SAFE AI / CDAC Network

The Dialogue convenes at a moment of unusual flux in global AI leadership. Conflicts escalating through 2026 have made AI's role in warfare, information operations, and crisis response visible in ways earlier debates treated as theoretical. New voices, including the Independent International Scientific Panel on AI, faith-based moral leadership, and Global South research institutions, are stepping into space established actors no longer dominate. The Dialogue's authority will depend on whether it speaks to that moment or around it. The Global Dialogue succeeds if it produces three things. 1. An explicit acknowledgement that AI assurance and independent audit infrastructure, not only principles, are required for trustworthy AI, and a commitment to support the development of sector-specific assurance and audit bodies generally. 2. Recognition of the 'right to know' when AI is being used, and to seek recourse when it causes harm, and that it extends beyond the interaction layer to the governance layer, covering both systems that shape decisions about people who never encounter the AI directly and AI-generated content that shapes the information environment those people rely on. 3. The integration of crisis-affected communities into the governance frame as a central case, on the basis that humanitarian settings are the hardest stress-test for AI governance: high stakes, low redress, weak market correction, and dependent populations. If humanitarians can't get AI right for the most vulnerable, trust will be jeopordised and we'll all be found failing. A concrete marker of success would be panel recognition of sector-specific assurance and audit infrastructure as a priority gap, with humanitarian action identified as a priority first-mover case. Without these outcomes, the terms of AI governance for crisis-affected populations will be set by default through fragmented commercial procurement by humanitarian organisations and national authorities, unilateral donor conditions, or absorption into regulatory regimes designed for other purposes.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

These four themes cover the operational loop that AI governance must close if it is to work in practice. Safety, security and trustworthiness are properties of AI systems in deployment. They depend on the assurance architectures of the organisations deploying them. Humanitarian action is the hardest stress-test case: high stakes, low redress, weak market correction, and populations for whom there is no alternative provider. Transparency, accountability and human oversight determine whether deployed systems can be known, challenged, and corrected. In humanitarian settings, AI most often shapes decisions about people who never interact with the system directly. The right to know must extend from the interaction layer to the governance layer. Interoperability of governance approaches asks whether the proliferation of AI frameworks produces coherent protection or fragmented compliance. The UNESCO Recommendation, NIST AI RMF, EU AI Act, Council of Europe Framework Convention, and the UN Model Policy on the Responsible Use of AI in UN System Organizations (2024) provide foundations. What is missing is the sector-specific operationalisation layer. The SAFE AI Framework, developed by CDAC Network with The Alan Turing Institute and Humanitarian AI Advisory and with founding investment from UK FCDO, is designed as that layer and launches May 2026. Social, economic, ethical, cultural, linguistic and technical implications of AI are where the stakes land on people. This includes information integrity: coordinated inauthentic behaviour, AI-amplified disinformation, and synthetic content are now documented as weapons of war in crisis settings, with direct protection consequences for affected communities and humanitarian responders. Together, these four themes map the governance loop from deployment conditions through oversight, interoperability, and community impact.

Three cross-cutting gaps merit explicit treatment. First, dataset repurposing and secondary use. Operational data collected by humanitarian and development agencies, including biometrics, protection case files, displacement records, and service uptake data, is being used to train, fine-tune and evaluate AI systems without a governance framework covering consent, purpose limitation, or community oversight. The same data is at risk of secondary use by actors with interests hostile to the people it describes. No existing regulatory regime addresses this at the global level. Second, the gap between the interaction layer and the governance layer. Most current transparency and oversight provisions assume the person affected by a decision is a user of the system. In humanitarian, public service and benefits settings, the person affected rarely interacts with the AI at all. Governance frameworks need explicit provisions for non-user affected persons, including community-in-the-loop requirements that extend the established human-in-the-loop principle to the governance layer. Third, the absence of sector-specific assurance and audit infrastructure. Principle documents proliferate, but there is limited independent audit capacity for AI deployments in high-stakes, low-redress contexts. The humanitarian sector is developing a first-mover response through SAFE AI Phase 2, which will establish an independent audit hub with sector-specific benchmarks and a cross-subsidy model enabling Global South participation. Building assurance and audit infrastructure now is cheaper than responding to the governance failures its absence will produce, as the Dutch childcare benefits scandal and comparable cases have shown. These gaps sit inside a live institutional reorganisation. The UN80 reform process is consolidating humanitarian coordination and framing AI as an efficiency mechanism for a leaner system. No accountability mechanism for AI use has yet been defined. The UN80 Data Quint and IASC Humanitarian Data Collaborative address data governance but not AI deployment governance. SAFE AI is designed to plug directly into that architecture

