2.1 Policy Process and Knowledge

In looking at policy process, three characteristics of the field of policy science are essential. First, policy science deals with applied knowledge, and it is oriented toward problem-solving. Second, this knowledge requires contextuality. Third, the knowledge produced by policy science is interdisciplinary (De Leon, Reference De Leon1981). Understood as specialized, applied, and interdisciplinary knowledge, work with public policy involves the government’s epistemic agency to solve problems. Placing policy science as an epistemic agency means that policy decision-makers mobilize their action based on knowledge of problems and the positive connection of these problems with knowledge about solutions. Problems, represented through information, mean inputs for policymakers’ actions, constituting a solution or outputs of the government agency in organizational contexts. Working with public policy involves agents understanding a problem and acting to build a solution and new knowledge about policy outcomes and dilemmas. The distinctive characteristic of policy science is to constitute applied knowledge that connects public problems to solutions constructed in contexts of collective decision-making.

Public policy embeds professional and technical knowledge and constant learning to create solutions to public problems (Heclo, Reference Heclo1974; Dunlop & Radaelli, Reference Dunlop and Radaelli2020). The epistemic action of public policy professionals is shaped and infused by an interdisciplinary practical knowledge that enables them to reason and design solutions. Knowledge in policy science is essential because, in many situations, it organizes how policymakers and managers select a purpose, create means of action, and diffuse policy objectives. The actors’ actions lead to discussions about policy science’s theoretical framework. This theoretical framework leads to reflection on different models of understanding the construction of knowledge and its practical application by professionals from governments and interest groups.

Public policy is an applied, interdisciplinary, and professional-oriented knowledge field. The public policy profession occurs in the organizational dimension. For example, the garbage can model criticizes the policy cycle’s conception by pointing out that knowledge of solutions precedes knowledge of problems, shaping decision-making as complex, infused by bounded rationality, and driven by fluid participation. Cohen, March, and Olsen (Reference Cohen, March and Olsen1972) recognize that decision-making, policy formulation, and implementation adhere to the phenomenon of organizations. Organizations are the ones that fulfill public policy’s purpose and shape professional practice. Policymakers understand that, within organizations, public policies are artifacts to solve problems that involve professional practice within an organizational context and create public values (Barzelay, Reference Barzelay2019).

Public policy’s professional practice involves solving problems, which requires coordinated activity to mobilize multiple knowledge domains. Organizations make decisions, and policymakers design artifacts in their context to solve problems. Public policy is designed in organizational contexts, from which professionals engender a complex chain of interactions driven by analysis, prototyping, testing, evaluation, and decision-making (Simon, Reference Simon1970; Howlett, Reference Howlett2019). In this dimension, public policy is a function of practical knowledge absorbed or produced by policymakers’ experience within organizations, which shapes a professional discipline.

Public policy relies on professional knowledge shaped in diverse political contexts, grounded in a conception of knowledge that combines scientific perspectives with practical political applications. Thus, social science influences the practice of public policy, while the practice of public policy also influences social science. Knowledge in public policy is politically shaped and constitutes an element of practical action in defense of a collective purpose (Weiss, Reference Weiss1979).

The epistemic challenge means understanding the public policy process – that is, how knowledge of problems translates into knowledge of solutions to accomplish a purpose. For many analysts, the policy process is linear between problems and solutions, respecting rational policymaking and aimed at optimization (Braybrooke & Lindblom, Reference Braybrooke and Lindblom1963). On the other hand, there are those in which the policy process is non-linear, in which policies depend on political, ideational, or psychological factors because of the idea of bounded rationality (Simon, Reference Simon1947; Lindblom, Reference Lindblom1959). Finally, there is an intermediate perspective of rationality. Knowledge is a factor that organizes decisions rather than precisely making rationality. Mixed scanning, for example, works with the idea that policy decisions are not linear, but that knowledge is a rationalizing factor for actors’ strategies in different situations involving collective decision-making (Etzioni, Reference Etzioni1968).

Research is conducted over a given period or based on past dynamics to overcome the problems of irrationality or limitations. As Baumgartner and Jones (Reference Baumgartner and Jones2015) point out, information plays an essential role in the policy cycle. The dynamics of policy agendas, for example, are unpredictable and complex and are inherently difficult to forecast. However, the production of knowledge is central to the policy process. Governance is a problem of information: getting it (or restricting it), ensuring it is reliable (which sometimes means reliably consistent with one’s priors), and acting on it. Political institutions are seen through this lens, focusing less on representation, partisanship, or organized groups than on information processing (Baumgartner & Jones, Reference Baumgartner and Jones2015). Increasing the volume and quality of information, therefore, represents crucial elements for improving the performance of public policy in society.

Knowledge mechanisms, their artifacts, and their usefulness in the policy process define many aspects of what we can call policy science. Conceived from a scientific and professional perspective, within organizations, public policy involves the interdisciplinary connection between knowledge and practice, thus shaping action within governments. Disruptions, however, form new perspectives and changes that transform knowledge production and practice. We are experiencing a gradual disruption of public policy practice due to changes in the information regime and the emergence of digital technologies such as AI (Dunleavy & Margetts, Reference Dunleavy and Margetts2024). The disruption lies in AI transforming the epistemic roots of policy science and changing the everyday practice of public policy. The framework we work with in this Element is that AI changes the epistemic bases of government action in society by conceiving information through data and agency effects through complex algorithms that form AI-based systems. In the next section, we clarify the terms of the epistemic change.

2.2 Epistemic Changes and AI Integration in Policymaking

Regardless of the types of policy knowledge and decision-making, policy science’s epistemology changes over time. Today, the idea prevails that we live in a disruptive world in which innovative technologies change the epistemic roots of policy science (Hartley & Kuecker, Reference Hartley and Kuecker2022). This disruptive world is an interesting perspective to portray a moment when the practice of public policy changes radically. Traditionally, policy advice and learning sources are consultancies, political parties, academia, and civil society organizations, which select, curate, and promote problems (Craft & Howlett, Reference Craft and Howlett2013; Craft & Halligan, Reference Craft and Halligan2020).

Currently, we are dealing with an epistemic change that occurs with the volume of information expanding on social life’s diverse aspects. The volume is ballooning because of the data amassed, shared, and processed in everyday life. First, information and evidence abound on problems due to large volumes of data (Kitchin, Reference Kitchin2014). Second, there is an expansion of knowledge sources that change policy advice dynamics (Safaei & Longo, Reference Safaei and Longo2024). In addition to consultancies, parties, academia, and civil society, the volume of information is expanding, whether it is the result of social media or expert systems that structure digital governments (Giest, Reference Giest2017).

Although there are multiple perspectives on how knowledge influences policy decisions and tasks, the idea of having a greater volume of information is a vital factor for policy rationalization and organization (Baumgartner & Jones, Reference Baumgartner and Jones2015). Digital technologies have long played an essential role in the policy process. Since the foundations of policy science, computational systems have activated an imaginary of qualification and rationalization of the policy process through the dimension of professional techniques. Lasswell’s imagination, for example, is constructing a professional perspective of public policy mobilized by scientific knowledge, in which problems and evidence guide the work of policy analysts and can be instrumented by computational technologies (Lasswell, Reference Lasswell1970).

In Lasswell’s argument, the policy work, guided by “computerized” information, is essential in a scientific nature and in constructing a technical image of its professionals. Positive public policy, as Lasswell argues, has data and information and communication technologies as the basis for building evidence and positive government action in society. Furthermore, cybernetic conceptions of government place a vital element in data (Hood & Margetts, Reference Hood and Margetts2007). The role that information plays in the policy process is part of a cybernetic conception of government action, in which actors understand the environment, react to information about the environment, make decisions, and establish value on the consequences of their decisions and actions implemented over time (Hood & Margetts, Reference Hood and Margetts2007; Peters, Reference Peters and Levi-Faur2012). This cybernetic conception of government has been introduced previously and coincides with AI’s emergence and construction (Samuel, Reference Samuel1962; Deutsch, Reference Deutsch1963; Schwember, Reference Schwember and Bossel1977). Currently, the disruption caused by AI in public policy is the result of changes in societies’ information regimes and the advancement of technologies capable of processing this information to generate knowledge and practice (Dunleavy & Margetts, Reference Dunleavy and Margetts2024).

This idea of strengthening policy analysts’ techniques and profession through digital technologies was reactivated by the possibility of using digital technologies throughout the public policy process. The use of blockchain to rearrange public organizations (Clifton & Pal, Reference Clifton and Pal2022), the Internet of Things (IoT) to collect data and information in urban space (Chetfield & Reddick, Reference Chatfield and Reddick2019), the use of large volumes of data in the policy process (Giest, Reference Giest2017), and the use of robots in public health, industry, and transportation (Willems et al., Reference Willems, Schmidthuber, Vogel, Ebinger and Vanderelst2022) are examples of how digital technologies are becoming pervasive in the government domain. Digital technologies in government aim to strengthen policy science and create a new dynamic of more technical and evidence-oriented professional practices. This imagination became hyped with AI’s reinvention (Filgueiras, Reference Filgueiras and Howlett2022a). More than reimagining the possibilities of applying AI to public policy, this technology transforms the roots of knowledge.

In all the main theoretical traditions of policy science, AI transforms the possibilities and frames of policy creation. In positivist perspectives, AI has established transformations in the way evidence is produced and influences the policy process. Because large volumes of data are analyzed by machines and have practical consequences, the use of AI constitutes a new arrangement for the construction and use of information, implying changes in the positive policy (Newman & Mintrom, Reference Newman and Mintrom2023). Policy analysis is what professional public servants do when they use information to create advice for policy decision makers (Wilson, Reference Wilson, Goodin, Moran and Rein2009). AI transforms the analysis of evidence, accelerating the production of information and changing advice practices. From a positivist framework, AI generates scientific knowledge about policies, as it is based on formal statistical models that scientifically predict and advise courses of action for different actors in policymaking contexts.

From a critical policy analysis, public policy is analyzed through narratives and discourses that focus on power dynamics, social structures, and the ways policies shape and are shaped by social and political contexts (Fischer, Reference Fischer2016). In many situations, the use of AI shapes synthetic social and political contexts. Algorithms shape and are shaped by discourses, creating frames of practical action by humans (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). The use of generative AI in urban policies, for example, shapes synthetic contexts that boost anticipatory forms of governance, through which state actors seek to predict the future and strategically intervene in the present. As Cugurullo and Xu (Reference Cugurullo and Xu2025) point out, the use of City Brain, a generative AI, enable policymakers to manage multiple urban domains, including transport, safety, health, and environmental monitoring. City Brain uses large language models (LLMs) to generate visions of urban futures: visions that are in turn used by policymakers to generate new urban policies. The use of this generative AI creates frames and discourses for action, creating new risks and perspectives for the formulation of urban policies (Cugurullo & Xu, Reference Cugurullo and Xu2025).

Another example is how generative AI transforms the actions of low-skilled workers, adding more emotional discursive layers to automated work with AI. Automation is a limited conception in terms of discursive production. Low-skilled workers are unable to embed their own problems in the larger social context of generative AI’s impact on the labor market, let alone to politicize these problems. The impacts of new emerging patterns of interaction between humans and machines create the need for policies aimed at adapting workers to the context of AI (Oder & Béland, Reference Oder and Béland2025).

In a critical vein, the integration of AI into public policy requires a more cautious approach since the context in which AI is shaped by global asymmetries, uncertainties about the values that are introduced into algorithms, and how the practices that emerge are contingent and shaped by interactions between humans and machines (Paul, Reference Paul2022). In all these situations, AI transforms knowledge about public policy, which automates and expands analytical capacities, but without the backdrop of human oversight.

