-saronei-to-dimosio-online/).

3.  See, for example, the case of the erroneous indication of fraud in benefit claims affecting vulnerable groups in the Netherlands (https://www.politico.eu/article/dutch-scandal-serves-as-a-warning-for-europe-over-risks-of-using-algorithms/), as well as the case of the U.S. criminal justice risk assessment software “COMPAS,” which assigned higher probabilities of “high-risk” recidivism to Black defendants compared to white defendants, even when profiles (prior offenses, socio-economic data, gender, age) were similar (https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing).

4.  In this article, when reference is made to algorithms, we primarily have in mind “deep learning” algorithms, which constitute a more advanced type of machine learning that uses multi-layered “neural networks” to process data through successive levels of abstraction. This type of AI appears to generate most of the challenges and risks discussed here due to the so-called “black box” phenomenon.

5.  A related issue is the exponential propagation of an error during data input into the system, particularly when decisions are made through interoperability using data from multiple public authorities, making it extremely difficult to trace the original error.

6.  See the recent case of Diella, the “first AI Minister” in Albania https://www.bbc.com/news/articles/cm2znzgwj3xo

7.  The field of public procurement is considered particularly important due to the specific nature of AI as an object of public procurement, given issues related to data ownership, algorithmic biases, and commercial confidentiality. Special care is therefore required by the administration when drafting the technical and operational requirements of a project, so that terms and specifications are incorporated to address, prevent, and mitigate potential risks (e.g., the creation and use of standardized procurement templates to ensure consistency, uniform application, and validity of AI system procurement; the obligation to provide explainability of the system’s results in each specific case; or proof of the contractor’s compliance with the principle of data impartiality as eligibility criteria for the project, etc.). For more on this specific issue, see McBride et al., 2024 (https://doi.org/10.4337/9781802207347.00013).

8.  These are situations involving the movement of senior executives between the public and private sector, creating the risk that public policies are shaped not according to the public interest but in the interest of private companies (“regulatory capture”; see also Articles 68 et seq. of Law 4622/2019).

9.  This competence could be assigned to constitutionally established independent authorities, such as the Ombudsman.

10.  For more on the issue of “audit washing” and the importance of external oversight, see Hartman et al., 2024, while for the gaps in the AI Act, particularly regarding cases where the provider itself submits a conformity assessment without requiring external auditing, see Wachter, 2024.