NetHope

NetHope is a consortium of ~60 of the world's largest international nongovernmental organizations, working collectively across 190 countries. Our members deliver education, healthcare, disaster response, and development programming to hundreds of millions of people. They are also AI adopters, deploying these systems in some of the most complex operating environments on earth: conflict zones, displacement settings, low-infrastructure contexts, and communities where power asymmetry between service providers and the people they serve is a defining condition. From this vantage point, a successful first Global Dialogue would accomplish three things: First, it would operationalize governance. NetHope's analysis of 53 AI governance instruments found extensive coverage at the regulatory and intergovernmental levels, but a critical "Missing Middle" in the sector-wide governance layer that translates those principles into practice (https://nethope.org/toolkits/ai-governance-and-the-nonprofit-sector-mapping-the-missing-middle/). Our members consistently identify this as their most urgent need: practical tools, shared standards, and implementation guidance that make existing commitments operational. The Dialogue should prioritize closing the distance between principle and practice. Second, it would establish formal pathways for the organizations deploying AI in high-stakes humanitarian contexts to contribute operational knowledge to governance design on a continuous basis. Data protection in fragile states, meaningful consent under dependency, algorithmic bias in development contexts, and AI deployment across jurisdictions are areas where our members hold hard-won expertise that global governance processes need. That expertise should inform governance structures, not arrive as testimony after structures are built. Third, it would commit to governance that enables responsible adoption, not only restricts harmful use. For organizations serving the world's most vulnerable populations, the risk of exclusion from AI's benefits is as consequential as the risk of harm from its misuse. Governance must build capacity alongside accountability.

• AI capacity-building
• Interoperability of governance approaches
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

These four priorities reflect what NetHope sees across a coalition of more than 60 of the world's largest INGOs, informed by a systematic analysis of 53 global AI governance instruments scored against 14 themes relevant to how nonprofits develop, procure, and deploy AI. AI capacity-building is the prerequisite for everything else. Only 8% of humanitarian organizations report AI as widely integrated, and fewer than a quarter have formal AI policies. Our members consistently identify their most urgent need not as more principles but as practical tools to operationalize existing ones. Governance without capacity is aspiration. The Dialogue must treat capacity-building as governance infrastructure, not a parallel workstream. Interoperability of governance approaches addresses a structural problem our members confront daily. NetHope's analysis identified what we call the "missing middle": the sector-wide governance layer between broad international frameworks and individual organizational policies barely exists. Only five instruments in our dataset of 53 attempt shared governance guidance across the nonprofit sector. Organizations operating across jurisdictions navigate contradictory compliance requirements alone, wasting scarce resources and producing inconsistent protections for the communities they serve. Protection and promotion of human rights must anchor this Dialogue, and the data shows why. Humanitarian principles alignment is addressed by only 19% of governance instruments globally. Power asymmetry and consent, 33%. Data protection in low-infrastructure settings, 20%. Binding regulations, AI developer frameworks, and national strategies all score zero on these themes. They were not designed for contexts where survival depends on a service provider, where consent frameworks become coercive, and where algorithmic outputs carry life-or-death consequences. The sector cannot outsource these governance needs. They must be built from operational experience. Transparency, accountability, and human oversight are essential because the nonprofit sector increasingly depends on AI systems it cannot independently verify. Organizations procure foundation models with no access to training data provenance, documented failure modes, or downstream use restrictions. Without shared standards for vendor due diligence, procurement accountability, and impact assessment, individual organizations bear risks they lack the technical capacity to evaluate. Transparency requirements create the minimum conditions for informed decision-making and meaningful accountability to affected populations.

Three cross-cutting issues that the listed thematic areas do not adequately capture deserve the Dialogue's attention. NetHope raises these based on direct operational experience across our member coalition and the governance gaps our research has documented. First, the governance of funding relationships around AI. Across every governance layer NetHope analyzed, this is the largest unaddressed gap. Foundations and institutional donors are signaling expectations around AI-enabled programming, but no shared framework establishes mutual responsibilities between funders and the organizations they support. Our members report pressure to demonstrate AI adoption to remain competitive for funding, sometimes before adequate governance is in place. This dynamic drives ungoverned adoption at scale. It sits outside traditional regulatory or human rights framing but determines how AI is actually adopted across civil society globally. The Dialogue should recognize funder-grantee AI governance as a structural issue requiring attention. Second, the governance of agentic AI systems. No governance instrument in NetHope's dataset addresses agentic AI by name. These systems collapse the window between output and action that current governance assumes exists. In earlier AI systems, a human reviews a recommendation before acting. In agentic systems, the system sequences its own actions, retrieves data, and triggers workflows autonomously. For organizations deploying AI where algorithmic outputs carry life-or-death consequences, this demands governance designed for speed and iteration, not static instruments that require years to update. Third, dual-use risks and AI in contested information environments. Humanitarian data documents vulnerable people's locations, identities, and needs. AI intensifies existing dual-use risks by increasing both the volume of data generated through operations and the sophistication of what can be inferred from it. Simultaneously, humanitarian organizations increasingly operate in environments where information itself is weaponized. AI systems that generate, translate, or summarize content in these contexts face compounding risks: manipulation of training data, repurposing of humanitarian AI infrastructure for surveillance or military objectives, and exploitation of AI outputs for disinformation. Standard safety frameworks do not address these intersecting threats.

