Artificial intelligence now intervenes in the least visible yet most decisive core of the modern state: administrative rationality. It does not act as a simple tool for efficiency gains, nor as a neutral support for public action. Rather, it affects the ways in which public order is constructed, maintained and justified, redefining the relationship between decision, computation and legitimation. In this sense, AI does not merely accelerate the machinery of the state, but reorients its operational principles, shifting the center of gravity of governance from procedural logics to predictive logics.

Classical administrative rationality was founded on general categories, abstract rules and formalized decision sequences. Its strength lay in predictability and impartiality, elements central to the legitimation of the rule of law. Artificial intelligence systems introduce a rationality of a different nature. They operate through statistical correlations, probabilistic assessments and continuous optimization. They do not simply apply norms, but anticipate behaviors, estimate risks and assign dynamic priorities. Public action thus shifts from an ex post model, based on rule application, to an ex ante model, based on prediction.

This transition is already observable in numerous domains. According to OECD data, more than 70 percent of governments in advanced economies use predictive analytics or decision support systems in sectors such as healthcare, welfare, security, taxation and urban management. In healthcare, algorithmic triage models are used to allocate hospital resources; in security, risk assessment systems orient intervention priorities; in social policies, scoring algorithms determine access to benefits and subsidies. In all these cases, the final decision formally remains public, but its space of possibility is delimited upstream by models that define what is efficient, sustainable or statistically justifiable.

What follows is a structural transformation of administration. Decision making power no longer resides solely in signed acts or final procedures, but in computational infrastructures that orient the entire process. The overcoming of twentieth century bureaucracy does not occur through the abolition of rules, but through their progressive subordination to adaptive systems that learn from the past and recalibrate over time. The office as the site of impersonal normativity gives way to diffuse decision networks, often invisible, embedded in the information systems that traverse the public apparatus. Administrative power is thus exercised in a less concentrated and more opaque manner, through models that precede choice and condition its form.

Efficiency is the value that justifies this transformation. AI promises cost reduction, rapid response and greater coherence in decisions. However, administrative efficiency is never neutral. Every acceleration implies a selection of what matters and what can be neglected. The reduction of decision times often entails a compression of the space for reflection, less attention to exceptions and a standardization of borderline cases. World Bank reports show that the adoption of automated systems improves average efficiency, but increases the risk of systematic exclusion for groups that do not fit the dominant patterns of historical data.

The central issue thus becomes responsibility. When a decision emerges from the interaction between datasets, models and risk thresholds, the attribution of political and moral responsibility fragments. Is responsibility borne by the administrator who approves the system, the technician who designs it, the institution that adopts it, or the algorithm itself as an operational entity? This dispersion of responsibility produces a form of institutional opacity that does not derive from secrecy, but from complexity. Even formally transparent systems often prove indecipherable to those without advanced technical expertise, reducing the possibility of effective democratic oversight.

The consequences for legitimacy are significant. Democratic legitimacy is not based solely on outcomes, but on the intelligibility and contestability of decision making processes. When public governance relies on complex models, political conflict tends to turn into a technical dispute. What is debated is no longer the choice itself, but the accuracy of the model; no longer the objective, but the quality of the data. World Economic Forum studies indicate that the perception of automated decisions reduces trust in institutions, even when such decisions produce statistically better results. Authority thus appears as a function of impersonal computation, rather than as the expression of responsible political judgment.

This dynamic recalls Hannah Arendt’s reflections on the risk of a rationality separated from judgment. Arendt distinguished efficient doing from responsible acting, noting how the absolutization of the former could empty the latter of meaning. Transposed to algorithmic administration, this distinction highlights a structural risk: that the state becomes extremely capable of optimization, but progressively incapable of explaining and justifying its choices in political and moral terms. Governance is strengthened on the technical plane, but weakened on that of legitimation.

The parallel with the private sector is instructive. Firms that have adopted advanced decision systems have experienced significant performance benefits, but also new vulnerabilities. Analyses by Harvard Business School show that excessive reliance on predictive models increases the risk of systemic errors when embedded assumptions are not questioned. More mature leaders have understood that automation does not eliminate responsibility, but makes it more demanding. Governing models requires critical capacity, understanding of limits and explicit assumption of the consequences of decisions. This lesson appears only partially absorbed within public institutions.

Artificial intelligence, in this sense, does not automatically strengthen the state. It tests its form. It introduces a powerful rationality that can become detached from judgment, a computational capacity that risks turning into a principle of public order in itself. The institutional challenge does not lie in choosing between bureaucracy and algorithm, but in constructing a balance in which predictive power does not erase the ethical and political dimension of public action. In a context in which computation tends to precede decision, the function of the state remains that of making choices explicit, assuming responsibility for their consequences and preserving the possibility of understanding why a decision is taken, and not only how it is optimized.

Global AI Observatory