The integration of AI and public policy has more specific implications regarding epistemic issues. Policymakers’ actions change in response to the presence of new artifacts that produce and expand knowledge about policy solutions and problems. Epistemic agency refers to an individual’s capacity to form, revise, and justify beliefs and to participate in knowledge-related activities (Scholosser, Reference Schlosser and Zalta2019). Epistemic agency in public policy emerges from epistemic communities that share knowledge within a specific area to build expertise (Dunlop, Reference Dunlop2017). In public policy, epistemic agency refers to the capacity of policymakers to manage their beliefs and construct knowledge. Policy epistemic agency refers to the construction of vectors of practical learning within the context of epistemic communities aimed at utilizing techniques and technologies that shape practical action. In this sense, epistemic agency concerns the analytical capacities of policymakers to understand problems and propose solutions. Analytical capacities encompass a set of skills for processing data and evidence and proposing solutions related to various policy functions (Howlett, Reference Howlett2015). AI amplifies analytical capacities, as it shapes and revises policymakers’ beliefs and actions in unexplored directions.

AI promises a disruptive change in the practice when integrated into public policy. On the one hand, AI reshapes the public sector’s entire organization, modifying public agents’ capacities, proposing a more vertical governance style of policy sectors, and the horizontal allocation of power and functions between organizations through state integration, common capacity, and a needs-based convergence of services (Dunleavy & Margetts, Reference Dunleavy and Margetts2024). On the other hand, AI provides a new epistemology and patterns of epistemic action (Coeckelbergh, Reference Coeckelbergh2023; Floridi, Reference Floridi2023a), reflected in new analysis and new public policy practices by its professionals (Safaei & Longo, Reference Safaei and Longo2024; Taeihagh, Reference Taeihagh2025; Jaidka et al., Reference Jaidka, Chen and Chesterman2025). Thus, AI allows policymakers to modify their knowledge and reasoning and think about government action in a complex and different world, with a new order of problems and innovative technologies to shape solutions.

For example, Open AI, an owner of ChatGPT, has a trained and customized policy advisor chat, which promises to do the following:

• break down policy proposals, legislation, or regulations to assess their potential impacts, benefits, and drawbacks;

• offer insights on how to approach policy challenges, engage stakeholders, and navigate the policy-making process effectively;

• identify opportunities for innovation within public services and suggest ways to implement new ideas;

• help understand and interpret data relevant to policy issues, including economic indicators, public health statistics, or educational outcomes; and

• provide examples of successful policies from other regions or sectors that could be adapted or learned.Footnote ¹

This solution promises to replace policymakers’ knowledge practice, constituting another advisory standard from which they learn about problems and solutions, create cognitive structures, and reason about their actions in the government organizational context. Furthermore, it changes the entire practice of policy advice and learning (Henman, Reference Henman2018; Safaei & Longo, Reference Safaei and Longo2024).

AI represents a knowledge instrument that changes practices and interacts with humans to solve different problems. In public policy, AI goes beyond the conventional classification of policy instruments. AI is a pervasive instrument that can produce nodality, regulation, organization, and define fiscal management mechanisms. For this reason, AI is beyond the digitalization process. AI is not just an automation engine. It falls beyond the taxonomy of substantive and procedural digital instruments (Howlett et al., Reference Howlett, Giest, Mukherjee and Taeihagh2025) and should be considered a pervasive knowledge instrument throughout the policy cycle.

Hood and Margetts (Reference Hood and Margetts2007) classify policy instruments into authority, treasury, nodality, and organization. The arrangement and mix of instruments to achieve a policy objective depend on an information regime. Epistemic communities produce information, which is an essential input for public policy design (Dunlop, Reference Dunlop and Zahariadis2016, Capano & Malandrino, Reference Capano and Malandrino2022). AI modifies the epistemic regime of public policy in dimension of inputs, throughputs, and outputs, augmenting and aligning information through technology (König & Wenzelburguer, Reference König and Wenzelburger2020; Dunleavy & Margetts, Reference Dunleavy and Margetts2024). AI represents an epistemic instrument that overlaps other policy instruments, modifying the dynamics of policy design. For instance, AI helps policymakers change nodality, creating more direct forms of communication between government and society (Margetts & John, Reference Margetts and John2024). Similarly, AI transforms the way governments collect and administer taxes, changing all instruments of treasury (Mökander & Schoereder, Reference Mökander and Schroeder2024). AI also transforms public organizations by introducing new patterns of agency and organizational functions (Neumann, Guirguis & Steiner, Reference Neumann, Guirguis and Steiner2022). Finally, AI transforms the regulatory practices of governments (Ghosh, Saini & Barad, Reference Ghosh, Saini and Barad2025).

In other words, AI shifts the epistemic regime of public policy. Epistemic regimes are stable arrangements of practices of knowledge production and social structures (Knorr-Cetina, Reference Knorr-Cetina1999). Policy analysts frame the policy practices from epistemic regimes. AI changes the practices of knowledge construction about public policy, creating a new layer of complexity in public policy practice. AI is both absorbed as a new practice in policy science and results in the framing of professional practices of policy analysts. By changing these practices, AI transforms the institutional arrangements that guide the agency of professionals in political contexts. The above example of ChatGPT acting as a policy analyst is a construction of this perspective of change. If widely adopted by governments, this large language model could produce knowledge in an increased and rapid manner, with prospects of meeting the demands of humans related to a policy field.

An illustrative example of this process is how the Intergovernmental Panel on Climate Change (IPCC) has successfully laid a baseline of scientific facts for global discussions on climate action. Based on multiple layers and AI algorithms, the dashboard modulates complex scientific information about climate change, infusing policymakers with a new knowledge structure that modifies policy practice (Cowls et al., Reference Cowls, Tsamados, Taddeo and Floridi2023). AI applied at the IPCC provides climate modeling, prediction and simulations, and the design of catastrophic scenarios, as well as chatbots to expand the communication of IPCC data and reports and monitor government actions. In addition to providing information about climate data, AI applied in climate change supports studies on past environmental change around displacement hotspots and delivers future projections to inform adaptation measures and anticipatory action for integration in humanitarian programming. Similarly, AI can support disaster prevention, tracking pollution, and building climate-resilient agrifood systems that are more efficient, sustainable, and adaptable to climate change challenges. Within the practices of climate policy, data and information are modeled to predict and simulate scenarios that guide government action in innovative areas. AI organizes different climate policy challenges. AI climate policy implies new recommendations for practical action taken by policymakers and bureaucrats (Muccione et al., Reference Muccione, Vaghefi and Bingler2024).

The IPCC and the embedded AI have changed the structure of knowledge and action of policymakers around the world, providing new patterns of political disputes framed by the possibilities of global warming. The IPCC illustrates how AI, by modifying the information regime, modifies political action in different directions depending on the framing resulting from the complex work with data. The AI systems embedded in the IPCC represent complex knowledge instruments, thus altering the dynamics of environmental policy. In all these situations, predictions, simulations, and AI modeling modify policymakers’ epistemic action in the context of climate policies. First, climate change policy requires global negotiation and outreach, strategic partnerships, and implementation processes beyond issues internal to the nation-state. Second, climate change policies require the creation of networks to adapt global infrastructure, intensely shifting production and consumption chains. Finally, climate change requires policymakers to implement mitigation actions, which involve direct organizational changes in bureaucratic agencies and infusing behavioral changes in society. Mitigation actions require new formats, knowledge, and instruments for policy design (Braunerhjelm & Hepburn, Reference Braunerhjelm and Hepburn2023). AI-generated knowledge in climate policy frames political disputes, in which policy knowledge is challenged by conspiracy theories that put political action beliefs at stake (van Prooijen et al., Reference van Prooijen, Šrol and Maglić2025).

These changes in the epistemic action of policymakers, encouraged by an AI-based system, mean that AI is providing new sociotechnical standards for public policy and has consequences in society. A sociotechnical conception of AI in public policy means how human agents interact with this technology and produce new action patterns in complex organizational contexts. AI creates a new context for knowledge production and action that infuses policy changes in dimensions of government and its institutions. Understanding what AI is and the framework needed to think about its consequences is essential in the composition of AI in the policy process. Furthermore, changes in the information regime and the production of knowledge through AI in public policy require a critical view regarding their placement in the policy cycle. AI, as we will see in the next section, brings opportunities and risks to the policy cycle, assuming an ambiguous character in relation to the sociotechnical reengineering of public policies.

3.1 What Is AI?

Artificial intelligence is not a thing or a singular technology. AI does not have a precise concept in specialized literature (Wang, Reference Wang2019). The conceptual bases of AI are in machine intelligence to perform a task or solve a problem, which can vary according to the concept of intelligence (Russell, Reference Russell2019). AI is a computational technology that imitates human intelligence but can vary depending on how intelligence is conceived. The philosophical foundation behind AI development is imitation (Turing, Reference Turing1950).

It is important to emphasize that AI can be deconstructed in two dimensions. The first has to do with the difficult concept of intelligence. If machines imitate human intelligence, the first step is to define what intelligence means. The concept of intelligence focuses on human agency and understanding the drivers of action in different contexts to solve problems. Humans are intelligent not because they have intellectual abilities but because they can intervene to achieve a particular purpose. Intelligence can be many different things. Intelligence can mean the idea of rationality, considering the ability of humans to realize their preferences in each context and maximize the greatest possible utility (Simon, Reference Simon1957).

The concept of preferences is based on the premise that humans know perfectly well what they want and that they will choose an optimal option in many situations. Intelligence can also be considered bounded rationality, which implies that human action is not based on perfect preferences but on heuristics that constitute an informational shortcut that leads humans to decide on the most satisfactory option and not the optimal option (Kahneman, Reference Kahneman2003). Intelligence also can mean the human capacity to learn and form cognitive structures that allow them to perform in the world (Russell & Norvig, Reference Russell and Norvig2010) or perhaps the capacity for abstraction. Abstraction is a crucial function that enables humans to perform different tasks (Minsky, Reference Minsky1985). Intelligence can be structured thinking, in which humans understand, plan, and reason to achieve specific goals and solve problems (Markram, Reference Markram2006). In a more simplified conception, human intelligence derives from the fact that humans are machines for constructing patterns and storing those patterns as information to act (Kahneman, Reference Kahneman2003).

There is no need here to define intelligence exhaustively, but this concept creates initial difficulties. Intelligence is a human capacity to act based on abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem-solving. In all these definitions, intelligence means some practical knowledge aimed directly at action in the world through social interactions and exchanges (Goldman, Reference Goldman2003). Intelligence involves how humans infer information and construct knowledge to adapt their behavior to the environment through exchanges that exist in social interactions (Simon, Reference Simon1983; Goldman, Reference Goldman2003). Intelligence presupposes a capacity of everyone. However, it also has a collective dimension established in meanings, common actions, and shared political knowledge. The collective dimension of intelligence is, by definition, a political dimension (Landemore, Reference Landemore2012).

The other side of the AI concept is that we have the issue of artificiality. According to Simon (Reference Simon1970), this dimension of artificial entails the idea that systems interact with humans to achieve a particular purpose. Here, we point out an essential premise for understanding AI. What is different about AI systems is that they are in permanent interaction with humans through computational interfaces designed for human affairs (Cross & Ramsey, Reference Cross and Ramsey2021). When interacting with humans, machines incorporate different ranges of human problems to solve them and perform tasks (Reid & Gibert, Reference Reid and Gibert2022). Systems only have a particular form and behavior because they adapt to their environment to accomplish objectives or purposes in interaction with humans. Thus, human artifacts, in terms of their behavior, are artificial. Simon (Reference Simon1970) characterizes an artificial system as an interface between internal and external environments. These environments belong to the “natural science” domain, but the interface that connects them is the “science of artificial” domain, which comprises an interdisciplinary dynamic that converges different forms of knowledge for constructing a given artifact. When an artificial system successfully adapts, its behavior mainly shows the external environment’s shape and reveals little about the internal environment’s structure or mechanisms (Simon, Reference Simon1970).

We can construct a general concept of AI using these two dimensions. Artificial intelligence is a computer system capable of operating, in interaction with humans, different techniques to calculate, predict, and simulate courses of action from large volumes of data to support humans in decision-making and carrying out tasks. This definition embeds five essential elements of AI, especially if we consider this definition for applications focused on public policy.