The most significant opportunity is shifting the nonprofit sector from consumer of AI to active participant in shaping it. Our members deploy AI in the world's highest-stakes environments, but they have almost no role in how those systems are built. Foundation models are developed for commercial markets, with training data that systematically underrepresents the populations and languages our members serve. Models underperform in low-resource languages, misclassify in development contexts, and embed assumptions designed for conditions that do not exist in humanitarian field settings. This is not inevitable. Humanitarian organizations can partner with model developers and research institutions on supervised fine-tuning for underserved contexts, red-teaming for high-stakes deployment scenarios, and pre-deployment evaluation in conditions that commercial testing cannot replicate. The Global Dialogue should create mechanisms for this kind of structured collaboration. It would produce safer AI systems and more legitimate governance. The most significant challenge is that the sector-wide governance infrastructure needed to support responsible adoption is being built in fragments. Promising work is underway: the SAFE AI initiative is developing compliance and independent assurance tools for humanitarian AI deployments. NetHope has built AI readiness assessments, adoption toolkits, and training programs for organizations at different maturity levels. The ICRC and EPFL have produced lifecycle guidance. But these efforts are not yet connected. Each addresses a piece of the governance puzzle, but no coordinated mechanism brings them together into shared infrastructure the whole sector can use. For smaller organizations in particular, navigating multiple disconnected tools and standards is itself a barrier to responsible adoption. NetHope is working to build a coordinated sector effort to connect these emerging solutions into an interoperable governance ecosystem that reduces duplication, enables mutual recognition, and makes responsible AI adoption achievable for organizations of all sizes and capacities.

The AI Dialogue's unique value is as an inclusive intergovernmental forum. Multiple speakers at the first consultation rightly noted that the Dialogue must complement, not replicate, existing processes. From NetHope's perspective, three roles would make the Dialogue indispensable. First, it can serve as the connective layer between governance processes that currently operate in isolation. The OECD AI Principles, the EU AI Act, the Global Digital Compact, regional frameworks across Africa, Asia-Pacific, and Latin America, and sector-specific initiatives all generate governance activity. No mechanism currently connects them or addresses the gaps between them. The Dialogue is uniquely positioned to map where governance layers reinforce each other, where they conflict, and where critical sectors fall between mandates entirely. For the nonprofit sector, which operates across all of these jurisdictions simultaneously, this coordination function is not abstract. It determines whether governance is navigable or paralyzing. Second, it can institutionalize the inclusion of operational expertise in governance design. Current governance processes are shaped predominantly by the actors who develop AI and the governments that regulate it. The organizations deploying AI in crisis settings, fragile states, and conditions of power asymmetry hold knowledge that these processes need: what governance looks like when infrastructure is unreliable, when consent is complicated by dependency, and when algorithmic errors carry life-or-death consequences. The Dialogue should create permanent, structural pathways for this expertise to enter governance processes, not as periodic testimony, but as a continuous input to how governance is designed and evaluated. Third, it can establish shared expectations around regulatory interoperability. Organizations operating across borders need governance approaches that are compatible enough to enable compliance without requiring bespoke interpretation for every jurisdiction. The Dialogue can advance mutual recognition frameworks and shared compliance guidance that reduce the burden on resource-constrained organizations while maintaining rigorous protections for affected populations.

The AI governance landscape is rich with initiatives. The Dialogue's added value is not to create new ones, but to connect existing efforts into a more coherent architecture. Several deserve specific attention. At the intergovernmental level, the OECD AI Principles and their monitoring framework provide the most mature normative baseline. The Global Digital Compact established commitments that need operational follow-through. The newly launched International Scientific Panel on AI offers an evidence function the Dialogue should draw on directly. Regional frameworks, including the African Union Continental AI Strategy, ASEAN's AI Governance Guide, and Latin American declarations from Montevideo and Santiago, represent governance approaches designed for diverse national contexts that the Dialogue must integrate rather than override. At the sector level, the humanitarian and nonprofit community has built significant infrastructure the Dialogue should recognize and connect with. NetHope's AI Working Group convenes more than 60 INGOs actively working through governance challenges, from shared procurement standards to funder-grantee norms. The SAFE AI initiative, a collaboration between the CDAC Network and the Alan Turing Institute, is developing independent compliance and assurance tools for humanitarian AI deployments. The ICRC and EPFL have produced operational lifecycle guidance. OCHA's Centre for Humanitarian Data provides coordination-system-level evidence. The Principles for Digital Development, adopted by over 250 organizations, offer a technology governance baseline. These initiatives represent the emerging sector-wide governance layer that global processes need to engage with directly. The Dialogue's added value is threefold. It can provide a coordination function that no single initiative can perform on its own, connecting governance layers that currently operate in parallel. It can create legitimacy for sector-specific governance approaches by recognizing them within an intergovernmental process. And it can ensure that the second dialogue in New York in 2027 builds on concrete progress rather than repeating the same consultations. The sector is building. The Dialogue should build with it.

