Public-sector AI could deepen data power and opacity in Kazakhstan

Kazakhstan’s rapid move to embed artificial intelligence (AI) into public administration could create hidden risks that are shaped less by faulty algorithms than by legal design, centralized data systems and weak accountability safeguards, researchers warn in a new study published in Frontiers in Artificial Intelligence on the country’s expanding digital state.

The study, titled “Potential negative effects of artificial intelligence in Kazakhstan’s public sector: an analysis of hidden risks,” examines Kazakhstan’s AI law, its National AI Platform and major digital governance projects, including Smart Data Ukimet and the Digital Family Card, while also drawing on a secondary qualitative analysis of 53 interviews with civil servants.

AI governance risks emerge before harms become visible

According to the research, the main risk in Kazakhstan’s public-sector AI agenda is not simply whether individual systems are accurate or efficient. The deeper concern is that risks may be built into the institutional structure before large-scale harms become easy to measure. The authors describe these as hidden or latent risks because they are embedded in legal definitions, centralized infrastructures and administrative routines rather than appearing as one clear technical failure.

Kazakhstan has moved quickly to formalize its AI strategy. The government approved the Concept for the Development of Artificial Intelligence for 2024–2029, created a Ministry of Artificial Intelligence and Digital Development, and adopted a dedicated Law “On Artificial Intelligence,” which entered into force in January 2026. The law introduces a risk-based logic and sets out principles such as legality, fairness, transparency, explainability, human well-being, data protection and confidentiality.

However, the study finds that these principles remain underdeveloped at the operational level. The researchers argue that the legal framework defines AI largely as a tool while leaving unresolved how citizens can contest AI-mediated decisions, how independent audits will work, who bears responsibility when harms occur, and how regulatory oversight will be separated from promotion of AI deployment.

AI in the public sector is not just another software upgrade. In government, AI can shape eligibility checks, service access, risk scoring, welfare classification, tax interpretation, appeals processing and citizen profiling. Once embedded in public administration, algorithmic systems can influence how rights, benefits and duties are allocated.

The authors point out that Kazakhstan’s case is especially important because AI development is taking place within a centralized digital governance environment. The state already operates integrated data systems, proactive service tools and platforms that consolidate large volumes of administrative data. These systems may improve efficiency, but they also create conditions for concentrated power, limited contestability and algorithmic decision-making that citizens may find difficult to understand or challenge.

The study avoids claiming that Kazakhstan is pursuing a deliberate strategy of AI-enabled social control. Its argument is more structural: AI systems can shift power and reduce accountability even without explicit coercive intent if they are deployed through centralized infrastructures and weak oversight arrangements.

Centralized data systems could deepen bias and weaken accountability

A major concern is the role of Kazakhstan’s centralized data infrastructures. Smart Data Ukimet consolidates and analyzes data from more than 120 government databases, while the Digital Family Card aggregates registry data to assess household welfare and trigger proactive social services and notifications. The emerging system also includes the National AI Platform, sovereign large language models and AI agents for e-government and administrative work.

The researchers argue that these systems can institutionalize risk because they shape how citizens become visible to the state. A welfare profile, household score or automated eligibility signal may appear neutral, but it depends on the data selected, the categories used and the thresholds applied. If those classifications are flawed, outdated or too narrow, vulnerable citizens may be excluded without knowing why.

The Digital Family Card is presented as a key example. It aims to identify families in difficult circumstances and connect them to support. But the study warns that multidimensional scoring can produce false negatives when households do not cross formal thresholds despite informal employment, debt, illness, unstable housing or other forms of vulnerability. In such cases, the absence of a service notification may look like an ordinary administrative outcome, making the decision hard to detect and contest.

The study also warns about cross-context use of household data. Data gathered for social support could later influence inspections, creditworthiness assessments, risk targeting, behavioral profiling or other forms of state interaction unless strict limits are imposed. That creates the possibility of algorithmic stratification, where citizens or households become fixed in data-defined categories that shape future treatment.