1. 1. AI depends on large amounts of data to generate analysis, learning, reasoning, and action (Kitchin, Reference Kitchin2014; DeSouza, Reference DeSouza and Jacob2017; Dunleavy, Reference Dunleavy, Stoker and Evans2016).

2. 2. Different AI techniques start from the premise of imitating human intelligence performed by algorithms (Russell & Norvig, Reference Russell and Norvig2010). As we have seen, intelligence has many definitions, each of which implies diverse ways of thinking and reasoning to solve a problem.

3. 3. AI systems depend on interfaces through which machines interact with humans (Simon, Reference Simon1995).

4. 4. AI systems are epistemic when processing large volumes of data to support decision-making and task accomplishment (McCarthy, Reference McCarthy, Webber and Nilsson1981).

5. 5. An AI system’s constitution has a purpose; therefore, this system intends to affect agency issues directly and can carry out some action (Barandiaran, Di Paolo & Rohde, Reference Barandiaran, Di Paolo and Rohde2009; Tomasello, Reference Tomasello2022).

Because AI is many different things and not a singular technology, the label “artificial intelligence” must be understood in its epistemic and not ontological dimension (Alvarado, Reference Alvarado2023). AI is a field of research and knowledge, shaped by different techniques. Among the best known, machine learning is an AI technique in which computer systems calculate a specific output from learning vectors calculated in a structured database for training. In other words, they are algorithms that let the machine learn from the training database and produce an output from the learned vectors (Samuel, Reference Samuel1959; Domingos, Reference Domingos2015). Algorithms shape this learning process, and it can be supervised or unsupervised. It is supervised when the machine learns from a human-categorized database. This structured database serves as supervision of the knowledge constituted by machines; it guides – or supervises – how outputs will be constructed. The other possibility is that the outputs are unsupervised. In this case, the machines calculate learning vectors autonomously and use them to create the desired outputs.

Machine learning is the main technique, but the field of AI expands into other techniques. Among them, we have ensemble learning that aggregates two or more learners (e.g., regression models or neural networks) to produce better predictions (Zhou, Reference Zhou2012). Deep learning is another offshoot and subfield of machine learning. Using premises from neuroscience, deep learning entails AI techniques in which multilayered neural networks are applied to simulate the complex decision-making power of the human brain (LeCun, Bengio & Hinton, Reference LeCun, Bengio and Hinton2015). Reinforcement learning focuses on decision-making by autonomous agents. It is mainly applied in robotics, where an autonomous agent is any system that can make decisions and act in response to its environment independent of direct instruction by a human user (Sutton & Barto, Reference Sutton and Barto2018).

Within machine learning techniques, different applications exist, such as natural language processing (NLP); within NLP techniques, there are large language models (LLMs). LLMs are embedded in systems such as ChatGPT and Gemini, enabling machine interaction with humans. LLM is a computational language processing system designed to generate sequences of words, codes, or other data (more recently, images) from an input sequence called “prompt” and processed by transformer algorithms (Eisenstein, Reference Eisenstein2019). Another set of techniques is AI systems based on computer vision, which use images or videos from various sources to train systems that perform tasks that depend on image recognition. Facial recognition applications in public safety (Chen, Surette & Shah, Reference Chen, Surette and Shah2021) or autonomous cars, for instance, use computer vision techniques to perform tasks and decision-making (Padmaja et al., Reference Padmaja, Moorthy, Venkateswarulu and Bala2023).

Finally, we have large action models (LAM) at the forefront of AI advancement and disruption with AI agents. This AI technique combines elements of neural networks inspired by brain architecture with aspects of symbolic AI, which focuses on logic and symbols (Yang, Wu & Jiang, Reference Yang, Wu and Jiang2007; Zhuo et al., Reference Zhuo, Yang, Hu and Li2010). This fusion of techniques helps LAMs capture and model the complex relationships between user intentions and the actions performed by AI.

In short, AI systems are designed and implemented with large volumes of data, different algorithms, and computational power embedded in interfaces that shape human-machine interactions. Data represents numbers, images, texts, and any element from which we construct information that can be collected, stored, and shared in critical infrastructures (Kitchin, Reference Kitchin2014). Algorithms represent the “rules of the game” (Turing, Reference Turing1950), from which courses of action are calculated and infused into the interaction between humans and machines (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). In many situations, algorithms are the elements that institutionalize rules and procedures, shaping human behavior and choices through diverse interfaces. Finally, AI systems are produced from large machines that amplify computational power in data storage and processing. Computer power is about the computing infrastructure where data are collected and processed, and algorithms perform complex calculations. This infrastructure is typically shared by cloud models that can assume private, public, or hybrid ownership. This infrastructure defines many elements of AI’s geopolitics (Rikap, Reference Rikap2021; Lehdonvirta, Reference Lehdonvirta2024).

The goal is for machines to discover patterns of action and generate knowledge and implications for human action from data that characterizes a given environment. In other words, AI is a tool that supports humans in building knowledge to act in increasingly complex, ambiguous, and uncertain environments. Modeling supports humans in making decisions and perform complex tasks, generating rapid knowledge and the possibility of interventions in society (Arcas, Reference Arcas, Lane and Sethumadhavan2024). A machine becomes intelligent when it can learn how to understand – or model – its environment (Arcas, Reference Arcas, Lane and Sethumadhavan2024). When the machine becomes able to model its environment, it generates information that defines action situations in interaction with humans, defines informational frameworks, and produces action scripts that aim to achieve an objective. Modeling is at the heart of AI, so it allows a machine to autonomously understand its environment according to models built by humans to achieve an objective.

3.2 AI as Policy Instruments

In public policy, AI defines and frames the political context, chooses political biases from input data and information, produces knowledge outcomes, and affects both the automation of tasks and human action resulting from the knowledge created (König & Wenzelburguer, Reference König and Wenzelburger2020). Thus, when humans interact with AI to make decisions or perform various tasks, in the context of policymaking, the machine chooses the political biases of decisions and tasks performed by policymakers. In this context of policymaking with AI, several problems emerge, such as biases, transparency, and accountability (Papadakis et al., 2024). Integrating AI into public policy means creating unfamiliar problems, dilemmas, and challenges related to the interaction between the technical and political elements associated with policymaking (Taeihagh, Reference Taeihagh2025). Policymakers and analysts need to understand the technical elements of AI and the emerging challenges of incorporating this technology into the policy cycle.

We can establish different applications in policy processes by integrating AI. For example, facial recognition systems are applied to implement public security actions, systems that perform administrative tasks automatically, policy evaluation and monitoring processes, knowledge of citizens’ preferences, learning, and cognition on the part of policymakers to understand problems and propose solutions (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). The essential task for integrating AI into public policy is to have large data sets and model these data to achieve a specific knowledge goal so that the machine can understand the environment and receive feedback from humans about this environment and the action.

The debate on AI is twofold. Each perspective in the debate defines the thought on AI and its social construction. In the first approach, AI is conceived as agents that decide autonomously. The first perspective is that AI means the constitution of autonomous agents. These perspectives on the AI debate have different consequences for its integration into the policy cycle. This definition, therefore, involves the fact that AI represents agents or multiagents that perform tasks and make decisions (Wang, Reference Wang2019). This first frame implies thinking of AI as agents in the policy process. Thinking of AI as agents makes some sense if we consider that these agents perform actions to make decisions and carry out tasks autonomously. If AI is conceived as an agent, in the public policy cycle, it is a new actor in the decision-making process scene.

Criticism of this conception of AI agency assumes that, to have an agency, technology must make autonomous decisions. Autonomy is one of the assumptions of agency, and it is constituted controversially. The tendency to attribute autonomy to AI systems disregards that these computational systems do not have consciousness (Floridi, Reference Floridi2023a). Autonomy presupposes consciousness and intentionality, which obscures the idea of AI as agents or multiagents. AI is an artifact built by humans in human contexts.

This discussion is foundational in the field of AI. Since the classic work of Joseph Weizenbaum (Reference Weizenbaum1976), the idea of autonomy and AI’s anthropomorphization has been discussed. By anthropomorphization, we refer to attributing human qualities to machines, especially emotional abilities and intentionality. Claims to avoid anthropomorphization are recurrent, especially concerning LLMs and chats (Bender et al., Reference Bender, Gebru, McMillan-Major and Shmitchell2021; Coeckelbergh, Reference Coeckelbergh2021; Shardlow & Przybila, Reference Shardlow and Przybila2023). The fact is that within this anthropomorphization process is how AI represents systems that interact with humans to solve problems. In this process of interaction, the tendency of anthropomorphization occurs and solidifies in the human imagination.

The second perspective is to conceive these computational systems from an instrumental reason (Weizenbaum, Reference Weizenbaum1976). Instruments such as computational systems provide new possibilities for the world’s imaginative reengineering. Or, more specifically, in this Element, the imaginative reengineering of policy science (Lasswell, Reference Lasswell1970). However, according to Weizenbaum, the way computational systems designers recreate the world implies the prominence of an instrumental reason. This frame conceives AI as an instrument for human action in different fields, constituting knowledge and changing the nature of human action.

In this second frame, we can conceive of AI in the policy process as an essential instrument for policy actors, in which intentionality and knowledge depend on human action in technology design (Filgueiras, Reference Filgueiras and Howlett2022a). In the policy process, AI is an instrument to perform regulatory, nodal, organizational, and financial functions, altering the capacities and skills of policymakers and actors in different policy domains. AI is an epistemic instrument primarily designed, developed, and deployed for use in epistemic contexts such as policy analysis and research. AI is specifically designed, developed, and deployed to manipulate epistemic content such as data, and it is applied to do so, particularly through epistemic operations such as prediction and analysis (Alvarado, Reference Alvarado2023).

In the context of this Element, we adopt this second perspective for two reasons. First, AI in the policy process is an epistemic instrument that supports humans in making decisions and performing various tasks. There is no reason to imagine an AI that presents intentionality. Intentionality in public policy is collectively created. Since public policy requires intentionality to define political objectives, no AI technology can define them. AI does not define these objectives independently without human intervention. Therefore, it can frame political choices and change power relations (Lazar, Reference Lazar, Lane and Sethumadhavan2024). Second, AI in public policy involves affordances. AI can perceive an environment designed by its developers, who use affordances to represent potential actions that exist as a relationship between an agent and its environment. Applied to public policy, affordances are essential to AI design. Designers and the modeling of available data frame the affordances in AI design. The choice of these affordances means that AI systems, when applied in the policy process, do not constitute agents but instruments that support human decision-making and tasks (Maragno et al., Reference Maragno, Tangi, Gastaldi and Benedetti2023).

AI applied in public policy is a type of knowledge technology with different epistemic effects on how policymakers think, reflect, understand, or reason policy problems and solutions through interactions between humans and machines. The disruptive point of using this technology in public policy is that it produces epistemic changes in policy practice, creates new dilemmas and problems, and changes organizational dynamics from the moment that practitioners and AI-based computer systems interact to solve problems, produce analysis, and apply this knowledge to the everyday policy practice. Understanding these interactions between humans and AI systems is the next component of the analysis undertaken here.

3.3 Human–Machine Interactions and Policy

One of AI’s main characteristics is that humans and machines interact to create knowledge and act to achieve a purpose. The determining factor in AI’s conception is that humans and machines interact and constitute a new form of intelligence, shaped from data and institutionalized through algorithms (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). The attribution of purpose to computer-based artifacts emerges from each human action that shapes an immediate machine reaction (Turkle, Reference Turkle1984). Interactions between humans and machines imply an exchange of meanings based on how users instill a need and how the machine responds. The human-machine interactions create a new cognitive structure based on normative meanings, preferences, or opinions interchangeable in interactions and situated in a context for action. The postulate is that the capacity for human agency constantly reconfigures based on interaction dynamics between humans and machines (Suchman, Reference Suchman2007) and becomes institutionalized in response to action situations, frames, and algorithmically defined action scripts (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023).

