The integration of artificial intelligence into public decision making processes profoundly alters the functioning of contemporary institutions. This is not a simple administrative innovation, but a systemic transformation that affects the way power is exercised, justified and made visible. Institutional decision making, historically conceived as an act of mediation among conflicting interests, regulatory constraints and political responsibility, is progressively anticipated by computational systems that select options, simulate outcomes and delimit the field of the possible even before deliberation takes place.
Algorithmic models enter policy making as support tools, with the promise of reducing uncertainty and improving the quality of decisions. This promise is based on a real capacity to process volumes of data that cannot be managed by traditional human structures. According to the OECD, more than 65 percent of governments in advanced economies already use predictive analytics or decision support systems in areas such as welfare, healthcare, taxation, security and public resource management. However, prediction is never a neutral act. Every model incorporates assumptions, priorities and relevance criteria that reflect a specific vision of the problem to be solved. In this sense, the algorithm does not merely inform the decision, but filters it, making certain options visible while rendering others marginal or impracticable.
The automation of administrative procedures accentuates this dynamic. Scoring systems for access to social benefits, risk assessment models for public credit or crime prevention, and algorithms for allocating healthcare resources operate according to logics of efficiency and internal coherence. Formally, the decision remains attributed to a public authority; substantively, the decision making perimeter is preconfigured. Analyses conducted by the European Commission show that, in several Member States, officials tend to adhere to algorithmic recommendations even in the presence of doubts, due to the difficulty of justifying divergent choices in a highly measured and auditable context. Power does not disappear, but shifts upstream, into system design, data selection and parameter definition.
This redistribution of power generates a new form of institutional opacity. This does not necessarily involve intentional secrecy, but structural complexity. Even when systems are formally transparent, their logic remains difficult to understand for those who do not possess advanced technical expertise. World Bank reports highlight that most public administrations depend on external providers for the development and maintenance of complex algorithmic systems, reducing internal capacity for control and audit. In this context, political responsibility tends to dilute: the decision is public, but its genesis is technical, fragmented and difficult to reconstruct.
The democratic implications are significant. Institutional legitimacy derives not only from policy effectiveness, but from the possibility of understanding decision making processes and contesting their outcomes. When algorithmic mediation becomes central, political conflict risks being transformed into a technical dispute. What is debated is no longer the objective, but the accuracy of the model; no longer the choice, but the quality of the data. World Economic Forum studies show that the perception of “automatic” decisions reduces citizens’ trust in institutions, even when such decisions produce statistically better results. The rationality of computation tends to present itself as necessity, naturalizing choices that remain deeply political.
This transformation recalls Max Weber’s reflections on bureaucratic rationality. Weber identified formal legal rationality as a source of efficiency, but also as a risk of growing separation between action and meaning. The algorithm represents a new phase of this rationalization, in which decision making is translated into a problem of continuous optimization. The “iron cage” is no longer made only of rules and procedures, but of models that learn from the past and project the future according to probabilistic criteria. The risk is not inefficiency, but a form of hyper efficiency that reduces the space for judgment and responsibility.
In the private sector, similar dynamics have already clearly emerged. Analyses by Harvard Business School show that the adoption of automated decision systems improves average performance, but increases vulnerability to systemic errors when the assumptions embedded in the models are not questioned. More mature firms have understood that models are instruments of orientation, not substitutes for judgment. Leadership does not consist in following algorithmic output, but in knowing how to interrogate it, limit it and integrate it into a broader vision. This awareness still appears fragile within public institutions, where pressure for efficiency and technical neutrality can encourage excessive delegation.
Algorithmic mediation does not eliminate moral dilemmas, but makes them less visible. Every public decision implies value choices: whom to include, whom to exclude, which risks to accept, which priorities to privilege. When such choices are embedded in models, they lose visibility without losing effectiveness. Power is then exercised as the capacity to define the frames within which a decision appears rational and inevitable. In this sense, the algorithm becomes a filter of power, rather than a simple instrument of power.
The central issue does not concern the use or rejection of algorithms in policy making, but the institutional capacity to govern their mediation. Making model assumptions explicit, strengthening internal competences, and preserving spaces of discretion and responsibility represent essential conditions to prevent automation from translating into a silent subtraction of democratic power. In a context in which prediction tends to precede choice, the role of institutions remains that of assuming the weight of decision making, recognizing that computation can orient the future, but cannot replace judgment about the present.
Global AI Observatory