Year: 2026
Type: dissertation
Abstract: What consequences will the widespread adoption of Artificial Intelligence (AI) in the public sector have on society?I was inspired by this question to conduct five studies using multidisciplinary and multi-methods research designs.While most research focuses on the private sector or pioneering countries in digitalization and public services, this work examines automated decision-making tools (ADM) and AI in Italian public welfare policies.Firstly, using qualitative methods, I explore how public employees perceive that AI adoption in public welfare policies can affect organizational and social factors such as administrative discretion and social inclusion.Secondly, relying on quantitative and experimental methods I examine two further social consequences: trust in AI applications and trust in institutions adopting AI.Besides advancing knowledge about this specific topic in Italy, the overall theoretical contribution highlights the role of social factors-including public values, social norms, and prior experience with institutions-in shaping the consequences of AI adoption rather than the technical characteristics of algorithms.Methodologically, this thesis generates a new dataset that combines survey items with experimental measures of trust in institutions and attitudes toward AI in Italy.Results from the qualitative study indicate that in Italy a fragmented and heterogenous AI adoption is occurring in the provision of welfare benefits and to detect fraudulent behaviours.These adoptions already imply relevant organizational and social consequences, reinforcing public employees' narratives of efficiency in terms of redistribution, social inclusion and the relation between the state and citizens.Survey items reveal balance between AI attitudes such as perceived benefits and risks, but contradictory patterns concern AI knowledge and awareness.Moreover, perceived benefits and social norms consistently predict higher trust in AI applications.Instead, perceived risks play a more heterogeneous role, particularly for high-stake domains such as healthcare or human resources compared to Chat-GPT like applications.Finally, a vignette experiment indicates that institutional trust is still shaped more by performance and prior experience than the decisionmaker-whether a public administrator, a hybrid system or an AI.However, context-dependent dynamics emerge, particularly between welfare-related and justice-related decisions.Despite the importance of the topic, our current knowledge is extremely limited.One broad view can be characterized by "technological optimists" who believe that the adoption of AI, and technology more generally, will lead to overwhelmingly positive outcomes for society in which disease and death are overcome, economies are revolutionized for sustainable production, and human welfare is maximized (Andreessen, 2023).These uncritically focus on the positive consequences of AI adoption.Conversely, "technological pessimists" anticipate and focus almost exclusively on the downsides of AI, with some of them arguing that AI poses non-negligible existential risks that arise from the problem of control (Bostrom, 2014; Yarvin, 2023).Concerns about discrimination and a lack of transparency have been commonly raised as potential issues.More concretely, what had initially attracted my attention some years ago was a scandal occurring exactly in Italy.Already in 2016, the Italian Ministry of Education introduced the "Buona scuola" algorithm that erroneously allocated teachers around the country (De Gregorio, 2017).Even more strikingly, the same was repeated in 2020 with the so-called GPS algorithm (Bizzini, 2023).The algorithm was an apparently easy automatic procedure to simplify teacher recruitment.Based on two rankings it evaluated CVs and cross-referenced teachers' preferences for location and classes with vacancies.However, the algorithm did not work.Thousands of teachers were subjected to discriminatory choices by being overtaken in the ranking by colleagues with lower scores.The errors caused confusion and even unemployment just before the beginning of the school year.Many people appealed and some cases ended up in courts.This resulted not only in big amounts of economic compensations but also in requests for transparency about the functioning of the algorithm.In the public debate the case was even known as the "Algofraud" (Algotruffa in Italian).papers lay out: this is the future (Bergmann et al., 2019; Berryhill et al., 2019; Farrell et al., 2023).Although AI is not a new concept, these cases coupled with some numbers highlight its recent renewed relevance.Already in 2020 it was estimated that the EU invested between 12.7 billion and 16 billion of euros in AI accounting for a relevant increase compared to previous years (figure 1-1) (Evas et al., 2022).The public sector accounts for less EU investments compared to the private one, 33% compared to 67%.Although the gap between the two sectors