As an epistemic tool, the contemporary world lives with hybrid intelligence, driven by humans interacting with machines to solve problems (Jarrahi, Lutz & Newlands, Reference Jarrahi, Lutz and Newlands2022). This characteristic of hybrid intelligence is how humans interact with systems. Hybrid intelligence is essential for understanding AI’s place in public policy, why AI changes the information regime and enables another action context. Humans and machines interact and create a feedback loop that transforms the epistemic conditions of public policy. In other words, humans and AI interact to develop policies through intelligence shaped by the interaction between humans and machines. Humans and machines co-evolve information, modifying the epistemic conditions of public decision-making (Pedreschi et al., Reference Pedreschi, Pappalardo and Ferragina2025). Under these conditions, humans insert inputs, receive outputs, react to them by giving new inputs, and receive new outputs in a loop that alters their understanding of the environment. By modifying the understanding of the environment, AI integrated into public policy changes the conditions of policy development and its consequences for society.

AI is a general-purpose instrument for use in different policy domains and as a different means to replace other instruments. AI can perform actions to exercise nodality, authority, treasury, and organization. Furthermore, AI transforms the way policy actors will act in the policy process. A hybrid intelligence’s character derives from the way AI, thought of as an instrument, changes the perceptions, reasoning, analysis, and policy work. The literature on policy instruments shows how they are not neutral because they produce specific effects regardless of the objectives pursued (Lascoumes & Le Galès, Reference Lascoumes and Le Galès2007). As a pervasive and general policy instrument, AI produces changes in the policy cycle, incorporating objectives defined politically by policymakers within a context of values, interests, and perspectives that shape the design of instruments and their reach in society.

For this reason, digital technologies such as AI, when integrated into public policy, use political affordances in their design. This means that the field of public policy must understand how instruments such as AI exert influence on its intended target audience (Hellström & Jacob, Reference Hellström and Jacob2017). The effects of AI when integrated into public policy are to facilitate action and learning. The notion of affordance involves the idea that action must be understood from the possibilities made available in object world, given the agents’ skills. Public policy developed with AI instruments defines the possibilities and scope of a public policy in society. AI changes the practice of public policy because this epistemic instrument defines, according to its design and modeling, the possibilities of actors in context of government action.

In different situations, humans interact with machines to design and build AI, and AI systems incorporate policy design and institutional dynamics to achieve policy objectives and augment public value. The disruptive point of AI for public policy is that these interactions between humans and machines change policy process dynamics, which are now carried out by other means, with different consequences and new epistemics for policy science. Interactions are constituted from meanings, and the sociological premise is that individuals interact by how they interpret these meanings, which they attribute based on norms (Blumer, Reference Blumer1986). These meanings are institutionalized through responses to action situations, frames, and action scripts that define motivations for human action based on algorithmically created knowledge (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023).

Interactions between humans and machines in the policy process will manifest in two dimensions. According to Daugherty and Wilson (Reference Daugherty and Wilson2018), these dimensions involve human dynamics in designing AI, on the one hand, and the way AI assists humans, on the other. This interactive dynamic between humans and machines depends on human inputs so that machines react to them and reciprocate feedback. Interactions between humans and machines are at the heart of AI, sparking discussions about machine intelligence (Searle, Reference Searle1980; Simon & Eisenstadt, Reference Simon, Eisenstadt, Bishop and Preston2002). Reciprocal feedback shapes new formats of knowledge, altering human action in society. Specifically in the field of public policy, interactions between humans and machines change the bases of the epistemic action of actors involved in the policy process because AI, in addition to being fed by diverse data, incorporates into its design political affordances that define the possibilities and alternatives of humans.

AI in public policy are systems that react to inputs posed by policy actors. The meanings attributed to this interaction between humans and machines in the policy process are still unknown. However, some clues exist in emerging experimental research. In the work by Alon-Barkat and Busuioc (Reference Alon-Barkat and Busuioc2023), AI’s adoption in public policy is driven by bureaucrats’ and decision-makers’ perceptions and stereotypes when they receive policy advice automatically. In other words, systems’ ability to augment data and information and constitute stereotypical policy advice based on the group’s biases engages AI adoption in government (Alon-Barkat & Busuioc, Reference Alon-Barkat and Busuioc2023). Human-machine interactions shape the meanings in the policy process, infusing ideas, norms, preferences, and perspectives that reinforce the biases of decision-makers and implementers. AI is dependent on the quality of the dataset, which may contain biases, omissions of excluded populations, or intentional designs toward certain target audiences. According to the experiment conducted by Alon-Barkat and Busuioc (Reference Alon-Barkat and Busuioc2023), policy decision-makers are more interested in the outcome and realize these results through their own biases.

In the policy process, these dimensions occur in the way developers and policy actors establish system design dynamics, in which humans assist machines so that they perform their functions. On the other hand, it matters how AI creates political and bureaucratic policy dynamics assisted by AI in policy formulation, implementation, and evaluation. In both situations, these interactions create normative, functional, and symbolic meanings that produce consequences for humans in the policy process and social behavior construction. AI reproduces old policy problems such as ideological bias, reproduction of prejudices and stereotypes, or established organizational dynamics that create implementation problems and design failures. Although AI is a rationalization mechanism for policy processes, it is not a mechanism of perfect rationality. Data sets limit the scope of AI, in addition to design problems in which humans set a goal for the machine. AI is closer to a machine of muddling through, in which decisions and tasks continue to be executed in contexts of uncertainty and bounded rationality (Cox, Reference Cox2019).

When AI is designed to be adopted in the policy process, it incorporates knowledge from various sources to instrumentalize public action, whether augmented or faster. Policy design requires internal dynamics from bureaucrats and policymakers regarding the choice of the mix of instruments and their disposal for coherence and consistency to achieve objectives (Howlett, Mukherjee, & Rayner, Reference Howlett, Mukherjee, Rayner and Howlett2018; Howlett, Reference Howlett2019; Capano & Howlett, Reference Capano and Howlett2020; Siddiki, Reference Siddiki2020). An AI that is applied in public policy involves defining the system’s scope, affordances, and characteristics, in which developers, bureaucrats, and policymakers interact to choose algorithm architectures, create databases, train systems, explain these systems, and evaluate their capacity and accuracy to fulfill policy functions, sustaining the system throughout the policy process. In both situations – policy design or system design – biases, stereotypes, values, and worldviews will be fundamental to understanding design choices (Forestal, Reference Forestal2022).

Designing AI means that humans constitute systems to perform a purpose. In many situations, systems design resembles policy design, depending on complex human interactions to choose algorithmic architectures and their instrumental requirements. System design means that AI systems can embody the entire policy design, perform functions as an implementing agent, or embody specific tasks by interacting with humans to solve problems. In this sense, the grounds for interactions between humans and machines is instrumental rationality to create AI systems to augment capacities (Weizenbaum, Reference Weizenbaum1976). In the policy cycle, AI is deployed to augment policy capacities (Filgueiras, Reference Filgueiras2022b).

In another direction, AI interacts with policymakers and bureaucrats in the policy process (Alon-Barkat & Busuioc, Reference Alon-Barkat and Busuioc2023). AI increases policy advice dynamics through recommendation algorithms, selection and hierarchy of preferences, clustering of social groups, and production of prediction and simulation on policy problems and solutions. AI is an instrument that produces organizational rationalization, new ways of thinking, and beliefs about objectives and future policy solutions. It is a pervasive technology driven by society’s desire to build a decision system through agents that can control the future and interact with humans differently, defining new modes of social action (Weizenbaum, Reference Weizenbaum1976; Nowotny, Reference Nowotny2021). Social action is the essential point of interactions between humans and machines in the policy process: AI transforms how policy actors understand and reason about problems and solutions through opaque technologies that support the entire construction of knowledge in different policy domains.

According to Daugherty and Wilson (Reference Daugherty and Wilson2018), the interaction between humans and AI in everyday life unfolds into three dynamics. First, AI amplifies analytical capacities and decision-making abilities. Policy decision-makers have bounded rationality because they cannot process all available information and then decide, but AI can look at everything that can be datafied, thereby increasing the human capacity to understand problems and solutions (Simon, Reference Simon1995). Second, AI makes it possible to interact with different audiences. In the policy process, policymakers and bureaucrats can interact with citizens, companies, and other stakeholders more comprehensively (Margetts & John, Reference Margetts and John2024). Interfaces amplify collaboration by the mediation of interactions (Campion et al., Reference Campion, Gasco-Hernandez, Jankin Mikhaylov and Esteve2022). Third, AI enables the embodiment of a robot that augments a human worker. With their sophisticated sensors, motors, and actuators, AI-enabled machines can recognize people and objects and work safely alongside humans in factories, warehouses, and laboratories. In the public sector, robotics is a future field involving automation and human interactions with robots that perform repetitive or complex tasks (Dunleavy & Margetts, Reference Dunleavy and Margetts2024). Amplification, interaction, and embodiment constitute the broader framework for AI in governments, with diverse consequences for performance of public policy and services. This discursive frame creates a technological hype that encourages governments to adopt technology uncritically and without knowledge of its limitations.

Considering these interactive dynamics between AI and policymakers, analysts, and bureaucrats, some questions arise. AI has a totalizing perspective on information, with the expectation of increasing and creating collaborative structures and embodied robots throughout the policy process. This idea is compelling in producing policy change, new modalities of policy learning and advice, new modes of institutionalization and governance, and new perspectives on working with public policy in an augmented way. Although there are many promises about AI in the policy process, doubts should remain about the capacity for practical improvements through technological change or whether new risks are emerging for old problems.

3.4 Risks, Problems, and Catastrophic Ambiguities from Epistocratic Policy

The epistemic changes infused into policy science by AI reinforce policy actors’ technocratic thinking (Hartley & Kuecker, Reference Hartley and Kuecker2022). This technocratic thinking derives from convergence defined by data patterns empirically, with greater or lesser receptivity to data and evidence. Technocratic bias is reinforced further when algorithms shape discourses, distribute resources, and increase or reduce people’s or groups’ visibility. Data and algorithms reproduce self-fulfilling prophecies in many situations through feedback loops, which forge innovative solutions to old policy problems (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). In this sense, using AI in the policy process is challenging precisely because it transforms the structure of knowledge, modifying the framing of problems, the knowledge of solutions, and the ways of intervening, governing, and producing changes in society. AI reinforces a governance style based on an epistocratic dynamic, which distrusts democracy as a problem-solving method.

This trend toward a public policy regime based on epistocracy means that the advancement of AI in the government sphere impacts the working of democracy. Epistocracy is “rule by the knowers” rather than “rule by the people” (Estlund, Reference Estlund2008). The hype of AI in governments spreads the idea that it can increase society’s trust in the functioning of government institutions, contribute to public integrity, and strengthen accountability (OECD, 2024). By integrating AI into public policy, governments assume a distrust of democracy and its capacity for collective decision-making. AI reinforces epistocracy and creates the conservative imaginary that machines deciding for humans is a better solution (Brennan, Reference Brennan2016).

The epistemic changes spread by AI occur through how it classifies, clusters, hierarchizes, correlates, produces causality, and performs tasks based on data. Following the logic of imitation, AI uses these procedures based on the philosophical premise that humans use these procedures to learn about things, people, and nature and, thus, make decisions and act. The problem is that we do not have complete information about the policy process. Actors’ biases shape decision-making. Ideas, values, and opinions shape policy process and sustain decisions and actions on a factual representation of the world rather than precisely on a background rationality (Béland, Reference Béland2019).

AI in public policy is a tool designed to change the perception and action of policymakers in diverse political contexts. In other words, AI integrated into public policy is a political instrument built from the biases and perspectives of its designers, who introduce affordances that define what is possible for the target audience of a given government intervention in society. Political, social, cultural, and economic biases are extremely important in composing the decision-making process and policy action (Lindblom, Reference Lindblom1959; Banuri, Dercon & Gauri, Reference Banuri, Dercon and Gauri2019). The fact that AI sustains its decisions on data does not mean that these decisions in the policy process are based on evidence, much less that it is a technical decision. Trial and error, decision-makers’ bias, and ideational heuristics are essential to AI design and turn technology into a political artifact. In this way, using AI in the policy process is closer to dynamics based on muddling through, requiring incremental decisions and constant learning (Cox, Reference Cox2019). Following the trail blazed by Lindblom (Reference Lindblom1959), integrating AI into public policy does not change the bounded rationality and complexity. Integrating AI into public policy may result in the reproduction of old problems in new knowledge artifacts.

