Accountability constitutes one of the least visible yet most decisive pillars of institutional legitimacy. It does not coincide with mere procedural compliance, but with the effective possibility of reconstructing the link between decision, responsible subject and resulting consequences. The entry of artificial intelligence into public decision making processes places precisely this link under pressure, not by introducing a new factor of arbitrariness, but by profoundly redefining chains of action. The decision does not disappear, but becomes distributed across technical, organizational and computational sequences that make it increasingly difficult to identify who decides, on what basis and with what margins of choice.

The difficulty of attributing responsibility represents the first structural paradox of algorithmic governance. When an administrative decision is the result of a predictive system trained on large volumes of data, integrated into automated procedures and validated through statistical thresholds, the figure of the traditional decision maker loses centrality. The official, manager or political decision maker no longer appears as the full author of the choice, but as the terminal point of a process that takes shape elsewhere. According to a 2023 OECD study on the use of AI in public administration, in more than 50 percent of the cases analyzed, automated or semi automated decisions lack a clearly documented chain of responsibility, making any form of ex post contestation complex.

In this context, algorithmic models assume a role that goes beyond that of simple tools. They orient priorities, delimit alternatives and exclude options even before the decision making process enters its visible phase. While not being moral subjects, they exercise a form of indirect agency that structures institutional behavior. Similar phenomena have been observed in the private sector, where capital allocation or risk management algorithms have progressively replaced managerial judgment in numerous strategic decisions. Analyses conducted by Harvard Business School show that, in large data driven organizations, more than 70 percent of relevant operational decisions are now decisively influenced by automated systems, with a consequent reduction in the traceability of individual responsibilities.

The crisis of traditional control mechanisms arises precisely from this transformation. Accountability architectures developed in the twentieth century presuppose discrete decisions, motivated and attributable to identifiable subjects. Audits, oversight committees, judicial appeals and parliamentary controls are designed to evaluate explicit and documentable acts. Artificial intelligence systems, by contrast, often operate as functional black boxes: they produce statistically reliable outputs, but are difficult to explain in detail. Control thus tends to shift from the content of the decision to trust in the system that generated it, transforming accountability into a verification of technical compliance rather than an instrument of substantive responsibility.

This shift has significant consequences. When control becomes systemic rather than case specific, the possibility of correcting structural errors is reduced. What is monitored is that the model functions according to technical specifications, not that it produces fair outcomes or outcomes consistent with broader social objectives. World Bank reports on the use of AI in welfare programs show that automated scoring systems have improved efficiency in resource allocation, but have also produced systematic exclusions of categories inadequately represented in historical data. In these cases, responsibility cannot be attributed to a single erroneous decision, but to an overall configuration that escapes traditional control mechanisms.

From this emerges the need to rethink the institutional architectures of responsibility. It is not sufficient to introduce ethical codes or guidelines on the use of AI. It is necessary to redefine the perimeter of accountability, shifting it upstream in the decision making process. Responsibility must include model design, data selection, definition of optimization objectives and the choice of success metrics. Some international experiences offer indications in this direction. In Canada and the Netherlands, for example, algorithmic impact assessment obligations have been introduced for automated public decisions, with the aim of making assumptions and risks explicit before system implementation.

This requirement refers to a broader reflection on the relationship between technique and power. Max Weber described modern rationality as an “iron cage”, efficient but potentially dehumanizing if separated from an ethics of responsibility. In the era of artificial intelligence, that cage takes the form of mathematical models and computational infrastructures. The risk is not arbitrariness, but systemic irresponsibility, in which decisions are not the result of conscious choice, but the emergent outcome of processes that no actor fully controls.

The parallel with the corporate world reinforces this interpretation. More mature organizations have understood that fully delegating decisions to systems is equivalent to renouncing leadership. Models can support, but not replace, strategic responsibility. When this occurs, the organization loses the ability to explain its choices and to correct them in a conscious manner. Similarly, a state that abdicates its capacity to assume decisions that are comprehensible and contestable risks eroding the capital of trust on which institutional legitimacy is founded.

Artificial intelligence therefore requires public accountability to be rethought not as ex post attribution of blame, but as ex ante design of responsibility. This means building institutions capable of coexisting with intelligent systems without being governed by them in an opaque manner. In a mature democracy, the problem does not lie in the complexity of decisions, but in the growing impossibility of tracing who is responsible for them. It is within this fracture, between the power of computation and political responsibility, that one of the most decisive institutional challenges of the post AI economy is played out.

Global AI Observatory