The Dialogue should be structured around the principle that governance is stronger when the people navigating its consequences help design it. Three structural recommendations: Organize stakeholder participation by governance function, not by stakeholder type. The standard model of government, industry, civil society, and academia panels produces predictable exchanges. Organizing sessions around specific governance functions, such as regulatory translation, operational tooling, capacity-building, and evidence generation, would surface practical knowledge from participants who are working on the same problems from different vantage points. Create submission pathways for operational evidence, not only policy positions. The written submission process appropriately solicits priorities and recommendations. It should also invite concrete evidence: deployment outcomes, governance failures, implementation lessons, and data on how existing frameworks perform in practice. This is the material that makes governance implementable. Organizations working at the sector level have this evidence. The process should ask for it explicitly. Ensure continuity between the Geneva dialogue and the 2027 New York follow-up. The co-chairs' summary should include specific commitments with named responsible parties and timelines, not only thematic conclusions. The sector organizations, coalitions, and networks that contribute to Geneva should have a structured role in tracking progress and informing the New York agenda. Governance processes that reset with each convening is not a good investment of any participant.

The nonprofit and humanitarian sector is systematically underrepresented in global AI governance discussions relative to its operational stake. More than 60 of the world's largest INGOs, collectively operating in 190 countries and serving hundreds of millions of people, deploy AI in conditions that stress-test every governance assumption: unreliable infrastructure, power asymmetry, cross-jurisdictional complexity, and life-or-death consequences for algorithmic errors. This operational reality is largely absent from the rooms where governance is designed. Within the sector, two groups are particularly underrepresented. First, local and nationally led organizations in the Global South. These organizations hold the most granular community-level data and have the least capacity for data protection. They are the most exposed to AI risks and the least resourced to govern them. Second, affected populations themselves. Accountability to communities is a foundational humanitarian principle, yet no global AI governance process has created meaningful mechanisms for the people whose lives are shaped by AI-driven decisions to inform how those systems are governed. The UN's global footprint offers a powerful pathway to close this gap. With offices in nearly every country, the UN system could use the period between the Geneva and New York dialogues for dedicated country-level engagement with Global South communities, local civil society, and national technology practitioners. Multilingual submission forms, AI-powered chatbot interfaces, and voice-based input collection could dramatically lower barriers to participation and ensure that governance reflects the perspectives of the communities it is ultimately meant to protect.

Two formats would meaningfully improve engagement. First, structured working sessions that produce outputs, not only discussion. Allocate time for cross-sector groups to draft shared language on specific governance challenges, such as regulatory interoperability, capacity-building benchmarks, or accountability mechanisms. Participants invest more when their contributions shape a tangible product. The co-chairs' summary should draw directly from these outputs. Second, pre-Dialogue collaborative drafting. Open specific sections of the co-chairs' summary for stakeholder input before the Geneva convening, not only after. This shifts participation from reactive commentary to co-creation, and it produces a stronger document because it incorporates operational knowledge at the drafting stage rather than attempting to retrofit it after the fact. Beyond format, the most important design choice is ensuring that participation carries consequences. If stakeholder input is visible in the outcomes, engagement will be substantive. If it is not, it will not be.

The following examples represent policies, practices, and platforms that are advancing effective AI governance, with particular relevance to the nonprofit and humanitarian sector. Governance frameworks and standards: The EU AI Act (eur-lex.europa.eu) established the first comprehensive risk-based AI regulation. The OECD AI Principles (oecd.ai) provide the most widely adopted intergovernmental normative baseline. ISO/IEC 42001 (iso.org) offers a certifiable AI management systems standard. The NIST AI Risk Management Framework (nist.gov) provides a flexible, voluntary risk governance structure. UNESCO's Recommendation on the Ethics of AI (unesco.org) is the first global AI ethics standard adopted by 194 member states. Sector-specific governance: NetHope's Humanitarian AI Code of Conduct (nethope.org) provides the most comprehensive sector-wide AI governance instrument for nonprofits. The Principles for Digital Development (digitalprinciples.org), adopted by over 250 organizations, offer a foundational technology governance baseline. The ICRC and EPFL's "From Principles to Practice" guide translates humanitarian norms into AI lifecycle guidance. The SAFE AI initiative (cdacnetwork.org/safe-ai), a collaboration between CDAC Network and the Alan Turing Institute, is developing independent compliance and assurance tools for humanitarian AI deployments. Operational tools and platforms: NetHope's AI Readiness for Nonprofits toolkit (nethope.org) provides assessments, training, and adoption guidance calibrated to organizational maturity. The OCHA Centre for Humanitarian Data (centre.humdata.org) provides coordination-level evidence infrastructure. The AI Incident Database (incidentdatabase.ai) enables cross-sector incident tracking and learning. Project Evident's Equitable AI Adoption Framework (projectevident.org) guides nonprofits through use-case-specific AI best practices. These examples demonstrate that effective governance is being built across multiple layers. The challenge is connecting them into shared infrastructure that the sector can use collectively.