AI, as an epistemic instrument, incorporates – through data – all these biases and conceptions of the world, causing policy advice and learning to occur in a way that discursively and rationally reinforces decision-makers’ biases (Eubanks, Reference Eubanks2018; Noble, Reference Nobel2018; Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). This incorporation of biases occurs from data, where bureaucrats and policymakers define algorithmic architectures and databases that will inform the decision-making and carry out tasks automatically. Training and operation databases for AI systems in public policy express – through documents and public records – the choices and decisions of bureaucrats and policymakers, making these systems optimizers of algorithmically institutionalized decision biases (Hong, Reference Hong, Paul, Carmel and Cobbe2024). AI in public policy requires caution, recognizing the challenges of biased or incomplete datasets and the potential for flawed models to generate misleading policy outcomes.

This dynamic design and use of AI creates a series of risks in the policy process. There are risks related to algorithmic injustice, organizational risks, and the possibility of rogue AI. Algorithmic unfairness arises from AI systems’ algorithmically institutionalized biases, relating to issues of gender, race, ethnicity, sexuality, and social groups. When algorithms reproduce racial bias, for instance, a person can see how certain conceptions of the world are assimilated in technical ways (Noble, Reference Nobel2018; Buolamwini, Reference Buolamwini2023). In many situations, AI algorithms reproduce biases or discrimination from a lack of diversity in the technological design process (Benjamin, Reference Benjamin2019; Hong, Reference Hong, Paul, Carmel and Cobbe2024). As a result, sexism, racism, and other forms of discrimination are built into the machine learning algorithms that underlie the technology behind many “intelligent” systems that shape how we are characterized and advertised.

In many instances, such discrimination stems from a computer industry comprised of few women, black people, or people of different sexual orientations, for example (Crawford, Reference Crawford2021). This issue of inclusion in the design of technologies applied in public policy is fundamental to avoid bias and ensure co-creation and transparency processes that enable correct application (Noble, Reference Nobel2018; Yuwono et al., Reference Yuwono, Tjondronegoro, Riverola and Loy2024). On the other hand, algorithmic injustice arises from data and the low visibility that data gives to topics related to social justice. Data incorporates society’s biases, and the extent to which it makes certain social groups visible or invisible is fundamental to AI’s performance in the policy process (Gitelman, Reference Gitelman2013). Public policy formulated and implemented with AI has the potential to create or reinforce inequalities. Algorithmic injustice requires that the design of AI applied to public policy introduce important innovations, such as inclusive co-creation of systems and algorithmic impact assessment focusing on inequalities in the policy process.

Organizational risks refer to the possibility of systems malfunctioning because of organizational failures. Simple bugs in an AI’s reward function could cause it to misbehave. For example, OpenAI researchers accidentally modified a language model to produce “maximally bad output.” Gain-of-function research – where researchers intentionally train a harmful AI to assess its risks – could expand the frontier of dangerous AI capabilities and create new organizational hazards. For example, in the pharmaceutical field, AI has a dual potential to discover drugs, depending on how human control and regulation are steered toward human well-being (Urbina et al., Reference Urbina, Lentzos, Invernizzi and Ekins2022). Public organizations such as the Federal Court of Accounts in Brazil use a ChatGPT-based instrument to optimize processes and perform repetitive tasks. Organizational failures and the absence of measures to control risks and avoid human errors produce catastrophic results for organizations. Public policy requires public organizations to build security measures and address organizational dynamics such as segregation of duties, internal controls, and systems security measures so that operations and policy tasks are unaffected. From a risk perspective, organizations demand actions in line with the principles of trustworthy AI and taking accountability to mitigate the risks (Curtis, Gillespie & Lockey, Reference Curtis, Gillespie and Lockey2023).

Finally, we have the risks of rogue AI. Rogue AIs are dangerous and powerful AIs that would execute harmful goals, irrespective of whether the outcomes are intended by humans (Bengio, Reference Bengio2023). Rogue AIs imply an idea of intentionally using powerful AI to produce social harm. The assumption is that there is a misalignment between the use of technology and the purposes for which it was designed. For example, there is the possibility of using deepfakes created with AI to influence elections on social media (Diakopoulos & Johnson, Reference Diakopoulos and Johnson2021), explore markets (Lin, Reference Lin2017), or generate political polarization (Jacobs, Reference Jacobs2024). Likewise, AI enables lethal autonomous weapons that identify human targets to eliminate them (Russell, Reference Russell2022) or to create bioweapons (Urbina et al., Reference Urbina, Lentzos, Invernizzi and Ekins2022). AI threatens humanity with catastrophic or even existential consequences in these situations.

AI is not a neutral instrument or an absolute technique in these risk situations. Humans designed and deployed it in diverse policy domains to accomplish diverse objectives based on data-driven knowledge. In this sense, AI is an ambiguous, dual, and opaque technology in the policy process, with diverse consequences for society. Conceived as a policy instrument, the possibility of optimizing the policy cycle is great in the sense of increasing information and the possibilities of organizational rationalization of governments, and augment policy capacity to achieve political objectives. The ambiguity lies in the fact that the design of AI systems can take on different facets and nurture the fantasies and beliefs of policymakers and bureaucrats (Filgueiras, Reference Filgueiras2022b), and in the same way, produce ambiguous results for society. Applying AI in the public policy cycle means taking risks and new possibilities to increase policy capacities.

4.1 Instrument Constituency and AI System Design

The debate over whether AI is an instrument or agent is substantial in the field of computer science. In terms of public policy, the definition of AI is as an epistemic instrument (Alvarado, Reference Alvarado2023) that transforms the way policymakers, decision-makers, bureaucrats, and networks make decisions and act based on knowledge generated with data. As a central epistemic instrument in the policy cycle, its constitution and design are essential. In the dynamics of interactions between humans and machines, modeling policy work implies a dynamic design of systems that will constitute actions so that AI can influence problem identification, formulation, decision-making, implementation, and evaluation. More specifically, in human-machine interactions, we address the way in which humans constitute AI-based instruments (Daugherty & Wilson, Reference Daugherty and Wilson2018).

Understanding how different agents act to create AI in public policy can be understood better in the dynamics of instrument constituencies. Instrument constituencies are networks of diverse types of institutions and organizations that share a common interest in promoting a specific policy instrument and related practices for their own benefit in material terms or sharing ideas (Voß & Simons, Reference Voß and Simons2014; Simons & Voß, Reference Simons and Voß2018). In general, instrument constituencies connect networks of scientists, design experts, consultants, public administrators, and technicians who design and deploy instruments in policy process. This actors’ network has an interest in developing, retaining, and expanding the instrument, and they work to institutionalize it in policy practice. In the instrument constituencies approach, we can see how to design AI to integrate it into the policy process. Political choices are at the heart of AI design.

In the case of AI conceived as an instrument, constituencies are formed by networks of the aforementioned actors plus developers in private companies that control the global communications infrastructure – big techs. Big techs are in the political stream, promoting AI as a solution to various social, economic, political, cultural, and environmental problems (Ulnicane & Aden, Reference Ulnicane and Aden2023; Khanal, Zhang & Taeihagh, Reference Khanal, Zhang and Taeihagh2025). Big techs are essential in the dynamics of instrument constituency because they approach a techno-solutionist perspective for different problems. Instrument constituencies act as a network in a policy context defined by the digital transformation package, with the potential to change all types of policy instruments (Margetts & Dorobantu, Reference Margetts and Dorobantu2019; Yeung, Reference Yeung2018; Dunleavy & Margetts, Reference Dunleavy and Margetts2024). In this policy context, public policy has its instrumental dynamics modified by the system design applied to automate and rationalize public organizations.

Constituting AI-based policy instruments is not a neutral activity, much less a technically constructed one (Russell, Reference Russell2019; Xavier, Reference Xavier2025). The AI instrument constituency involves networks of scientists, policymakers, bureaucrats, industry, design experts, consultants, and software developers who interact politically. These actors dedicate themselves to articulating and promoting solutions regardless of the problem context, with the aim of producing technologies. This network provides the encounter between the solution and the problem, mobilizing articulations around solutions in search of problems (Béland & Howlett, Reference Béland and Howlett2016). The legitimization of this process occurs within a context of functional evaluation of the instruments, extending their application in society. Moreover, this framework relies on the nature of the problems and how epistemic lenses underlie institutional and social relations, as well as the role the instruments can play within such structures (Lascoumes & Lé Galès, Reference Lascoumes and Le Galès2007).

Instrument constituencies deploy AI as a pervasive instrument across the policy cycle and change all policy instruments – nodality, authority, treasury, and organization. Instrument constituencies change the configuration of government, being a point of deep hype about AI’s potentials and challenges, and the visible and invisible interests that permeate political articulations. The design of technologies applied in governments is politically motivated, given the economic interests in retaining technological instruments (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). The legitimacy of these instruments is based on an order of efficiency as AI increases and amplifies the productivity of the public sector (Dunleavy & Margetts, Reference Dunleavy and Margetts2024). Discourses on the efficiency of AI policies abound, but evidence from productive sectors suggests that productivity gains are modest, with AI primarily serving to automate repetitive activities (Acemoglu, Reference Acemoglu2024). The point we argue in this Element is that AI does not necessarily amplify productivity but transforms the epistemic bases of policy instruments.

Although an epistocratic style of governance emerges from AI, this does not mean it is a neutral and effective instrument. Its constitution is permeated by economic interests that shape the perspective from which the government will adopt AI. For example, leading companies in AI development, such as NVidia, Amazon, Microsoft, SpaceX, and Meta, are major solutions suppliers to the US government and other governments worldwide. Currently, these companies engage in deep political debates with specific ideological perspectives according to the economic interests of a given situation. Since they control the global communication infrastructure, their power is strongly constitutive of the framing of technological solutions applied in governments (Rikap, Reference Rikap2021). The intersection of these instrument constituencies with the government organizes the AI hype and the framings of the policy objectives that AIs intend to achieve when integrated into public policy. In other words, big techs influence policy content because they integrate into political, policy, and problem streams and forge windows of opportunity to create policies (Khanal, Zhang & Taeihagh, Reference Khanal, Zhang and Taeihagh2025).

Instrument constituencies involve design AI systems from policy modeling, defining systems’ abstract objectives, conceptualizing and choosing algorithmic architectures, modeling databases, defining techniques, and coding and assessing accuracy, as well as validating systems for making decisions and carrying out policy tasks.

AI instruments have political implications as machine learning algorithms mean “a way of gathering and ordering society’s knowledge that fundamentally transforms how the state and society come to understand each other” (Amoore, Reference Amoore2022, p. 21). Furthermore, AI instrument constituencies act by monopolizing knowledge and turning data and information into commodified resources (Rikap, Reference Rikap2021). As a result, we have programmable and codified institutions that change policymakers’ behavior in a broader context (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). According to Louise Amoore (Reference Amoore2014), working with AI requires a different type of ability from designers who are more imaginative and intuitive, applied to governing society innovatively. AI instrument constituencies act by defining the steering of public policy, embedding all the policy cycle in systems that perform policy interventions and impact society.

In other words, AI-based sociotechnical systems are produced on an industrial scale and constructed on algorithms embedded in data processing platforms, shared or marketed in the cloud, with ready-made architectures that are customizable for different problems, in both public and private sectors. Scientists and developers promote the choice of algorithmic architectures based on trial and error concerning the problem, choosing the solution that presents the best accuracy and optimization (Amoore, Reference Amoore2022). The design dynamics of AI systems, in many ways, emulate the dynamics of policy design: elements of rationalization and optimization are sought to implement a politically shaped idea. AI instruments are politically constructed and guided by interests, perspectives, and opinions like policy instruments. The idea of austerity in economic crises infuses a digital transformation perspective (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). Furthermore, austerity makes governments dependent on private cloud-based data infrastructures that give big techs an intellectual monopoly on political ideas (Rikap, Reference Rikap2021).

