The systematic entry of algorithmic models into public decision making processes marks a structural transformation of contemporary governance. This is not a simple improvement of decision support tools, but a shift in the very principle through which power is exercised. In post AI economies and institutions, decisions increasingly emerge as the result of computational procedures rather than as the outcome of political confrontation among normative alternatives. This transformation directly affects the relationship between economy, institutions and democratic legitimacy.
Predictive models and simulation systems are introduced with the declared aim of reducing uncertainty and improving the effectiveness of public action. Ex ante analysis of fiscal policies, algorithms for allocating health resources, risk assessment systems in social security or public safety are now widespread across many advanced administrations. According to the OECD, more than 70 percent of member countries currently use data analytics and machine learning tools in at least one area of public policy, with particularly rapid growth in healthcare, welfare and urban management. This quantitative expansion produces qualitative effects on the very form of decision making.
When policy becomes algorithmic output, the decision making process undergoes a profound reconfiguration. Decisions are no longer constructed through a dialectic among worldviews, social interests and institutional mediation, but emerge as optimized solutions relative to parameters defined upstream. The political moment thus shifts from visible deliberation to a phase less exposed to public scrutiny, in which relevant variables, assigned weights and acceptability thresholds are established. Power manifests not so much in the final choice, but in the definition of the model that renders certain choices conceivable and others invisible.
This shift has direct consequences for the quality of democratic debate. If a decision appears as the “necessary” result of a model, the space for dissent tends to shrink. Contesting a policy means contesting data, statistical assumptions and technical architectures that require specialized expertise and are rarely accessible to public debate. The language of politics is progressively replaced by that of performance and predictive accuracy. Studies by the MIT Governance Lab show that, in contexts where decisions are highly automated, parliamentary debate tends to focus on procedural aspects, leaving value based questions in the background.
The standardization of decisions further reinforces this dynamic. Algorithmic models function more effectively the more they are replicable and scalable. This produces a tendency toward homogenization of policy responses, even in deeply different social and economic contexts. Welfare policies, criteria for access to public services or systems for prioritization are applied according to uniform schemes, reducing space for contextual judgment. Social complexity is compressed into formal structures that privilege what is measurable, penalizing qualitative dimensions that are difficult to translate into data.
This process feeds a form of structural depoliticization. Governing through models means shifting conflict from the level of values to that of procedures. Decisions are no longer justified as expressions of political choice, but as consequences of technical optimization. Power thus assumes an indirect form that does not impose, but normalizes. In economic terms, the algorithm acts as a coordination mechanism that reduces the range of options considered rational, producing effects analogous to those observed in highly standardized markets.
The economic implications of this transformation are significant. Algorithmic governance tends to privilege criteria of allocative efficiency and risk reduction, often at the expense of redistributive or long term objectives. According to the World Economic Forum, the adoption of predictive systems in public spending has led in several countries to operational cost reductions between 10 and 20 percent, but also to increased inequalities in access to services where models did not incorporate explicit social correctives. Calculation, in the absence of clear political orientation, tends to reproduce and amplify existing structures.
Max Weber’s reflection on rationalization offers an interpretive key that remains relevant. The progressive organization of the world according to criteria of calculation increases efficiency, but risks turning into an “iron cage” that constrains political action. In the era of artificial intelligence, this cage takes on a digital form, more flexible and adaptive, yet no less binding. The risk is not explicit authoritarianism, but a governance that renounces conflict and choice in the name of measurable performance.
The parallel with the corporate world confirms this ambivalence. Organizations that have fully delegated strategic decisions to predictive systems have often achieved short term efficiency gains, but shown reduced capacity to adapt to unforeseen shocks. McKinsey analyses indicate that firms with a balanced use of models and managerial judgment display more resilient performance than those adopting a purely data driven approach. Models excel at projecting the past into the future, but struggle to grasp radical discontinuities.
The central question is therefore not whether to use artificial intelligence in policy making, but how to redefine the relationship between calculation and decision. Models can inform, simulate and make visible consequences that would otherwise remain opaque. They cannot replace political judgment, which remains a normative act linked to values, priorities and responsibility. Governing through models ultimately means redefining who decides and on what basis. If this redefinition occurs without explicit political accountability, power does not disappear, but retreats into less visible spaces, harder to contest and more fragile in terms of legitimacy.
Global AI Observatory