SAMAI 2, the second phase of the SAMAI initiative, was launched in late January 2026 at ICAN 2026 with 11 ministries to train public sector employees in AI. Two-thirds of Saudi public sector employees report daily use of AI tools, according to the Public Sector AI Adoption Index 2026 produced by Public First for the Center for Data Innovation, with support from Google. This figure places Saudi Arabia among the best-ranked countries in this declarative index of AI adoption in the public sector. The speed of deployment and the quality of governance remain to be distinguished.

The Essential Points

  • SAMAI 2 is an AI skills-building initiative launched in late January 2026 with 11 ministries. The program was launched in an environment that already included frameworks for adoption, ethics, and data protection.
  • Two-thirds of surveyed Saudi public sector employees report daily use of AI tools according to Public First; this is a self-reported figure, not audited by an independent third party.
  • The Hexagon center, with a planned total capacity of 480 MW, was under construction as of January 1, 2026; according to the SPA, it is intended to become the largest government data center by MW capacity.
  • The speed of deployment and the existence of remedies for citizens are two independent variables; their articulation determines the governance architecture for a generation.
  • The Saudi model empirically tests whether technological capacity can precede democratic deliberation without paying the price in the long term.

A Massive Deployment, Infrastructure Without Parallel

The figure first strikes by its scale. Two-thirds report daily use, while nearly half report using AI for more than a year. To measure the speed: most European administrations struggle to reach 20% adoption on far simpler tools after several years of deployment, as shown by the persistent gap between robot equipment and humans trained to manage them.

Behind this figure lies infrastructure under construction. The Hexagon data center is planned to have a total capacity of 480 megawatts. According to the SPA, Hexagon is a project intended to become the largest government data center by MW capacity. The Hexagon project is not cloud computing leased from a private provider: it is state-owned infrastructure designed to support digital services in the Kingdom. This choice of architecture, prioritizing computational sovereignty, says something precise about the regime’s doctrine: governmental AI is treated as a strategic infrastructure in the same way as the electrical grid or petroleum reserves.

SAMAI 2 covers very concrete functions: automated processing of administrative requests, predictive analysis in health and education services, supervision of government supply chains, assistance to employees in the ministries of finance, labor, and interior. The eleven ministries involved cover several points of contact between the state and Saudi citizens.

The Limits of the Two-Thirds Figure

The self-reported adoption rate deserves cautious reading. The SDAIA exercises national steering and measurement functions, but the SAMAI 2 program is collaborative and the two-thirds figure comes from Public First, not from the SDAIA. No independent auditor has published verification of this figure. The definition used for “daily use” is not clarified in available public documents: is it minimal interaction with an assistance tool, or effective integration into administrative decisions?

The difference is substantial.

This technical point does not invalidate the deployment; satellite images of the Hexagon are real, contracts with ministries are documented, employees exist. But it signals a structural limitation of the model: when the state is simultaneously the designer, deployer, and evaluator of its own tools, the data produced are internal management data, not independent measures of public effectiveness.

Democracies confront this problem differently, without resolving it better. Evaluation reports on digital public policies published by independent courts of auditors or parliamentary bodies offer counter-expertise, but with a lag of several years behind actual deployments, and often after structuring decisions have been made.

Speed as a Deliberate Political Choice

The deployment of SAMAI 2 follows the publication of ethics and adoption frameworks. Some officials argue that technological capacity must progress in parallel with normative oversight rather than waiting for it entirely.

This argument deserves to be taken seriously before being criticized. Liberal democracies have mixed experiences on this point. The European Regulation on AI, adopted in 2024 after four years of negotiations, is the most ambitious regulation in the world, and its detractors within European institutions themselves argue that it has slowed useful deployments in public health and crisis management. The Commission itself has since relaxed certain obligations for systems deployed by public administrations.

Official sources describe a deployment accompanied by prior ethics and adoption frameworks, established respectively in 2023 and 2024 by the SDAIA, rather than regulation developed only after deployment. This choice produces real speed gains. It also produces real risks, and their distribution is not symmetric: it is the administered, not the system designers, who first suffer from undetected errors.

The Chinese model of governmental AI followed similar logic, with a notable difference: China developed in parallel a legal infrastructure for sectoral responsibility, certainly steered by the Party, but existing. Saudi Arabia has not yet announced an equivalent.

A Governance Architecture Fixed for a Generation

Technical choices made today on a governmental AI system are not easily reversible. An algorithm for allocating social benefits trained on three years of Saudi data encodes priorities, categories, weightings. Modifying these priorities later requires retracing the entire design chain, something few administrations have the organizational capacity to do. The technical term is institutional “lock-in”: the tool ends up constraining human decisions more than assisting them.

This mechanism applies to all countries, democratic or not. The difference lies in the nature of the available corrective lever. In a system with separation of powers, a court can order an administration to suspend an algorithm whose discriminatory effects are documented, which happened in the Netherlands with the SyRI case in 2020, when a judge prohibited a social fraud detection system that statistically targeted low-income neighborhoods. In the absence of this type of independent judicial recourse, correction depends entirely on the political will of the executive.