The risk extends to sovereign large language models and AI assistants. Kazakhstan’s sovereign-model strategy includes systems designed to operate in Kazakh, Russian and other languages and to support public services, legal information and administrative workflows. The authors argue that such tools may privilege official or institutionally curated interpretations if they rely heavily on state documents, national websites and approved data sources.

This does not mean the models are necessarily designed to suppress alternative views. The concern is that a single official AI interface may narrow the range of explanations citizens receive. If citizens increasingly rely on state-integrated AI agents for legal and public-service information, alternative interpretations, minority positions and critical viewpoints may become less visible.

The study identifies accountability as another weak point. AI-mediated public decisions involve developers, platform operators, user agencies and citizens. Without detailed rules for audit, logging, model version tracking, human intervention and appeal, responsibility may become fragmented. Developers may point to technical compliance, agencies may point to system outputs, and citizens may be left unable to reconstruct how a decision was made.

Civil servants interviewed for the study reinforced this concern. Many described digital systems as useful but also as hierarchical, opaque, fragile and dependent on centralized platforms. Respondents raised concerns about formalism, system overload, data leakage, limited flexibility and weak trust in internal contestation. These findings suggest that AI is entering an administrative culture where responsibility is already difficult to challenge and where digital systems may be treated as authoritative even when their logic is not fully understood.

Researchers call for audits, appeal rights and stronger human oversight

The study sets out five major groups of hidden risk: political-legal and institutional risks; data and technology risks; organizational and managerial risks; explainability and opacity risks; and market and environmental risks.

• Political and legal risks arise from the way authority is distributed. The researchers warn that the same institutional center may hold regulatory, operational and promotional roles, creating potential conflicts between fast AI deployment and strong accountability. If the body promoting AI also shapes rules and controls infrastructure, external oversight may be weakened.
• Data and technology risks come from centralized databases, AI platforms and proactive service systems. These tools can improve coordination, but they can also scale errors across agencies, reinforce bias, and create detailed citizen profiles that are hard to challenge.
• Organizational risks include deskilling and dependence on AI interfaces. Tools such as eGov AI assistants, tax assistants, legal support systems and agency-specific AI agents may help civil servants process information faster. But over time, the study warns, workers may rely more on prompt formulation and system outputs than on independent legal, social or administrative judgment. This could weaken professional expertise, especially among younger civil servants entering an AI-heavy workplace.
• Explainability risks arise because transparency can become symbolic. A label stating that content or a decision involved AI does not explain how AI shaped the outcome, which data were used, which assumptions mattered, or whether an alternative result was possible. The study argues that meaningful explainability must include system-level and case-specific information, especially in decisions affecting rights, benefits and public obligations.
• Market and environmental risks are less visible in official debate but remain important. The National AI Platform could become a dominant gateway to government data and computing resources. Without fair access rules, tariff neutrality and safeguards against preferential treatment, independent developers, researchers and smaller market actors may face disadvantages. At the same time, large-scale AI infrastructure raises energy and resource concerns that Kazakhstan’s AI regulation does not yet fully address.

The authors propose several policy responses. High-risk AI systems should be subject to independent algorithmic and data audits, with attention to bias, robustness and legal compliance. Citizens should have a clear right to appeal algorithmically mediated decisions, especially in social protection, taxation and justice. Authorities should be required to provide intelligible explanations and to review contested cases without relying on the same AI system.

The study also calls for separating the roles of regulator, operator and technology promoter. Independent bodies responsible for data protection, human rights and competition should have stronger roles in oversight. AI systems should include lifecycle logging so that model versions, datasets, parameters and human interventions can be reviewed by courts and regulators.

For proactive tools such as the Digital Family Card, the researchers recommend human-rights impact assessments, analysis of classification errors and strict limits on cross-context use of household profiles. They also call for training civil servants to use AI critically, not passively, and for mandatory human involvement in decisions affecting rights and freedoms.