UCL

A successful first Global Dialogue on AI Governance would produce outcomes that change who is in the room, what knowledge counts, and how it is used. It should establish new, durable channels of communication that go beyond major AI firms and their home governments, ensuring meaningful participation from low‑ and middle‑income countries, civil society, technical communities, and groups most affected by AI systems. The Dialogue should also generate a shared understanding that governance practices are inseparable from context and actors. Rather than compiling a de‑contextualised "menu" of tools, it should document who did what, for whom, under which political, economic, and institutional conditions. This would help participants avoid copying policies that only work under very specific circumstances, and instead support adaptation that is sensitive to local realities. Finally, as a rare convening space at UN level, the Dialogue should curate and critically assess governance approaches that have demonstrably succeeded or failed in multicultural, multilingual, and multisector settings, with explicit reference to how they advance or undermine the Sustainable Development Goals. Success would mean leaving with a focused set of high‑leverage practices and open questions, not just a restatement of existing principles. If the Global Dialogue cannot clearly show how it adds this kind of practical, equity‑oriented value beyond existing forums, it will have missed its opportunity.

• Protection and promotion of human rights
• Transparency, accountability, and human oversight

While all these themes are interconnected, we believe it is important that the UN provide a human-centred view and place the impact on people and society at the heart of conversations on technological development.

In data‑sensitive, multi‑stakeholder environments such as during humanitarian responses, governance gaps stem less from a lack of high‑level principles and more from failures to translate them into situated practices that reconcile divergent stakeholder values. Existing risk‑centric approaches emphasise technical failures (hallucinations, security, adversarial attacks) but underplay misalignments between developers, humanitarian agencies, and affected communities over what counts as "safe enough", who is accountable, and whether tools entrench structural inequalities. In humanitarian and crisis‑affected settings, these gaps manifest acutely: tools decide who receives aid or information, yet those most affected have the least influence over design and governance. Equity‑related rifts are particularly significant, with stakeholders disagreeing on whether AI tools reproduce data colonialism and neocolonial narratives about conflict‑affected populations, or instead help redress power imbalances. Accountability is further complicated when humanitarian organisations are held responsible for AI‑mediated decisions generated by proprietary systems they cannot fully audit. In terms of opportunities, through practice-based research, Shivaang Sharma and Angela Aristidou have developed frameworks like SHARE (safety, humanity, accountability, reliability, equity) provide concrete diagnostic questions that humanitarian organizations and their technology partners can use to surface these misalignments, or what they term 'AI responsibility rifts' (AIRRs) early, include marginalised voices, and re‑design tools or processes before harms crystallise. This points to an opportunity for sectors operating in fragile contexts to move from principle‑setting to structured, participatory, and iterative governance practices.

Global cooperation should complement the dominant AI risk lens with an explicit focus on AI responsibility rifts (AIRRs) and the lived experiences of those governed by AI systems. An effective AI Dialogue could institutionalise this shift by making AIRRs and SHARE‑like dimensions a standing pillar of multilateral discussion, alongside safety, security and innovation. First, the AI Dialogue could promote common methodologies for identifying and addressing AI responsibility rifts across borders, e.g. shared self‑diagnostic tools, stakeholder mapping templates, and minimum expectations for participatory design in data‑sensitive deployments. Second, it could broker learning between regions and sectors where these challenges are most acute (such as humanitarian action in Gaza and other conflict‑affected settings) and jurisdictions developing binding regulation, ensuring that global frameworks are informed by operational realities rather than only by corporate or Global North regulatory perspectives. Finally, by design, the AI Dialogue could require that deliberations include representatives of affected communities and local implementers, not only states and technology firms, thereby embedding AIRR‑sensitive governance into the UN architecture itself.

Beyond UN processes, partnerships and networks such as those focused on responsible AI governance, humanitarian innovation, and community‑led AI oversight offer concrete tools and case studies on multi‑stakeholder engagement, accountability infrastructures and rights‑based approaches. The SHARE framework developed by Shivaang Sharma and Angela Aristidou specifically provide an empirically grounded, cross‑cutting lens – safety, humanity, accountability, reliability and equity – that can be applied across sectors to interrogate how existing principles are operationalised. The AI Dialogue's added value would be to: • Elevate AIRR‑sensitive approaches like SHARE from project‑level tools to reference models for UN‑level governance discussions. • Connect humanitarian and other high‑risk sectors with regulators and standards bodies so that operational lessons (e.g. around equity‑related rifts and data colonialism) feed back into global rule‑making. • Provide a platform for convergence across initiatives, distilling a small set of interoperable expectations on participatory design, multi‑stakeholder accountability and equitable impact assessment that can be recognised across regimes. In essence, the AI Dialogue can turn a fragmented landscape of principles and pilots into a more coherent, practice‑oriented ecosystem anchored in both global norms and local experiences of AI governance.

Underrepresented voices include: crisis‑affected communities, local civil society and grassroots organizers, front‑line humanitarian workers, ethicists embedded in local institutions, and representatives of groups historically subject to surveillance and discrimination. Their perspectives are essential to evaluating humanity‑related rifts (e.g. dehumanisation, loss of meaning in work) and equity‑related rifts (e.g. whether tools deepen or reduce inequalities). To include these perspectives, the AI Dialogue could: • Mandate participatory AI design and evaluation practices (as reflected in the SHARE framework's equity and humanity dimensions) as a norm for projects discussed in UN fora. • Require that case studies presented to the Dialogue document how local stakeholders influenced design choices and how disagreements (AIRRs) were surfaced and addressed over time. Such steps would move global AI governance away from purely expert‑driven narratives towards shared, negotiated understandings of responsible AI grounded in lived experience.

Sharma and Aristidou's work offer an approach rather than a formal policy: the AI responsibility rifts lens and the SHARE framework, which together function as a structured practice for diagnosing and addressing gaps in responsible AI operationalisation. SHARE's five dimensions - safety, humanity, accountability, reliability, equity - are implemented through a self-diagnostic questionnaire tailored separately for non-developer stakeholders and AI developers, prompting reflection on data safeguards, sense of distance from decision contexts, clarity of accountability, fail-safes and explainability, and inclusiveness of affected voices. This tool is already being used by humanitarian organizations to identify where disagreements block deployment and to prioritize governance interventions, such as during crisis response in Gaza. Broader approaches also include: • Comprehensive, multi-stakeholder safety evaluations that incorporate vulnerable and non-technical stakeholders from the outset, rather than relying solely on technical benchmarks. • Participatory AI design processes that treat local communities and ethicists as co-designers of data collection and usage protocols, helping preserve human meaning in work and respecting cultural values. • Visible monitoring and logging mechanisms that clarify accountability for AI-assisted decisions in complex workflows and enable contestation. • Equity-oriented impact reflection, explicitly probing whether tools deepen or reduce structural inequalities and colonial legacies, rather than treating "bias mitigation" as a purely technical afterthought. These practices can be translated into governance requirements (e.g. procurement standards, audit criteria, or conditionalities in donor funding) that make the identification and management of AIRRs an integral part of AI system lifecycle management, especially in high-risk, data-sensitive sectors. Citation of this study: Sharma, Shivaang and Aristidou, Angela (2025) "How Stakeholders Operationalize Responsible AI in Data-Sensitive Contexts," MIS Quarterly Executive: Vol. 24: Iss. 2, Article 4. Available at: https://aisel.aisnet.org/misqe/vol24/iss2/4