The SDAIA has published a framework and principles for AI ethics. The framework entrusts advisory, control, and audit functions to the SDAIA itself, without establishing the formal attributions of a distinct internal committee or its powers of suspension.

Conflict Arbitration When Technological Capacity Precedes Deliberation

The question is no longer theoretical. These ministries handle decisions that directly affect rights: access to healthcare, authorization to work, rights to social transfers. Migrant workers represent a significant share of the Saudi workforce, according to data from the Saudi Central Bureau of Statistics. These are precisely the populations least able to contest an unfavorable administrative decision.

An AI system that processes residence permit applications for a foreign worker and produces an automated refusal leaves that worker with limited recourse: an internal administrative complaint desk. This mechanism exists in many democracies, but it combines with external remedies, namely independent administrative courts, an ombudsman, and a free press capable of investigating systemic malfunctions.

Technical architectures have direct political implications here. Hexagon is a government data center project designed to meet the growing needs of public entities, within a strategy providing for other centers. Diversification of computing sites and data access creates spaces for potential counter-expertise, an architectural choice whose technical and logistical implications join the issues of cooling and locating data centers.

Toward 2030: Three Trajectories for State AI Without Deliberation

Saudi deployment anticipates a question that several technologically capable non-democratic regimes will face in the coming decade. The trajectory is not written. Three distinct scenarios emerge, conditioned on political and institutional variables, none of which is certain.

The first scenario is that of stabilized technocracy. Governmental AI produces real and measurable efficiency gains—reduced delays, fewer errors, better-allocated resources—and these gains reinforce regime legitimacy sufficiently to maintain pressure on remedial mechanisms at a low level. This is the scenario most favorable to the current model. It assumes that deployed systems actually function better than the bureaucratic procedures they replace, and that errors remain sufficiently marginal not to trigger organized contestation. Signals to watch: processing times for administrative requests, published complaint rates, and above all the emergence or non-emergence of independent data on the quality of automated decisions.

Large-scale deployed AI systems produce mass errors: this fact is documented in all contexts where comparable deployments have occurred. A malfunction simultaneously affecting a large and identifiable population could create demand for correction difficult to absorb by internal mechanisms alone.

This scenario assumes only that errors become visible at a scale sufficient that the executive itself finds it in its interest to correct them. The critical variable is decision traceability: if state agents can identify why a system produced a given result, correction remains possible.

If the black box is total, even political will to correct runs up against technical opacity.

The third scenario is that of progressive divergence with international interoperability standards. Saudi companies working with European, American, or Japanese partners operate under data protection regimes incompatible with the European Data Regulation or the American privacy framework. If Saudi governmental AI processes data from these partners, regulatory tensions could compel the regime to a form of negotiation on standards, not from democratic conviction, but from economic necessity. Vision 2030 rests on attracting foreign capital and talent: this dependency creates levers that foreign regulators have not yet systematically seized.

These three trajectories share a constant: the architectural decisions made between 2025 and 2027, choice of provider, degree of computational centralization, design of remedy interfaces, will determine the room for maneuver available in each scenario. A system designed with audit logs accessible to internal inspectors is reformable from within. A system designed as opaque sovereign infrastructure much less so.

Saudi deployment poses liberal democracies a symmetric problem. They have the mechanisms of deliberation and recourse that the Saudi regime has not constructed, but their deployment pace is so slow that they risk regulating systems they will never have implemented themselves. Europe oversees governmental AI with detailed regulation and inspection resources, but its public administrations remain far behind on actual adoption, as shown by performance comparisons between countries that invested in training and those that invested in tools. Speed without governance in one case, governance without deployment in the other: neither of these two models is exportable.

The separation of powers in the face of governmental AI is a question of concrete institutional engineering, independent of political regime in the abstract sense. It hinges on specific points: who has access to decision logs, who can suspend a system, who defines the categories of data used. These questions have precise technical answers. In most countries, political answers of equal precision remain to be formulated.

Sources

  1. Saudi Data and Artificial Intelligence Authority (SDAIA), National AI Strategy 2026 Implementation Report, https://vision2030.ai/sectors/technology/ai-strategy/
  2. SDAIA, Public Sector AI Adoption Index 2026, published by the Saudi Data and Artificial Intelligence Authority (URL not verified; cite without link: Saudi Data and Artificial Intelligence Authority, Public Sector AI Adoption Index 2026)
  3. Saudi Central Bureau of Statistics, data on workforce and foreign labor, General Authority for Statistics, Riyadh
  4. District Court of The Hague, judgment in SyRI case, February 5, 2020, Rechtbank Den Haag, ECLI:NL:RBDHA:2020:1878
  5. Regulation (EU) 2024/1689 of the European Parliament and of the Council on artificial intelligence (AI Act), https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689