An example of this process is how, in a context of strong conservative opposition to immigration policies, the use of AI to control borders has grown aggressively. This example of AI use is for them to classify, divide, and demarcate boundaries in data. These AIs use deep learning techniques that involve much more than the deployment of technologies at the border. These AIs are reordering what the border means and how to imagine the boundaries of the political community (Amoore, Reference Amoore2024).

Machine learning algorithms are solutions looking for problems within broader organizational processes (Filgueiras, Reference Filgueiras2022b). Although there are technical measurements of an algorithm’s accuracy, it does not allow checking the degree of adherence, coherence, and consistency when applied in the policy process. The technical accuracy of algorithms does not consider policy objectives, public values, justice criteria, or the solution’s effectiveness because systems designers consider the accuracy of knowledge, not its epistemological status, aroused in confronting reality. AI-based policy modeling means that the network of instrument constituencies interacts politically with a view to achieving a goal. Industry, developers, and scientists benefit materially from the industrial development of AI, defining actions aimed at the permanence and retention of AI-based instruments and their use in public policy.

The permanence and retention of AI-based policy instruments mean a struggle to adapt systems to policy contexts infused with broader political and institutional orientations based on interests and values. Ideas, therefore, count in the broader framework of technology design (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023). Thus, AI design is a political struggle to realize an idea framed in discursively expressed political interests of instrumental constituencies. In this case, AI design in public policy fits into techno-solutionism (Paul, Reference Paul2022; Sætra, Reference Sætra2023). Techno-solutionism is the metaphysical power of advanced technology to transmute the universe’s complex, indeterminate nature into obedient, mechanical certainty to be manipulated by the fantasies and fads of policymakers. The AI hype in public policy stems from techno-solutionist ideas that enable constituents to retain policy instruments (Khanal, Zhang & Taeihagh, Reference Khanal, Zhang and Taeihagh2025).

For example, the digital welfare state is based on the premise of redesigning all welfare policies through digital instruments. Machine learning algorithms can perform different tasks related to the welfare state, promoting greater access for society (Coles-Kemp et al., Reference Coles-Kemp, Ashenden, Morris and Yuille2020). However, they often are developed in line with a logic of control and dispositions around surveillance and efficiency that challenge careful engagements (Zakharova, Jarke & Kaun, Reference Zakharova, Jarke and Kaun2024). For example, the Danish government built a surveillance behemoth, dedicated to increasing surveillance against citizens receiving welfare state benefits (Kayser-Bril, Reference Kayser-Bril2020). In the case of Danish reforms between 2002 and 2019, the public sector transferred responsibility for key infrastructure to private actors through digitalization. As Collington (Reference Collington2021) points out, the main objective of public sector digitalization has been the growth of Denmark’s nascent digital technology industries as part of the state’s wider export-led growth strategy, adopted in response to functional pressures on the welfare state model. Fiscal stability drives the reforms in Denmark and has produced a retrenchment of critical assets and capacities (Collington, Reference Collington2021).

Instrument constituencies therefore are central to defining the frames of AI applied in public policy. The networks of actors participating in the instrument constituency strive to realize ideas that create the frames from which AI technologies will be deployed to realize policy ideas. The network of instrument constituencies strives to define the objectives, data modeling, choice of algorithmic architectures, coding, and framing of system outcomes, to compose all or part of the policy design through AI. Incorporating AI into policy design happens by trial and error, so that the policy analysis that results from using AI confirms and disseminates the results expected by constituents and their partners in an opaque and unaccountable manner. In many situations, AI applied to public policy confirms policymakers’ bias in a network of business and control of central infrastructures to develop opaque and unaccountable technologies (Alon-Barkat and Busuoic, Reference Alon-Barkat and Busuioc2023).

4.2 Policy Advice in the AI Era

Policy advice is the other side of the AI coin when applied to public policy. In the instrument constituency dynamic, we deal with how public policy actors and industry interact to build and feed AI systems. On the policy advice side, we consider how policy actors receive AI outputs and recommendations like advice to act in policy practice. Policy advice is the set of activities that support policymakers’ decisions by analyzing problems and connecting them with solutions and recommendations (Hallingan, Reference Halligan, Peters and Savoie1995). Policy advice activities start from problems and define courses of action for decision-makers (Althaus, Reference Althaus2013). Policy advice encompasses the epistemic nature of policy science and applies to the entire policy cycle (Wilson, Reference Wilson, Goodin, Moran and Rein2009). Thus, policy advice is a special type of policy work that connects problems and solutions. Furthermore, policy advice is a complex system that provides decision-makers with a political perspective through values and beliefs, on the one hand, and knowledge through evidence and information (Veselý, Reference Veselý2017). Disruption in public policy arises from the fact that AI provides policy advice, both in a macro dimension of policy overall, and in a micro dimension, which involves small decisions and tasks.

Sharing information, knowledge about problems, and policy recommendations for action is the heart of policy advice. Typically, multiple actors and multiple levels provide policy advice. Actors provide information, knowledge, and recommendations for action to policymakers. These actors include both individuals and organizations. Among multiple actors, policy advice encompasses the actions of consultants, academics, scientists, third-sector organizations, philanthropic organizations, or international organizations. There are levels of policy advice that include the influence of the knowledge generated on policymakers’ and bureaucrats’ action, on the one hand, and the broader organizational action of the government (Veselý, Reference Veselý2017). Furthermore, policy advice relates more directly to policy capacities, especially those related to analytical capacities (Craft, Head & Howlett, Reference Craft, Head and Howlett2024).

The practice of policy advice is fundamental in constructing knowledge and in converting it into policy action, both at the individual and organizational levels (Wilson, Reference Wilson, Goodin, Moran and Rein2009; Craft, Head & Howlett, Reference Craft, Head and Howlett2024). Policy advice builds knowledge through the analysis of outcomes, production of evidence, use of data in business intelligence, studies focused on a policy topic, or recommendations that emanate from experts and scientists in a given field of knowledge. Recommendations, therefore, are central to policy advice and they connect problems with changes in policy work (Veselý, Reference Veselý2017). Modeling AI systems for the entire public policy cycle transforms the dynamics of policy advice. The heart of AI’s use is to transform the entire epistemic basis of public policy through systems that perform predictions and simulations quickly and reliably. Using recommendation algorithms is fundamental to generate and disseminate content and change policy actors’ course of action. Algorithmic recommenders are systems aimed at generating meaningful recommendations for content or products that might interest a given set of users. Algorithmic recommendation systems’ main function is to estimate a utility function that automatically and mathematically predicts, ranks, and presents the user’s top preferences for a specific content or product (Schrage, Reference Schrage2020).

Government platforms embed recommendation algorithms and produce content for policy analysts. Another alternative is to create LLM that uses different techniques that enable the generation of knowledge and influence on action. For example, the Federal Court of Accounts (TCU) in Brazil developed ChatTCU, a chatbot based on the use of GPT-4 and the retrieval-augmented generation (RAG) technique to integrate all the specialized knowledge produced by auditors and system users in Brazil that recommends or generates audit and policy evaluation content. This content recommendation standardizes knowledge among auditors and represents a radical change in auditing and monitoring public policies implemented by the Brazilian federal government (Silva et al., Reference Silva, Santos, Miranda, Bezerra and Miranda2024). AI’s use adds a new interactive dynamic of knowledge production and infusion of action. Policy advice dynamics go beyond human interactions and incorporate relationships between humans and machines to shape policy content. The example of ChatTCU in Brazil means that knowledge emanating from public policy monitoring and accountability processes disseminates a new type of public action driven by interaction with chatbots.

LLM’s usage has the potential to produce automatic policy advisers, increasing science’s capacity to support the practice of public policy. AI has the potential to produce evidence syntheses in all fields of knowledge, such as medicine and health (Nowak, Reference Nowak, Lidströmer and Ashrafian2022), environment (Wani et al., Reference Wani, Rahayu and Ben Amor2024), and education (Ifenthaler et al., Reference Ifenthaler, Majumdar and Gorissen2024). AI has interacted with scientists and experts and changed how science operates, creating difficulties and disruption in the construction of knowledge. Large language models and other AI systems could be excellent at synthesizing scientific evidence for policymakers. However, this use requires appropriate safeguards and humans in the loop (Tyler et al., Reference Tyler, Akerlof and Allegra2023; Dwivedi et al., Reference Dwivedi, Kshetri and Hughes2023).

AI’s use shifts the logic of policy advice, increasing and accelerating knowledge and modifying the practice of policy work. AI transforms knowledge construction, transforming, in turn, individual actions, organizational frameworks, and institutional dynamics. AI will not replace policymakers, but it can enable a comprehensive, faster, and more efficient approach to policymaking in the short run and potentially offer policymakers different options. The premise is that in the dynamics of the policy process, humans and machines will interact to generate new forms of policy advice for recommending and shaping government action. Assuming the inherent risks, policymakers, bureaucrats, lobbyists, consultants, members of civil society organizations, and citizens will interact with AI to synthesize evidence, understand the problems, propose solutions, evaluate alternatives, and model the entire institutional and organizational architecture to policy implementation.

This dynamic occurs in a broader dimension of policy advice. It also occurs in a micro dimension, in which AI creates evidence, information, and knowledge that influences policy practice among street-level bureaucrats. For example, using predictive policing in security policies radically changes how security agents act in society. Predictive policing involves collecting a broad variety of data to estimate, through several correlations, when and where crime is likely to occur, thereby more efficiently employing existing resources to avoid it (O’Neil, Reference O’Neil2016; Meijer & Wessels, Reference Meijer and Wessels2019). Usually, machine learning makes predictions by artificial neural networks that shape police action. In public security, AI advises on micro actions fundamental to policy implementation, which changes the relationships between governments and citizens.

AI transforms the entire logic of policy advice, both in a macro sense, in the dynamics of formulation, and in a micro sense, shifting the implementation actions. Considering the way instrument constituencies act in the design of AI instruments, their deployment as an essential knowledge structure in implementation changes political decision-making’s entire composition and public action’s construction in society. Transforming policy advice’s structure means having AI instruments that change the knowledge dynamics of policymakers, in turn changing the outcomes of interventions and the impact on society.

4.3 Coda: AI Disruption in Policy Knowledge

AI in the policy process is a transversal knowledge policy instrument. This perspective challenges the foundations of policy science because AI transforms the way policy analysts produce knowledge, and it has practical consequences. AI is integrated throughout the policy cycle, creating knowledge about problems and the policymaking. AI is a knowledge instrument that is transversal to all other policy instruments, requiring that its integration be dependent on a new type of policy work: data modeling and algorithms for AI systems.

The transformations in policy work arise from interactions between policymakers and AI systems, adding a new layer to the policy cycle. On the one hand, policymakers participate in the modeling of AI systems, participating in instrument constituencies to solve problems. On the other hand, policymakers utilize the knowledge generated from these AI models to address problems of varying magnitudes, amplifying knowledge for policy formulation and evaluation, and automating elements and actions of implementation.

From a positivist perspective, we have the idea that policy scientists no longer produce knowledge. The idea that machines manipulate and operate policy knowledge creates a techno-solutionist imaginary, often present in Lasswell’s ideas (Reference Lasswell1970). On the other hand, the integration of AI reproduces many of the problems associated with muddling through (Lindblom, Reference Lindblom1959). In contexts where AI is integrated into the policy process, it does not change the bounded rationality, causing this integration to reproduce the idea that policies are incremental and driven by trial and error. The integration of AI constitutes machines of muddling through, in which knowledge is constituted without clarity of policy objectives, creating difficulties in the policy decision-making process.

