AI-Enabled Digital Public Infrastructure: Responsible AI Governance and DPI Safeguards

 AI-ENABLED DIGITAL PUBLIC INFRASTRUCTURE

An Operational Framework for Responsible AI Governance, Human Oversight and 

DPI Safeguards

Author

Nasir Rehman

Independent Technical Working Paper

Version 1.0 | August 2026

Purpose

An independent practitioner contribution examining how responsible AI governance can complement DPI safeguards as AI becomes embedded within population-scale digital infrastructure.

Status

This is an independent publication. It is not an official publication of the United Nations, UNDP, the DPI Safeguards Initiative or any other international organisation.

Executive Summary

Digital Public Infrastructure (DPI) is becoming foundational to how societies establish digital identity, make payments, exchange data and deliver public and private services at population scale. At the same time, artificial intelligence is moving from standalone applications into the systems, workflows and decision processes that operate on top of digital infrastructure.

This convergence creates an important governance question: How can AI be integrated into DPI without weakening the safety, inclusion, agency, accountability and trust that DPI requires?

The Universal DPI Safeguards Framework provides a rights-based and practical foundation for addressing this challenge. This paper does not seek to replace or reproduce that framework. Instead, it proposes an additional operational layer focused specifically on the governance of AI when AI systems interact with or operate on top of DPI.

Central proposition: AI governance should become an operational component of DPI safeguards wherever AI can materially influence access, decisions, services, data, security, economic participation or public trust.

The proposed framework introduces six components: AI-DPI impact assessment; risk-proportionate governance; human oversight architecture; technical and procurement controls; accountability and redress; and continuous monitoring, measurement and learning.

These components operate across five DPI lifecycle stages: conception and scoping; strategy and design; development; deployment; and operations and maintenance. The paper also introduces an AI-DPI Control Loop:


Identify → Assess → Design → Test → Approve → Deploy → Monitor → Intervene → Learn.

The objective is not zero risk. The objective is managed, visible, accountable and continuously improving risk.

  1. The Emerging AI-DPI Convergence

The first generation of digital transformation focused primarily on digitising processes. The second connected systems, institutions, people and data. The emerging phase is increasingly characterised by intelligent systems that can interpret information, generate content, make recommendations and perform actions.

AI can become a decision-support layer, service-delivery layer, fraud-detection layer, cybersecurity layer, identity-verification layer, data-analysis layer, automation layer or autonomous action layer.

The opportunity is significant. AI can improve public-service efficiency, detect anomalous transactions, assist officials, improve accessibility and support better allocation of scarce resources.

But the risks change when AI operates through infrastructure used by millions of people. A model error in an isolated enterprise application may affect a department. A model error integrated into foundational DPI may affect an entire population.

Principle: The governance intensity of AI should increase as the scale, autonomy, interconnectedness and consequence of the system increase.

  1. Relationship with the Universal DPI Safeguards Framework

The Universal DPI Safeguards Framework is a multi-stakeholder framework intended to guide safe and inclusive DPI implementation and to address risks at individual and societal levels. This paper is deliberately positioned as a complementary AI governance layer, not a competing framework.

The broader DPI safeguards architecture establishes the foundation. This paper asks what additional controls are required when AI becomes embedded within that architecture.

DPI Safeguards → AI-DPI Risk Assessment → AI Governance Controls → Technical Implementation →

Monitoring & Evidence → Human Accountability & Redress

The DPI Safeguards Initiative describes its approach as minimising risks across technical, normative and organisational layers while making implementation secure, inclusive, practical and adaptable. The Initiative also describes a Discover → Design → Implement → Evolve process, reinforcing the importance of learning from implementation. (DPI Safeguards Initiative, 2026).

  1. The Six-Layer AI-DPI Governance Model


Layer 1: Purpose and Public Value

Before introducing AI, establish the problem, beneficiaries, potential harms, public value, alternatives and consequences of failure. AI should not be introduced simply because it is technologically available.

Layer 2: Risk and Impact Assessment

Assess individual, institutional and societal impacts, considering scale, severity, reversibility, autonomy and interconnectedness.

Layer 3: Human Oversight Architecture

Design meaningful oversight with awareness, authority, capacity and intervention. A person merely appearing in a workflow is not sufficient.

Layer 4: Technical and Procurement Controls

Translate safeguards into architecture, data governance, security, auditability, supplier obligations, contracts and lifecycle controls.

Layer 5: Accountability and Redress

Assign decision rights and create accessible mechanisms to challenge consequential AI-assisted outcomes.