The humanitarian sector's accountability to affected populations has long been incomplete. It operates without market correction, without legally enforceable rights for affected people in most settings, and without independent audit infrastructure of the kind found in regulated sectors. Normative commitments such as the Core Humanitarian Standard exist without enforcement architecture. The position resembles aviation sixty years ago, before independent external assurance was built alongside the industry. Affected people have no legal recourse, no electoral leverage, and no advocates with structural power. AI in humanitarian action now needs that architecture. AI deployment is running ahead of oversight. Operational systems for forecasting, targeting, eligibility assessment, biometric registration, case prioritisation, and information moderation are in use across UN agencies, INGOs and pooled funds. No donor or humanitarian pooled fund grant agreement currently requires independent AI assurance as a condition of funding. The funding closest to affected people is currently the least likely to carry AI accountability requirements (McElhinney et al., 2026). Sector incentives favour rapid adoption over accountable adoption. The information integrity dimension is acute. CDAC Network's analysis of Sudan's information war documents how coordinated inauthentic behaviour, bot networks, AI-amplified hate speech, impersonation of humanitarian sources, and language-specific disinformation now operate as weapons of war, obstructing aid delivery and endangering responders (CDAC Network, 2025). In Khartoum, community kitchen volunteers were targeted and killed within hours of Facebook misinformation identifying them with one of the warring parties. SAFE AI is being stress-tested in this environment through a live collaboration with technical partner Valent, co-designed with Sudanese Mutual Aid Groups and Local NGOs. The consortium behind SAFE AI has consulted more than twenty humanitarian organisations, multiple technology think tanks, academics and industry, held discussions across 18months of communities of practice, convened a regional dialogue in Nairobi in June 2025 with Norwegian Refugee Council facilitation for local organisations to input, and run a beta-testing cohort including for example Save the Children, the British Red Cross and Internews at the Alan Turing Institute in October 2025, and completed community co-design fieldwork with FilmAid Kenya in Kakuma refugee camp in Northern Kenya. The evidence base is now sufficient to launch a sector-wide framework in May 2026 and to move into the second phase: an independent humanitarian AI audit function. The window for establishing this as a standard before fragmented approaches fill the gap is narrow.

The Dialogue's distinctive contribution is to move international cooperation on AI governance from principle convergence, which is already advanced, to operational convergence, which is not. Principle convergence has produced the UNESCO Recommendation, OECD AI Principles, the NIST AI Risk Management Framework, the EU AI Act, the Council of Europe Framework Convention, and the UN Model Policy on the Responsible Use of AI in UN System Organizations (2024). These establish substantial normative alignment. Operational convergence, meaning sector-specific tools, independent assurance infrastructure, and shared audit capacity, has not kept pace. The Dialogue can advance cooperation in four concrete ways. First, by recognising sector-specific operationalisation layers as a category of governance infrastructure worth naming in the AI standards charter proposed in Governing AI for Humanity (UN, 2025), rather than leaving each sector to negotiate its own translation of global principles. Second, by endorsing independent assurance and audit as the mechanism that converts principles into enforceable practice, drawing on the architecture set out in the PAI AI Assurance Ecosystem whitepaper (2026) and NIST AI 800-4. Third, by embedding community-in-the-loop participation as a governance requirement rather than an optional consultation stage, extending the established human-in-the-loop principle from the interaction layer to the governance layer. This matters particularly for deployment contexts where affected people do not interact with the AI directly. Fourth, by establishing connective tissue between the UN AI governance architecture, including IPAIS, and live reform processes such as UN80 and the Humanitarian Data Collaborative. AI governance is currently absent from these reforms despite AI being embedded in the coordination infrastructure they are rebuilding.

