International peace and security frameworks on artificial intelligence — including United Nations General Assembly (UNGA) resolutions 79/239 and 80/58, the Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, and its subsequent Blueprint for Action — have developed a growing body of norms governing AI in the military domain, with particular attention to autonomous weapons systems, reflecting broader efforts to preserve meaningful human control, accountability, and traceability across decision-making and related operational contexts. This milestone is legitimate and necessary. But it tends to produce a structural imbalance in global AI governance: accountability expectations for military applications continue to evolve, while civilian AI systems (non-military), which already mediate access to rights and services, remain comparatively overlooked.
This imbalance is not merely conjectural. The U.N. Secretary-General’s High-level Advisory Body on Artificial Intelligence warned in 2024 of a systemic accountability deficit in the deployment of advanced AI, including AI for civilian purposes, and cautioned that automated decision-making is advancing faster than existing mechanisms of oversight and responsibility can absorb. The gap is particularly consequential in fragile and post-conflict contexts, where limited institutional capacity amplifies the destabilization risks associated with opaque algorithmic decision-making — and where eroded public trust can compound the fragility that governance systems are meant to address.
From Military Control to Civilian Governance
International governance norms on autonomous weapons have increasingly converged around a common principle: decisions with significant consequences for human life and security should remain subject to meaningful human responsibility, accountability, and oversight, as reflected in UNGA resolutions 79/239 and 80/58, which emphasize human-centric, accountable, safe, secure, and trustworthy AI in the military domain. That standard, however, has not been extended with comparable urgency to civilian AI systems — used in welfare administration, regulatory delivery, law enforcement, and risk management — that are increasingly embedded in core state functions, shaping access to public goods through automated decision-making processes that remain difficult to contest.
This shift is not simply one of technological adoption. More specifically, it represents a deeper reconfiguration of how state authority is exercised. When automated systems mediate access to social protection, determine eligibility thresholds, or classify populations for intervention, they become instruments of distributive power — with consequences for legal rights and democratic accountability that extend well beyond administrative efficiency.
Ukraine’s Diia ecosystem exemplifies both the potential of autonomous or agentic systems in sustaining delivery capacity and the governance risks that accompany this transition. Diia is a digital government services platform that integrates public services and identity documents into a mobile application, enabling the continuation of public administration under wartime conditions, including access to social benefits, identity documentation, and essential services during severe institutional disruption. Yet its wartime deployment also exposes the vulnerabilities and challenges that arise when core state functions depend on digital infrastructures operating under stress: maintaining service continuity, protecting data integrity and cyber resilience, ensuring interoperability across agencies, and preserving public trust when the state itself is under attack. These are not edge-case concerns; they are the conditions under which welfare-critical AI is increasingly expected to function.
The global regulatory picture reflects a deepening mismatch. The Global Index on Responsible AI, assessing 138 countries, reveals that 67% of nations scored 25 points or less out of 100 on responsible AI governance, leaving nearly six billion people without adequate human rights protections and accountability safeguards in the age of AI (see p. 57). At the same time, governments in two-thirds of OECD countries now use AI in government services. The divergence between rapid deployment and the slower development of accountability mechanisms is widening — and, in weaker institutional contexts, that divergence is not cost-free.
Sisbén IV and Algorithmic Social Classification
Colombia’s post-2016 social policy architecture offers a concrete illustration of how AI-mediated systems can quietly reshape the exercise of distributive state authority in post-conflict settings — and why that matters for the legitimacy of peace implementation itself.
Following the 2016 peace agreement, which placed rural development and social inclusion at its center, the System for the Identification of Potential Beneficiaries of Social Programs (Sisbén) became the central infrastructure for allocating social assistance and repairing social-systemic injustices. Socioeconomic classification was transformed into a primary gateway to welfare provision, with far-reaching implications for the rural and conflict-affected populations the agreement was intended to reach.
Over successive iterations, Sisbén evolved from a system relying predominantly on observable living conditions — collected through household surveys — to one incorporating increasingly sophisticated forms of statistical inference drawn from integrated administrative data. Following its modernization, Sisbén IV introduced more granular classifications of vulnerability and “probabilistic” assessments of socioeconomic status based on predictive algorithms and alternative data sources — such as credit histories, loan information, financial activity, and telecommunications data — to refine social protection targeting. This was not an abrupt rupture, but rather an intensification of existing proxy means-testing practices, accelerated by the growing availability of administrative datasets.
The cumulative effect, however, was a significant epistemic shift: from descriptive assessments of poverty to increasingly inferential forms of social classification, in which eligibility for basic support is mediated by statistical representations that are partially transparent and difficult for beneficiaries to contest. Administrative appeals mechanisms remain formally available, but asymmetries in information and limited visibility into classification criteria constrain what meaningful participation actually looks like for affected citizens.
Government justifications for limiting transparency have typically emphasized the risk of manipulation of classification criteria and strategic gaming of the system by potential beneficiaries. This concern was not without merit: CONPES 3877, the 2016 policy document laying the groundwork for Sisbén IV, identified manipulation and misreporting of information as a key challenge undermining the integrity and effectiveness of the targeting system.
This governance model was further consolidated during the COVID-19 pandemic. In 2020, Colombia established the Social Registry of Households (Registro Social de Hogares, RSH), an information system administered by the National Planning Department (DNP) to support targeting, policy design, and the monitoring of household living conditions through the integration of public and private datasets. The Colombian Ingreso Solidario emergency cash transfer program relied on this expanding data architecture to identify eligible beneficiaries at speed. That approach enabled rapid disbursement — a genuine operational achievement under crisis conditions — but it also deepened reliance on a governance architecture built around extensive data aggregation and eligibility criteria accessible only incompletely to public scrutiny.
Opacity of scoring criteria was, in part, a response to a real integrity problem. But that justification has also served to normalize a broader pattern in which algorithmic design choices and eligibility thresholds operate with limited independent oversight — and in which the reduction of contestability is treated as an acceptable feature rather than a governance failure. Colombia’s longer institutional history, including the 2009–2011 Chuzadas DAS surveillance scandal — in which intelligence services conducted illegal monitoring of journalists, judges, and opposition figures — offers a reminder that state data practices have not always commanded public trust, and that confidence must be earned through accountability mechanisms and participatory frameworks, not assumed.
Conclusion
The Colombian case is instructive not only because it represents exceptional misgovernance, but also because it illustrates structural dynamics that recur across many contexts in which welfare-critical AI is deployed: the gradual reduction of contestability, the normalization of inferential classification as a basis for rights access, and the widening gap between deployment pace and accountability infrastructure.
In Colombia, these dynamics carry particular weight. The social policy instruments mediated by algorithmic classification are not politically neutral; they are integral to the implementation of a peace agreement whose legitimacy also depends on delivering tangible benefits and resources to conflict-affected communities. When those groups cannot meaningfully understand, challenge, participate, or appeal the decisions that determine their access to rights, the accountability deficit is not merely an administrative problem. It is a political one; with implications for the perceived legitimacy of the peacebuilding process.
The international community has invested significant normative energy in ensuring that autonomous weapons systems remain subject to meaningful human oversight. That principle — that consequential decisions require transparency, traceability, and the possibility of challenge — should not remain confined to the military domain. It applies equally to the algorithmic systems through which states exercise authority in civilian domains. Extending accountability norms from military to civilian AI governance systems is not a theoretical exercise. In fragile, post-conflict, and reconstruction settings undergoing digital transformation, the governance of civilian AI systems is already a defining condition for sustaining peace.