In many situations, the integration of AI into the policy process creates new opportunities for mixed scanning (Etzioni, Reference Etzioni1968). AI enables a broad review of the decision-making field without requiring a detailed analysis of each alternative. This review enables the examination of alternatives created by prediction or simulation, leading to informed decision-making. Finally, the use of AI in the policy process challenges a critical perspective, as it creates discursive frames, particularly with generative AI (Oder & Béland, Reference Oder and Béland2025). The disruption of AI in public policy lies in its integration into policy analysis.

The disruption occurs in the foundations of policy knowledge because the use of AI throughout the policy process modifies the foundations of policy analysis. First, we have the pillar of knowledge in data and the construction of knowledge through sophisticated mathematical models (Floridi, Reference Floridi2023b). Even when we consider texts and speeches within critical policy studies as empirical objects, the use of generative AI creates new possibilities in the treatment of data and the construction of policymakers’ action frames. Second, disruption occurs because prediction and simulation become central to this dynamic, enabling the creation of fast machines to generate scenarios, predict problems, and inform solutions. Third, disruption occurs because AI becomes integrated as an automation tool, modifying the entire organizational structure where policymakers operate. In short, disruption occurs because humans and machines interact to generate policy knowledge and perform policy work.

The development and application of AI in policy demands that policymakers take actions to frame and outline AI policies, with special attention to issues of transparency, accountability, and responsibility. In this sense, modeling must adopt a series of procedures and instruments that we will analyze in the following section. AI applied in policy processes has an infinite range of recommendations for the action of policymakers, bureaucrats, citizens, and companies in contexts of decision-making and collective action. As previously highlighted, integrating AI in policy processes has risks and limitations, including data quality, algorithmic architecture choices framing, and development and coding in complex political contexts. AI changes the practice and scientific approach to public policy because it changes the knowledge structure and the situations, frames, and scripts for the policy action. Given AI’s limitations, problems, and risks, governance mechanisms for policies modeled with AI are necessary, which we will analyze in the next section.

5.1 Ethical Dilemmas and AI in the Policy Cycle

Digital technologies’ advancement tends to create an imaginative vision of greater political neutrality and accuracy of public decisions based on data (Esko & Koulu, Reference Esko and Koulu2023). Using AI in public policy is surrounded by an imagined world of policies implemented in a faster, more effective, safer, and more neutral way. AI applied in policy enables imagination to augmenting the productivity of public service and benefits for society. This imaginary place creates a tech-solutionist frame for AI in public policy, where we rely on and pursue using technology as the foremost solution, without heeding the social and political dilemmas that emerge. For example, in 2019, the Dutch government’s tax authority used AI based on a machine learning algorithm to create risk profiles to spot fraud among people applying for childcare benefits. In practice, this AI would recommend to bureaucrats whether families would be eligible to receive the benefit. Thus, authorities penalized families over a mere suspicion of fraud based on the system’s risk indicators. Tens of thousands of families were pushed into poverty because of exorbitant debts to the tax agency. Reports indicate that some victims committed suicide. The scandal resulted in more than a thousand children being taken into foster care (Newman & Mintrom, Reference Newman and Mintrom2023).

This case of the Dutch government demonstrates how AI can incur ethical dilemmas from the perspective of its use and application in the policy process. This techno-solutionist vision provides a frame of opaque technologies that infuse human actions by defining individual and collective choices. Within human-machine interactions, these frames are constructed discursively and based on ideas and values that make AI-based instruments ambiguous and powerful in defining action (Khanal, Zhang & Taeihagh, Reference Khanal, Zhang and Taeihagh2025). The Dutch government’s goal was to reinforce austerity policies and improve the provision of public benefits with the moral duty of fairness and honesty. It is not possible to discuss here whether the Dutch government acted well or badly. The fact is that the use of technologies is driven by discursive frames that instill values and norms into the use of technologies (Mendonça, Filgueiras & Almeida, Reference Mendonça, Filgueiras and Almeida2023).

The sociotechnical reengineering of public policy has a series of implications for society and the political system, centering debate around the use of AI as a complex set of social dilemmas, normally discussed from a risk perspective. A social dilemma is a situation of interdependence between people in which there is conflict between doing what is best for oneself versus what is best for the group. The social dilemmas of AI in public policy emerge from human-machine interactions and the mode of how algorithms institutionalize policy practices and knowledge. By delegating the task of solving problems and making decisions to an AI, a false image is created that social dilemmas in public policies have been overcome. Sociotechnical reengineering of public policy with AI produces social dilemmas because of ethical issues.

The first dilemma is epistemological. AI introduces a dilemma related to its predictive power and how it resonates with human action. Deploying AI in a policy cycle has the goal of predicting and simulating all decision-making and tasks. According to Nowotny, we use AI to increase our control over the future and uncertainty, while AI’s performativity – the power it has to make us act in the way it predicts – reduces our agency over the future (Nowotny, Reference Nowotny2021). The way policymakers delegate to AI the power to decide and perform tasks reduces the margin of human control over technology. AI creates the sense that public policy governance is data-driven and neutral. However, algorithmic governance does not supplant democracy because of an epistemic impossibility. Even if AI were to exercise algorithmic governance, it would not supplant democracy because humans continue to construct data and information that feed decision-making (Innerarity, Reference Innerarity2024).

The spectrum of complexity and uncertainties in public policy means that defining problems and constructing solutions depends on a new type of knowledge, which is comprehensive and uncertain, yet granular and focused on individuals. However, the modeling policy work based on AI imposes many challenges for governments in designing and implementing effective policies to govern AI. We do not fully understand the problems posed by AI, which makes the technology itself unpredictable, intractable, and nonlinear, making it onerous for governments to create an institutional framework and correct objectives for their policies (Gasser & Almeida, Reference Gasser and Almeida2017).

The second dilemma is controlling technology. Not knowing the problems and uncertainties related to AI makes the paths forward ambiguous, as we expect greater knowledge with AI’s deployment in public policy, while at the same time we do not know its dynamics for producing knowledge. Social dilemmas arise in the way in which the framing of AI in public policy implies interventions in human life, without us having control over it (Russell, Reference Russell2019). An example of this issue of control is the discussion on lethal autonomous weapons (LAW) at the United Nations Office for Disarmament Affairs. AI is embarked in drones and robots that search for, identify, and eliminate targets autonomously, without human intervention (Russell et al., 2022). The United Nations creates debates on banning LAW. However, these debates were hampered by veto issues from countries that are major players in this technology, particularly the United States, Israel, China, and Russia. Recently, the Lavender system, a LAW developed by Israel, was widely used in the conflict in Gaza. Lavender was developed to identify people linked to Hamas or Jihad, with an accuracy of 90%. There are reports that the use of Lavender in Gaza resulted in the elimination of civilian targets in their homes, which were systematically bombed.Footnote ²

Because AI significantly reduces human control over decisions, it creates new challenges for ascribing responsibility and legal liability for AI’s harms imposed on others. AI is a technology that learns and adapts to the environment by following the rules set out in algorithms, without humans being able to control the results and impose responsibility on systems. So, the unpredictability of machine learning-based decisions implies that many erroneous decisions made by AI are beyond the control of and cannot be anticipated by society (Lim & Taeihagh, Reference Lim and Taeihagh2019).

The ethical challenge of AI in public policy concerns the extent to which governments can construct interventions in society, prioritizing people’s lives, while using technologies whose decisions are unpredictable and uncertain. These decisions are subject to diverse forms of algorithmic injustice (Eubanks, Reference Eubanks2018), invisibility of identities or visibility of prejudice (Noble, Reference Nobel2018), or even policy failures that cause harm in society. The Dutch case is an example in which the harms caused by AI create ethical dilemmas and require innovations in governance processes that go beyond traditional patterns in policy theory. The ethical dilemma of the Dutch case lies in the fact that wanting to do the right thing with AI – regardless of what we think is right or wrong – produces unforeseen effects on society, making it necessary to have a governance framework that is also uncertain and experimental.

The epistemological dilemma posed by Nowotny (Reference Nowotny2021) extends into uncertainties and the way in which delegating decisions to an AI system produces consequences for society. According to Floridi, the world of AI produces a divorce between agency and intelligence. Epistemologically, AI as a new form of agency can be harnessed ethically and unethically (Floridi, Reference Floridi2023b). When AI is applied throughout the policy process, this divorce is amplified. On the one hand, it produces epistemic changes in policy work, in which policymakers, bureaucrats, citizens, and corporations delegate decision-making and task performance to AI systems, changing the entire policy advisory system and forging actors’ actions shaped with an artificial (rather than a properly human) intelligence. Within the interactions between humans and machines in the policy process, AI as an epistemic instrument can be modeled and deployed ethically and unethically. In this sense, faced with the ethical dilemmas of AI, we need governance frameworks that make it possible to instill practices and procedures that ensure ethical AI development.

AI governance in public policy starts from principles that frame AI development and deployment to governments and companies, creating a discursive and normative frame. AI ethics require principles that guide innovation and technological deployment. AI ethics do not have a consensus about the principles. However, international organizations such as the Organization for Economic Co-operation and Development (OECD) disclose and disseminate principles. The AI principles from OECD include:

• inclusive growth, sustainable development, and well-being;

• human rights and democratic values, including fairness and privacy;

• transparency and explainability;

• robustness, security, and safety; and

• accountability.

AI principles are not a “one size fits all” approach, but rather action guidelines that frame institutional arrangements to governance. Similarly, organizations such as UNESCO advocate an ethical perspective on AI that is grounded in strengthening democracy. The cornerstone of the United Nations Educational, Scientific and Cultural Organization’s (UNESCO’s) “Recommendation on the Ethics of Artificial Intelligence” is the advancement of fundamental principles such as transparency and fairness, while always remembering the importance of human oversight for AI systems. The principles formulated by UNESCO comprise a framework for AI development aimed at the following:

• Proportionality and do no harm: the use of AI systems should not go beyond what is necessary to achieve a legitimate objective. Risk assessment should be used to avoid harm that may result from such uses.

• Safety and security: unintended harm (security risks) as well as vulnerabilities to attacks should be prevented and addressed by AI actors.

• Right to privacy and data protection: privacy should be protected and promoted throughout the AI lifecycle. Adequate data protection frameworks should also be established.

• Collaborative, adaptive, and multistakeholder governance: international law and national sovereignty should be respected in the use of data. Furthermore, multistakeholder engagement is necessary for inclusive approaches to AI governance.

• Responsibility and accountability: AI systems should be auditable and traceable. There should be oversight, impact assessment, auditing, and due diligence mechanisms in place to avoid conflicts with human rights standards and threats to environmental well-being.

• Transparency and explainability: AI systems’ ethical deployment depends on their transparency and explainability (T&E). The level of T&E should be appropriate to the context, as there may be tensions between T&E and other principles such as privacy, safety, and security.

• Human oversight and determination: Member States should ensure that AI systems do not displace ultimate human responsibility and accountability.

• Sustainability: AI technologies should be assessed against their impact on sustainability, understood as a set of evolving goals, including those set out in the United Nations Sustainable Development Goals.

• Awareness and literacy: Public understanding of AI and data should be promoted through open and accessible education, civic engagement, digital skills training and AI ethics, and media and information literacy.

• Fairness and non-discrimination: AI actors should promote social justice, impartiality, and non-discrimination, while adopting an inclusive approach to ensure that the benefits of AI are accessible to all.

From this frame, AI ethics guides on the reflection of human-machine interactions, in which the discussion on the moral status of AI is essential. The AI ethics framework addresses the moral implications of AI to society, assessing the moral status of AI instruments and their social, economic, cultural, and political effects. The principles guide human action to design AI systems and the working of these systems in human-machine interactions. AI ethics’ primary claim is that humans create mechanisms and instruments to control AI. This claim is appropriate to public policy, in which the technological deployment is challenging and disruptive. The main claim translates these principles into practical action to create governance instruments for AI in public policy. Likewise, it creates mechanisms that assign humans in control, especially in public policy. Keeping humans abreast of AI deployment in the policy cycle is the main challenge in public policy, as the mode of knowledge created by AI shapes human behavior and government action.