The Dialogue should connect with initiatives already operating across three layers. Global governance layer. The Independent International Scientific Panel on AI (IPAIS); the International AI Safety Report 2026; the OECD.AI Policy Observatory; UNESCO's monitoring of its Recommendation on AI Ethics; and Council of Europe Convention implementation. These provide the evidence base and normative infrastructure the Dialogue should draw on rather than duplicate. Assurance infrastructure layer. The Partnership on AI (PAI) AI Assurance Ecosystem work (2026); NIST AI 800-4 on assurance; the UK DSIT Trusted Third-Party AI Assurance Roadmap and AI Assurance Innovation Fund; frontier lab responsible scaling commitments. This layer is where operational convergence will be won or lost. The Dialogue can accelerate it by recognising independent assurance as a priority gap. Sector-specific operationalisation layer. Initiatives translating global principles into deployment conditions for specific high-stakes sectors. SAFE AI (Standards and Assurance Framework for Ethical AI) is the humanitarian sector example: a published governance framework (May 2026 launch) developed by CDAC Network with The Alan Turing Institute and Humanitarian AI Advisory, with founding investment from UK FCDO, building toward an independent humanitarian AI audit function. Comparable operationalisation work is emerging in health AI, children's AI (UNICEF Policy Guidance on AI for Children), and public sector AI procurement. The Dialogue's added value is to establish this three-layer architecture as the frame for international cooperation, and to ensure connection between the layers. Principle documents without assurance infrastructure generate compliance theatre. Assurance infrastructure without sector operationalisation cannot reach the deployment conditions that matter. Sector operationalisation without connection to global governance risks fragmentation. The Dialogue's added value also includes connecting AI governance to live humanitarian reform

The Dialogue's structure should reflect the three-layer architecture set out above: global governance, assurance infrastructure, and sector-specific operationalisation. Stakeholders contribute differently at each layer, and format should follow function. States and regional bodies should contribute through the global governance layer: committing to the AI standards charter proposed in Governing AI for Humanity, supporting the institutional home within the UN system, and making compliance with assurance standards a condition of public procurement and development financing. The single most consequential contribution is to embed independent assurance as a condition of funding. Frontier AI developers and deployers should contribute through the assurance infrastructure layer: supporting the independent audit function, committing to sector-specific assurance where their systems are deployed in high-stakes contexts, and participating in deployment-level transparency requirements. Sector actors and operational agencies should contribute through the operationalisation layer: developing, testing and running sector-specific assurance frameworks, and providing the evidence base that anchors global cooperation in deployment reality. Crisis-affected communities and affected-population representatives should contribute as governance participants, not consultation inputs. This requires resourced, structural participation rather than symbolic inclusion. Community-in-the-loop requirements should be embedded in the Dialogue's own process design as a precedent. Format recommendations: a standing structure with three working layers, not one plenary, with each layer publishing outputs that the next Dialogue reviews; sector-specific sessions anchored in deployment evidence, including humanitarian, health, public benefits, and AI in conflict settings, which are where operational convergence will be tested; a written submissions track that feeds synthesised inputs into the plenary, so the diversity of evidence from sector-specific work is not lost; and structural participation by Global South institutions as co-chairs of layers, not invited speakers.

Four sets of voices are currently underrepresented in global AI governance discussions. Crisis-affected communities. The people whose lives are most directly shaped by AI in the hardest deployment contexts are systematically absent from the governance frame. CDAC Network's SAFE AI work, grounded in community co-design fieldwork with FilmAid Kenya in Kakuma (193 participants, December 2024), is one of the few governance design processes to integrate this voice structurally. The Dialogue should establish mechanisms for affected-population representation as governance participants rather than subjects of consultation. Global South AI research institutions. Organisations including the Distributed AI Research Institute (DAIR), African university-based AI governance researchers, and regional civil society bodies working on AI in development and humanitarian contexts are producing evidence that rarely reaches global policy fora. The Dialogue should create a dedicated channel for Global South research contributions and fund participation, consistent with the third condition in Governing AI for Humanity. Frontline humanitarian and public service practitioners. Those deploying AI in needs assessment, eligibility determination, health triage, and information response have operational evidence of where governance breaks down. They are rarely in policy rooms. Sector-specific Dialogue sessions should include practitioner panels drawing directly on live deployment experience, including in active conflict settings. Local information integrity actors. In crisis contexts, grassroots verification networks, youth volunteers, local and community media, and local civil society already operate as interpreters and validators of information that affects life-critical decisions. Their practices are documented in CDAC Network's 2025 Sudan research. They are invisible in most AI governance discussions. The mechanism for inclusion matters more than the invitation. Resourced participation, interpretation, travel and digital access support, and co-chair roles for underrepresented actors should be embedded in the Dialogue's structure rather than offered as accommodations.