Layer 6: Continuous Monitoring, Measurement and Learning

Measure safeguard effectiveness, detect drift and incidents, and feed evidence back into governance and system design.

  1. AI-DPI Risk Classification

Level 

Characteristics 

Minimum governance 

1: Limited 

Low consequence, reversible, limited users 

Basic documentation, security and monitoring 

2: Material 

Meaningful operational or individual impact 

Formal risk assessment and accountable owner 

3: High 

Significant rights, financial or service-access implications 

Enhanced testing, independent review and human approval 

4: Systemic 

Population-scale, highly consequential or difficult to reverse 

Senior approval, continuous monitoring, tested intervention and independent assurance 

Classification should remain context-sensitive. A system that is low-risk in an isolated environment can become high-risk when connected to foundational DPI.

  1. Mapping DPI Safeguards to AI Controls

DPI safeguard area 

AI-DPI operational control 

Do no harm 

AI impact assessment, harm modelling and risk thresholds 

Non-discrimination 

Bias testing, representative evaluation and disparity monitoring 

Non-exclusion 

Accessibility testing, alternative channels and inclusive user journeys 

Transparency and accountability 

AI disclosure, decision ownership and audit trails 

Rule of law 

Legal basis, regulatory review and documented decision authority 

Autonomy and agency 

User choice, meaningful human review and contestability 

Community engagement 

Participatory testing and stakeholder feedback 

Remedy and redress 

Appeals, complaints, correction and human review 

DPI safeguard area 

AI-DPI operational control 

Future sustainability 

Lifecycle cost, institutional capacity and environmental assessment 

Evolving with evidence 

Continuous monitoring, audits and model reassessment 

Privacy/security by design 

Data minimisation, access controls, threat modelling and secure architecture 

Inclusive governance 

Multi-stakeholder oversight and documented decision rights 

Financial viability 

Sustainable operating model and lifecycle cost assessment 

Open/interoperable assets 

Open standards, interoperability and portability where appropriate 


  1. AI Controls Across the DPI Lifecycle

The AI governance layer should operate across the same lifecycle rather than being treated as a separate compliance activity.

Lifecycle stage 

Key AI-DPI questions and controls 

Primary output 

Conception & Scoping 

Is AI necessary? Who could be affected? What services or rights could be impacted? 

AI-DPI Impact Assessment 

Strategy & Design 

Define governance, data, security, human oversight, procurement, monitoring and redress. 

Governance & Technical 

Design 

Development 

Test data, model performance, security, fairness, accessibility, failure modes and auditability. 

Verified Release Candidate 

Deployment 

Controlled rollout, capacity-building, monitoring, communication, redress and incident response. 

Approved Operational 

Deployment 

Operations & Maintenance 

Monitor drift, incidents, complaints, security, model changes, impact and retirement readiness. 

Continuous Safeguard 

Assurance 


  1. The AI-DPI Control Loop

     

AI-DPI 

Control Loop

Identify → Assess → Design → Test → Approve → Deploy → Monitor → Intervene → Learn

The lifecycle provides the governance structure. The Control Loop provides the operational mechanism. The loop should repeat whenever material model, data, use-case, risk or context changes.

A responsible AI system must have an explicit pathway for intervention, escalation, modification or retirement. A theoretical override that cannot be executed under operating conditions is not a meaningful safeguard.

  1. Responsible Authorities and Accountability

Responsible authority 

Core AI-DPI responsibility 

Government / DPI owner 

Public purpose, policy and ultimate accountability 

Responsible authority 

Core AI-DPI responsibility 

DPI programme leadership 

Programme governance, coordination and risk management 

Regulator 

Rights, legal compliance, competition and independent oversight 

Technology provider 

Technical safeguards, security, testing and documentation 

Procurement authority 

Contractual safeguard requirements and supplier accountability 

Funder 

Safeguard conditions and sustainability requirements 

Civil society / advocates 

Independent scrutiny and community representation 

Independent assurance 

Testing, audit and evidence-based challenge 

Service operator 

Operational monitoring and incident response 

Users / communities 

Feedback, participation and contestability 

The purpose of this model is to prevent accountability from disappearing behind the statement that “the algorithm made the decision.” Algorithms do not replace institutional accountability.

  1. Responsible AI-DPI Procurement Control Set

Procurement is an early opportunity to make safeguards enforceable. AI-enabled DPI procurement should address:

  • Data: ownership, permitted processing, retention, deletion and secondary use.

  • Model: provenance, versioning, material changes, performance expectations and limitations.