is still considerable, investments in the public one grew by 21%.At member level, Italy invested over 1 billion euros, like leading countries in AI investments such as Germany and France (figure 12).Despite these considerable sums, someone question whether countries are truly prepared to integrate AI into their institutional, economic, and social frameworks.The International Monetary Fund (IMF), for instance, has Despite this evidence, little is known at academic level.The scientific literature reporting adopted technological tools, examining existing case studies, and assessing social consequences is limited.While theoretical frameworks are scarce and borrowed from different fields of research, empirical evidence is often fragmented and mixed.For instance, most of the research analyse cases in Nordic and continental countries (e.g., Ammitzbøll Flügge et al., 2021; Andersson et al., 2022; Kuziemski & 1 https://www.imf.org/external/datamapper/datasets/AIPIThe remainder of this introduction presents the research questions, aims, multiple theoretical perspectives and methodological approaches applied in the various chapters.Before doing this, however, I provide a brief discussion concerning the terms ADM and AI, that have caused to me considerable conceptual challenges throughout this work.Institutional sources often, though not always, distinguish between ADM and AI.ADM refers broadly to processes in which decisions are initially, partially or fully delegated to automated systems (Chiusi et al., 2019).In this sense, ADM can be implemented with relatively simple technologies.A sub-setThe OECD similarly defines an AI system as (Berryhill et al., 2019): "A machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments.AI systems are designed to operate with varying levels of autonomy.In addition, AI are "machines performing human-like cognitive functions".These definitions stress three aspects.First, the functional and operational role of AI in producing outputs relevant for decision-making.Second, the different degrees of autonomy that these systems can have.And, importantly, they also highlight that AI encompasses a variety of dimensions.This includes rule-based methods, but also approaches such as machine learning, where systems detect patterns in large datasets ("Bigdata") and improve their performance without following explicit, human-coded rules (Berryhill et al., 2019).The boundaries across categories are clearly blurred.Even within AI, multiple classifications and perspectives exist.A common distinction is between General AI, also referred to as "strong AI" or "Artificial General Intelligence" (AGI) and Narrow AI or "weak AI" (Berryhill et al., 2019)."Weak AI" performs narrow, specialised tasks, while "strong AI" aims at general, human-like intelligence.This latter form remains the rarest if not still completely hypothetical.This is also the kind of AI represented in novels and movies, often dramatized in dystopian scenarios where machines outperform and control people, potentially fuelling scepticisms.Narrow AI, instead, is the one providing existing cases and tangible benefits and risks.It focuses on AI capacity to be more performative in terms of data elaboration compared to humans.In other words, it refers to technological methods that can elaborate larger amounts of data in shorter periods of time compared to humans.These AIs lack characteristics that are typical of people: handling ambiguous situations, intuition, creativity or emotions.Put it simply, the OECD affirms that "at present, all AI is Narrow AI" and it lacks the type of comprehensive intelligence typical of humans.Another classification is between rational systems, which follow fixed, logical rules, and learning rational systems, which can improve their performance through data exposure and training (Hallensleben & Hustedt, 2020).Similarly, the literature often differentiates between symbolic AI, Bocca, 2023).In this way, they are often conceptualized and intended based on their functions and implications.In the academic literature ADM and ADS are generally conceived in their broad dimension that can incorporate various techniques, including machine learning and forms of so-called narrow AI.In other Author technology definition Araujo et al 2020 automated decisionmaking "Decisions by technological means without human involvement".The process through which the ever-growing amount-and variety-of personal data "are subsequently processed by algorithms, which are then used to make (data-driven) decisions" Chapter 1 Desiere and Struyven, 2021 profiling models Rule-base profiling uses administrative eligibility criteria, such as jobseekers' age, educational level, and/or unemployment duration to classify jobseekers into client groups.Statistical profiling uses a statistical model to predict the likelihood of work resumption.AI Principles (OECD, 2019) stress
Authors: Pavolini, E., Guerci, M., Prelle, G.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7171604283