5.2 Humans in the Loop and AI Governance in the Policy Cycle

Next, we consider the humans-in-the-loop approach, which reframes an automation problem from human–machine interactions (Amershi et al., Reference Amershi, Cakmak, Knox and Kulesza2014). Reframing an automation problem means that humans create controls on AI to calibrate outcomes and to involve humans in the system design. The application of the humans-in-the-loop approach perspective in public policy is that humans calibrate AI instruments by exercising supervision and controlling the flow of input data, and then evaluating the output data’s outcomes and impacts. Humans-in-the-loop approach enables an innovative way of calibrating policy instruments when AI is integrated into the policy cycle. Similar to traditional policy instruments, humans can calibrate the instruments so that they are more reliable in achieving policy objectives.

The humans-in-the-loop is an approach that instrumentalizes by governance the calibrations of outcomes and designing and redesigning the AI systems. In AI-shaped public policy, the humans-in-the-loop approach means that humans can calibrate both the design of systems and the constituency of AI-based policy instruments, on the one hand, and the outcomes of systems that reorganize policy advice. This means that the humans-in-the-loop approach meets the instrument constituencies, requiring compliance with procedures such as transparency, open training databases, or privacy and data protection. Meanwhile, the humans-in-the-loop approach also meets the systems outcomes, providing procedures that require algorithmic audits, system validation, or regulatory sandboxes in a democratic frame.

Naturally, this approach demands governance instruments, which are experimental and emerging in the policy landscape. Calibrations from human feedback are essential to develop AI in public policy, as AI systems are based on machine learning algorithms that adapt to their users and environment. Guided by the AI principles, humans can calibrate outcomes and review the entire dynamics of AI systems in the policy landscape. The humans-in-the-loop approach and AI governance emerge from AI regulations delivered recently by governments. Contexts of uncertainty and profound technological change shape emerging regulatory and governance approaches.

The European Union Artificial Intelligence Act, for example, requires that AI designers allow human control to achieve effective human oversight. Under article 14, AI systems should be designed in a path that can be overseen by people in the AI lifecycle. The objective is to formulate policies to compel AI designers to integrate the human control function as part of the safeguard against AI risks and malfunctions. The EU AI Act develops a regulatory approach based on the risks of AI. It defines a framework of rights that allows humans to request a review of decisions made by AI systems, as well as transparency and control requirements. Based on the risks, AI systems are classified as low, limited, high, and unacceptable risk. Based on this classification of an AI system, there are different requirements for different rights. Low-risk AI, such as video games or spam filters, requires AI developers to follow codes of conduct. Limited-risk AI, such as chatbots or deepfakes, implies transparency requirements. High-risk AI, such as those applied in educational processes, requires compliance and algorithmic impact assessments. Finally, unacceptable-risk AI, such as social scoring or facial recognition, is prohibited under European legislation. Each country in the European Community defines a national authority to enforce the AI Act, and a central authority coordinates all activities. The European Union prioritizes its norms or rather its market power (Ulnicane, Reference Ulnicane, Hoerber, Weber and Cabras2022).

China, in contrast, takes a state command-and-control approach to AI regulation and governance. China’s emerging regulation is based on central planning by the State Council, which outlines a timeline and strategy to advance AI development and build out AI regulations in China. In addition, the State Council has established an expert group to draft a law that aims to promote innovation in AI technology, develop a healthy AI industry, and regulate AI products and services. It contains provisions for managing AI systems, safeguarding national security and public interest, and protecting the rights and interests of individuals and organizations. In addition, there are several recommendations, such as the one from the China Academy of Information and Communications Technology on building trustworthy AI. In addition, there are specific recommendations on the development of generative AI. This emerging command-and-control perspective in China signifies the state’s ability to provide regulatory instruments based on direct command to developers, with oversight by the State Council (Liu, Zhang & Sui, Reference Liu, Zhang and Sui2024).

In the United States, there is no comprehensive legislation on AI regulation. The regulatory approach in the United States is built around self-regulation, infusing voluntary guidelines and best practices for AI systems, reflecting a cautious approach to regulation aimed at fostering innovation without imposing strict mandates. In this framework, the US government is concerned about losing competitiveness with China in the AI market, fragmenting legislative proposals into topics such as copyright, AI in education, AI robocalls, biological risks, and AI’s role in national security. In general, the perspective of self-regulation in the United States is based on the idea that users can calibrate AI outcomes, opening the technology to direct intervention by society. Regulation in the United States is far from the need for empowered authorities to impose rules on the constitution and use of AI-based systems in different social sectors (Judge, Nitzberg & Russell, Reference Judge, Nitzberg and Russell2025).

These emerging perspectives on AI regulation and governance hope to enable human control over AI, so these approaches sound uncertain in contexts of profound technological change. Humans-in-the-loop approach frames AI governance. Creating human control over AI is essential to align technologies and human objectives. Alignment issues in AI emerge when assessing the performance of an AI in normative terms. Alignment is the normative non-anomalous AI behaviour with or without the incorporation of human workers and users in these operations (Gabriel, Reference Gabriel2020). Alignment lies in the fact that an AI can respect the foundational principles of technological development, such as those disseminated by UNESCO and the OECD.

With alignment issues and emerging regulatory frameworks on the horizon, three challenges are foundational to AI governance, particularly when applied in the public policy cycle. First, information asymmetries between people and instrument constituencies. Instrument constituencies engage in AI development and deployment in a policy context shaped by information asymmetries. Second, the lack of normative consensus increases the complexity of technology control and obscures AI’s potential. Third, the government mismatches the design of effective, efficient, and legitimate means (strategies, approaches, tools, and so forth) to resolve the substantive issues, concerning the conditions of uncertainty and complexity in the AI ecosystem. Governments have failed to design policies regarding AI because of the limits on traditional approaches to law and policymaking in the digital age (Gasser & Almeida, Reference Gasser and Almeida2017).

AI governance is structured in layers, with different problems and approaches. The technical, ethical, social, and legal layers put in the AI landscape are challenges to institutionalized governance for emerging technologies deployed in the policy cycle. From here, AI governance proceeds in two ways: first, AI is a substantial instrument in the policy cycle applied in policy formulation, implementation, and evaluation; second, AI governance requires a set of procedural instruments that steer the mode of constituencies and design dynamics, adaptation to the institutional framework, alignment with ethical principles, accountability, transparency, and technical requirements. Procedural instruments, typical in AI governance, are used to indirectly but significantly affect policy processes and outcomes (Bali et al., Reference Bali, Howlett, Lewis and Ramesh2021). In the different layers that comprise AI governance, all instruments are procedural, with the objective being to steer actions and organizations to govern emerging technologies. Table 2 offers a synthesis of procedures concerning all layers of AI governance.

Table 2AI governance’s layers and procedural instruments in the policy landscape.

Source: Gasser & Almeida (Reference Gasser and Almeida2017) and own elaboration.

Table 2Long description

The table contains three columns and approximately thirty-four rows. The first column identifies the primary layers of AI governance, the second column introduces key topics within each layer, and the third column lists specific policy instruments that operationalize governance practices.

• The Ethical Layer contains topics such as ethical principles and awareness among people involved in AI systems. Related instruments include codes of ethics, guides for inclusive AI design, training programs for public servants, and national AI strategies.

• The Technical Layer consists of topics such as data governance, algorithmic accountability, and standards for system development. Instruments under this layer include policies for data collection, data sharing, and interoperability; open data initiatives; algorithmic audits; responsibility frameworks; public registries of algorithms; cybersecurity infrastructures; and standardized technical documentation and validation processes.

• The Social and Legal Layer includes broader normative and regulatory structures. Topics involve rights, legal principles, and mechanisms of system oversight. Instruments encompass regulatory sandboxes for experimentation, periodic assessments of AI systems, privacy and data protection frameworks, and intellectual property mechanisms.

AI governance in public policy is about procedures concerning ethical, technical, and regulatory requirements to assign humans in the loop with technologies that make decisions and conduct tasks in the policy process. Thus, AI governance in public policy is about defining who can do what when, where, and how with requirements to action to align AI with government objectives, democracy, and public values (Korinek & Balwit, Reference Korinek, Balwit, Bullock, Chen and Himmelreich2023; Innerarity, Reference Innerarity2024). Integrating the main layers of AI governance is essential to align technology development and social good (Wirtz, Weyerer & Sturm, Reference Wirtz, Weyerer and Sturm2020). AI governance’s procedural logic is asserted by the necessity to maintain humans in the loop with AI applied in the policy process. The main objective is designing an institutional framework that integrates ethical, technical, and social and regulatory layers (Gasser & Almeida, Reference Gasser and Almeida2017), delivering procedural instruments that create a logic of appropriateness for AI development (Filgueiras, Reference Filgueiras2022b). Between the procedural instruments for AI governance, Table 2 enumerates the main instruments applied to AI development and deployment. Data governance procedural instruments, such as interoperability, data sharing, data collection, and qualification are essential to AI in public policy. In the same vein, regulatory procedural instruments such as sandboxes or privacy and protection are essential to align AI with social goods. Finally, defining ethical principles and codes of conduct for AI designers and developers is essential to create a public value perspective.

The challenge of AI governance is that the procedural instruments are beyond the government. The AI governance landscape requires approaches that are beyond governments’ power to regulate and define procedures. Another alignment in AI governance is a global digital ecosystem with local needs and institutions. For example, diverse AI are designed and deployed in cloud systems, without government capacities to regulate and impose control (Filgueiras & Almeida, Reference Filgueiras and Almeida2021). In AI governance, the institutional framework to define rules and procedures to develop and deploy AI systems is labeled in a multistakeholder perspective. From this perspective, the objective is to couple global and local institutions with emerging practices to govern AI. The challenge is that humans supervise decisions and tasks performed by AI and then create a perspective that evolves and creates knowledge from human feedback framed in data, controls, and regulations. Yet the humans in the loop approach involves the micro and macro institutions (Table 3).

Table 3Layers and levels of AI governance.

Source: own elaboration.

Table 3Long description

The table is organized into four columns and two rows, where the first column identifies the governance level, and the remaining columns represent the three layers. 1. At the Micro Level: In the Ethical Layer, principles directly shape how AI systems are developed and used. In the Technical Layer, interaction occurs through continuous feedback between humans and algorithmic systems. In the Social and Regulatory Layer, humans retain supervisory responsibility to oversee and intervene in system operation. 2. At the Macro Level: In the Ethical Layer, governance occurs through broad frameworks that define AI goals and responsibilities. In the Technical Layer, governance is implemented through standards, protocols, and technical rules. In the Social and Regulatory Layer, governance is expressed through norms, rights, and legal protections.

The macro level concerns more global and systemic norms and practices, which require multistakeholder constructions to define technical standards, internationally shared norms and rights, and frameworks for governance. At the micro level, AI governance deals with institutionalizing principles and how they translate into practical action. Similarly, at the micro level, it deals with how reciprocal feedback between humans and machines will be processed and how humans supervise systems’ work. The micro level deals with institutionalizing practices related to compliance with fundamental procedures that need to be performed by developers, companies, governments, and other organizations.

In the human–machine interactions, AI governance is essential to policy process. The challenge is how to govern disruptive technologies with large uncertainty and ambiguities. AI governance requires a policy perspective based on procedural instruments to align technological development in the policy process and outcomes in terms of policy knowledge, practices, ideas, analyses, and work. This challenge is strange – to create a policy environment adapted to disruptive AI technologies applied in policy process. Policymakers and bureaucrats have limited knowledge concerning how it works and why, as well as what the possible applications and consequences are for its deployment. Furthermore, policymakers and bureaucrats have a policy environment shaped by uncertainty and structural power dynamics framed by big techs (Taeihagh, Ramesh & Howlett, Reference Taeihagh, Ramesh and Howlett2021). AI in policy cycle and structural power dynamics creates a governance context based on uncertainty, making it difficult for governments to design policies and regulatory perspectives.