  • Security: security standards, vulnerability management, incident reporting and access control.

  • Accountability: audit rights, documentation, assurance requirements and responsible personnel.

  • Continuity: availability, disaster recovery, business continuity and transition arrangements.

  • Interoperability: open standards, APIs, portability and avoidance of unnecessary vendor lock-in.

  • Human oversight: intervention requirements, escalation procedures and suspension mechanisms.

  • Exit: data return/deletion, transition support, knowledge transfer and replacement readiness.

Procurement principle: safeguards should become contractual obligations rather than post-deployment aspirations.

  1. Inclusion, Agency and Redress

AI can unintentionally increase barriers for people already experiencing digital exclusion. An AI-DPI Inclusion Test should examine language, literacy, disability, age, connectivity, affordability, digital skills, alternative channels and contestability.

The objective is not merely equal access to technology. It is equitable access to the outcome the technology is intended to enable.

For consequential decisions, transparency should support contestability. Affected people should understand, where appropriate, whether AI is involved, what role it plays, relevant limitations, who is accountable and how an outcome can be challenged.

This approach is consistent with the Global Digital Compact's emphasis on a responsible, accountable, transparent and human-centric approach across the technology lifecycle, with effective human oversight, and with its emphasis on equitable AI governance and capacity-building, particularly for developing countries (United Nations, 2024).

  1. Security, Privacy and Technical Assurance

AI-enabled DPI creates a combined privacy and cybersecurity challenge. Controls should include data minimisation, purpose limitation, access controls, retention limits, privacy impact assessment, threat modelling, encryption, authentication, vulnerability management, monitoring, incident response and appropriate adversarial testing.

AI-specific threats can include prompt manipulation, data poisoning, model extraction, adversarial inputs, unauthorised model changes, training-data leakage and excessive autonomous action.

Security and privacy should be reassessed as systems evolve. The OECD AI Principles similarly emphasise human agency and oversight, transparency and explainability, robustness, security and safety, and accountability throughout the AI lifecycle (OECD, 2024).

  1. Sustainability

Sustainability should be assessed across four dimensions: financial, institutional, technical and environmental.

  • Financial sustainability: Can the system be funded over its full lifecycle?

  • Institutional sustainability: Does the country or organisation have the capability to operate and govern it?

  • Technical sustainability: Can the system be maintained, upgraded and integrated without excessive dependency?

  • Environmental sustainability: What are the infrastructure, compute, energy and resource implications?

AI-enabled DPI should be evaluated not only on whether it can be launched, but whether it can remain safe, useful and governable over time.

  1. Measurement, Evaluation and Learning

Dimension 

Example indicator 

Governance 

% of AI-DPI systems with named accountable owners 

Risk 

% completing documented impact assessments 

Human oversight 

% of high-impact systems with tested intervention mechanisms 

Testing 

% completing appropriate pre-deployment testing 

Transparency 

% of relevant services disclosing AI involvement 

Redress 

Average time to resolve AI-related complaints 

Inclusion 

Performance disparity across relevant user groups 

Accessibility 

% of priority user journeys passing accessibility testing 

Security 

Number of material AI-related security incidents 

Privacy 

% of systems completing privacy assessment 

Procurement 

% of critical suppliers meeting safeguard clauses 

Monitoring 

% of high-impact systems subject to periodic reassessment 

Sustainability 

Lifecycle cost and environmental assessment completed 

Learning 

% of material incidents resulting in documented control changes 

Indicators should be adapted to national context rather than treated as a universal compliance checklist. The objective is to connect safeguard principles with observable evidence and learning.

  1. AI-DPI Deployment Readiness Assessment

Before population-scale deployment, the responsible authority should be able to demonstrate:

  • Purpose: clear public/service purpose; AI necessity established; alternatives considered.

  • Risk: impact assessment completed; high-risk populations identified; residual risk explicitly accepted.

  • Human oversight: accountable authority identified; human review established; override tested.

  • Technical safeguards: security testing, privacy controls, auditability and model/version management.

  • Inclusion: accessibility tested; relevant language needs addressed; alternative channels considered.

  • Accountability: supplier responsibilities documented; regulatory obligations addressed; escalation established.

  • Redress: complaint mechanism available; appeal process defined; human review accessible.

  • Monitoring: indicators and intervention thresholds established; continuous monitoring activated.

  • Sustainability: lifecycle financing, institutional capacity, technical maintenance and environmental implications assessed.

If critical controls are absent, deployment should be delayed, restricted or redesigned.