The most effective formats reduce the distance between governance decisions and deployment evidence, and between governance fora and the people AI systems affect. AI as participation infrastructure. Real-time interpretation in the working languages of regions most affected by AI deployment decisions, accessible summarisation, asynchronous tools, and search across submissions are infrastructure the Dialogue should invest in, and an important signal about whose voices it is designed for. But synthetic stakeholder engagement and automated outputs presented as community input should be named as risks. Community testimony streams. Structured inclusion of testimony from crisis-affected communities, with mechanisms for that testimony to influence session outputs. Precedent exists: community video testimony from Kakuma was brought into an FCDO-convened Wilton Park gathering in 2024 as the only direct voice from affected populations in the room. Publish inputs, not just summaries. Submissions, testimony and working group outputs should be made available in full, not only as a synthesised report. When diverse inputs get blended into one official narrative, the disagreements and minority views that matter most are the first thing lost. Humanitarian coordination at scale has shown this repeatedly. Sharing inputs openly, and inviting independent scrutiny of how they are read, is the participation design that holds up. Listen to the outliers. Single testimonies, edge cases and minority positions should be treated as early warnings about where governance is breaking down, not smoothed out by aggregate analysis. Where large-scale analysis of submissions is used, its job is to check for bias in how the Dialogue is hearing its evidence, not to override what the smaller signals are saying. Red-team sessions. Stress-testing draft Dialogue outputs against specific deployment contexts before finalisation. This is standard practice in AI safety. It should be standard in AI governance. Hybrid working formats. Written submissions, moderated online discussion and in-person plenary combined so participation does not require travel. This addresses Global South access and the cost barrier for civil society.

Published sector-specific assurance frameworks: The SAFE AI Framework (CDAC Network, The Alan Turing Institute, Humanitarian AI Advisory; UK FCDO founding investment; May 2026 launch) translates global AI governance principles into operational tools for humanitarian deployments, covering problem definition, impact assessment, transparency documentation, technical assurance, and deployment monitoring. Freely available. Designed to interoperate with UNESCO, NIST, OECD, EU AI Act, Council of Europe Convention, and the UN Model Policy on Responsible AI. A published governance gap analysis sets out the architecture (McElhinney, Mazumder and Tjalve, 2026). Community-in-the-loop as a lifecycle requirement. Extending the established human-in-the-loop principle to require affected community participation across problem definition, design, development, and deployment. Embedded in SAFE AI and replicable principle for high-stakes sectors. Independent assurance infrastructure. The UK DSIT Trusted Third-Party AI Assurance Roadmap and AI Assurance Innovation Fund establish the premise that independent assurance is what makes AI trustworthy in high-stakes contexts. PAI's AI Assurance Ecosystem whitepaper (2026) and NIST AI 800-4 set out the architecture. SAFE AI Phase 2 proposes the first sector-specific application of this architecture to humanitarian deployments, with Global South co-governance and sector-specific audit benchmarks. Co-designed community participation in AI governance. SAFE AI's Kakuma fieldwork with FilmAid Kenya and Nairobi regional dialogue are examples of grounded co-design methodology. Procurement conditionality. Embedding assurance compliance as a condition of public procurement and development financing is the enforcement mechanism that converts voluntary frameworks into sector standards. References McElhinney, H. with Mazumder, A. and Tjalve, M. (2026) SAFE AI: The Governance Gap in Humanitarian AI. Addressing the structural gap between global frameworks and operational reality. CDAC Network, The Alan Turing Institute, Humanitarian AI Advisory. Available at: www.cdacnetwork.org/resources/the-governance-gap-in-humanitarian-ai. CDAC Network (2025) Sudan's information war: How weaponised online narratives shape the humanitarian crisis and response. Available at: www.cdacnetwork.org/resources/sudan-information-war-2025