  1. Implementation Roadmap and Maturity

Step 

Action 

1. Establish 

Create AI-DPI governance ownership and identify existing DPI risks. 

2. Assess 

Map AI systems against DPI principles, risks and lifecycle stages. 

3. Embed 

Integrate safeguards into architecture, procurement, contracts and operating procedures. 

4. Measure 

Introduce indicators, assurance mechanisms, user feedback and independent review. 

5. Adapt 

Use evidence to modify governance, technology and policy continuously. 

A practical maturity model is: Reactive → Policy-Based → Managed → Adaptive. The strategic objective is to move from compliance with safeguards to institutional capability to operate safely.

  1. Strategic Recommendations

  1. Treat AI as a DPI safeguard issue wherever AI can materially affect DPI outcomes.

  1. Establish accountability before deployment and identify an accountable authority for high-impact systems.

  1. Make human oversight operational by defining authority, capability, intervention and escalation.

  1. Embed safeguards into procurement so critical controls become enforceable supplier obligations.

  1. Design for inclusion and test systems against real-world barriers rather than assuming equal access.

  1. Make redress part of the architecture for consequential AI-assisted decisions.

  1. Measure safeguard outcomes and not only policy completion.

  1. Build local technical, legal, governance and operational capability alongside infrastructure.

  1. Design for financial, institutional, technical and environmental sustainability.

  1. Govern continuously and evolve controls with evidence, technology, threats and societal needs.


Conclusion

Digital Public Infrastructure has the potential to become one of the most important foundations of inclusive digital economies and public services. Artificial intelligence may significantly increase that potential. It may also increase the consequences of failure.

The answer is not to prevent innovation, nor to treat safeguards as a compliance exercise after technology has already been deployed. The answer is to make safeguards part of the architecture of digital transformation.

From principles to operational controls.

From policy documents to accountable implementation.

From human presence to meaningful human oversight.

From equal access to equitable outcomes.

From procurement as purchasing to procurement as governance.

From static compliance to continuous learning.

The ultimate test of responsible AI is not how intelligent the system becomes. It is whether people remain protected, informed, empowered and able to seek meaningful human accountability when the system affects their lives.

Responsible AI is not an obstacle to digital transformation. Properly designed, it is part of the infrastructure that makes transformation sustainable.


References

DPI Safeguards Initiative (2026) Universal DPI Safeguards Framework, Version 2.0. Available at: https://www.dpi-safeguards.org/framework (Accessed: 16 August 2026).

DPI Safeguards Initiative (2026) DPI Safeguards Initiative. Available at: https://www.dpi-safeguards.org/ (Accessed: 16 August 2026).

Organisation for Economic Co-operation and Development (2024) OECD AI Principles. Paris: OECD. Available at: https://www.oecd.org/en/topics/ai-principles.html (Accessed: 16 August 2026).

United Nations (2024) Global Digital Compact: Annex I to the Pact for the Future. New York: United Nations. Available at: https://www.un.org/pact-for-the-future/en/annex-i-global-digital-compact (Accessed: 16 August 2026).

United Nations Office for Digital and Emerging Technologies (2026) Global Digital Compact. Available at: https://www.un.org/digital-emerging-technologies/global-digital-compact (Accessed: 16 August 2026).

United Nations (2025) Resolution A/RES/79/325: Independent International Scientific Panel on Artificial Intelligence and Global Dialogue on Artificial Intelligence Governance. Available through the United Nations Global Digital Compact AI resources. (Accessed: 16 August 2026).

United Nations (2026) Global Digital Compact: AI Panel and Dialogue. Available at: https://www.un.org/global-digital-compact/en/ai (Accessed: 16 August 2026).

Author's Note: 

This working paper is an independent practitioner contribution. It is not an official publication of the United Nations, UNDP, the DPI Safeguards Initiative or any other international organisation. It does not claim authorship, endorsement or institutional affiliation with the Universal DPI Safeguards Framework. Its purpose is to contribute a practitioner perspective on the governance of AI-enabled systems within the broader DPI safeguards ecosystem.

Research and citation note

The publication was reviewed against current public materials available on 16 August 2026. Claims about the DPI Safeguards Initiative, the Global Digital Compact and OECD AI Principles were checked against primary institutional sources. The paper intentionally distinguishes the author's proposed framework from official frameworks and does not represent the proposed controls as UN-endorsed recommendations.

Comments

Popular posts from this blog

Human-Centered AI in Robotics: Balancing Automation with Ethical Sustainability

Ethics in the Age of Gen AI