Queen Mary University of London

The first Global Dialogue on AI Governance will be a success if it confronts the structural conditions shaping AI development rather than concentrate on ex-post regulation of AI harms. A successful Dialogue would go further on three fronts. First, it would reframe scale. Current industry pressures push relentlessly toward larger models, larger datasets, and greater use of environmental resources and energy. However, increasing evidence, including from our research, shows "bigger" is not better. Scaling down — building bounded, context-specific AI systems shaped by relevant communities — produces better, less resource-intensive, and more accountable outcomes. Domain-specific models are outperforming general-purpose ones in fields like medicine and journalism. Community-led datasets, built with care and consent, are more accurate than those assembled through automated scraping. A successful Dialogue would recognize that the "scale" of AI development is a choice, not a technical necessity. Second, it would develop strategies for meaningful public engagement in AI development and governance. Substantive participation requires creating bounded spaces — defined by geography, community, or domain — in which affected people can shape AI development from the start, not after key decisions have been made. This means facilitating longer timescales that respect the rhythms of deliberation and trust-building with diverse publics, rather than prioritizing the speed of industry deployment. Governance processes that privilege efficiency exclude exactly the voices most affected and undermine meaningful public accountability. Third, it would address questions of ownership and infrastructural power directly. The current landscape of AI development concentrates technical infrastructure and decision-making in the hands of a few corporations. A successful Dialogue would explore models — cooperatives, public institutions, commons-based approaches — that distribute both ownership and decision-making. Pushing forward across these fronts is paramount for building sustainable AI governance frameworks that confront harms and support democratic accountability.

• Transparency, accountability, and human oversight
• Open-source software, open data and open AI models
• Social, economic, ethical, cultural, linguistic and technical implications of AI

These three priorities reflect a shared concern: AI development is increasingly concentrated among a small number of well-resourced private actors who are also shaping the governance frameworks meant to hold them accountable. Addressing this requires going beyond industry consultation or ex-post regulation. Transparency, accountability, and human oversight are an urgent priority: Current practices make it extremely difficult for publics, regulators, or civil society to scrutinize how AI systems are built and deployed. Critically, regulatory power is also being shaped upstream: industry actors are actively contesting what kinds of governance are considered legitimate, substituting "industry-led standards" for binding public oversight. Accountability mechanisms must be substantive, enforceable, and insulated from regulatory capture. Social, economic, ethical, cultural, linguistic and technical implications of AI is also a key issue: General-purpose models trained at internet scale tend to extract what is common across contexts while erasing what is particular - disadvantaging communities whose languages, cultural frameworks, and social contexts are underrepresented. Bounded, community-led approaches can produce safer, more accurate, and more accountable systems. These lessons must inform governance, not just technical development. Open-source software, open data and open AI models are critical tools for challenging infrastructure concentration. When properly governed, openness enables wider participation, supports local adaptation, and reduces dependence on proprietary systems. Governance should actively support non-commercial and cooperative alternatives, including through strategic public investment.

Two structural issues cut across all the listed themes and deserve explicit attention. First, the concentration of AI development. The listed themes largely address how AI is used and governed, but not who builds it, who owns it, and on what terms. The dominant trajectory toward larger models, greater compute requirements, and centralized infrastructure systematically favors private actors with access to significant capital and infrastructure - and those same actors are actively shaping regulation and governance to protect their position. Closing digital divides and achieving the SDGs requires directly addressing who controls AI infrastructure. The Dialogue should explore alternatives such as cooperative ownership models, public compute infrastructure, and strategic public investment in community-led AI development, rather than directing public resources toward deepening corporate concentration. Second, the limits of consultation as participation. Company-run participatory initiatives have largely amounted to consultation on discrete implementation decisions, prioritizing scale of participation over depth or duration of dialogue, without substantively shifting decision-making power or ownership. Meaningful inclusion - particularly for the Global South and affected communities - requires not just a seat at the table but the capacity and authority to shape AI development from the start. Evidence from community-based processes shows that bounded, participatory approaches, given adequate time and resources, produce better and more accountable AI systems than those developed at speed and scale by a handful of corporations. Without addressing these dynamics, progress on every listed theme risks being undermined by the concentration of power that shapes how AI is developed and who benefits from it.

We write from academic research institutions in the Global North, working on AI development and governance. The governance gaps we have identified are acutely felt in this context. The concentration of AI development has reshaped academic research itself. Industry control over AI supply chains means that academic institutions increasingly work within an industry-defined problem space, shaping which questions are deemed worth asking and which answers result in grants, awards, and tenure. The percentage of AI PhD graduates hired by industry increased from around 40% in 2011 to 70% in 2022, representing a significant drain of public investment in research talent toward private ends. Industry players co-author research papers, fund academic work, and organize leading conferences, with measurable effects on the values that dominate the field: studies show an emphasis on performance, generalization, and efficiency, with comparatively little attention to informed consent, autonomy, or participation. This matters for governance. Knowledge production shapes what problems are visible, what solutions are imaginable, and whose harms are counted. Governance frameworks that rely heavily on industry-adjacent research risk inheriting these blind spots.

The Dialogue should build on initiatives that demonstrate, in practice, what democratic and community-led AI development looks like, rather than starting from scratch or defaulting to industry frameworks. Data Against Feminicide (DCF) is one such initiative. A South-North participatory action research and design project, DCF has spent five years collaborating with civil society groups across North America, Latin America and Sub-Saharan Africa to co-develop AI tools that support data activism against gender-related violence. Its AI-powered email alert platform — trained through participatory processes with civil society communities and available in Spanish, English, Portuguese, and Swahili — is currently used by dozens of projects across more than 15 countries. DCF illustrates several principles relevant to the Dialogue: that bounding AI development to specific communities and contexts produces more accurate and more accountable systems; that participatory development requires time, trust-building, and ongoing collaboration rather than one-off consultation; and that AI tools can be designed to support and not replace the expertise of affected communities. DCF also highlights persistent challenges that governance frameworks must address, including the difficulty of sustaining community-oriented tools when funding agencies favor innovation over maintenance. (Project: https://datoscontrafeminicidio.net/en/home/) Other relevant initiatives include the Masakhane African Languages Hub, which builds AI tools for African languages through community-led, participatory processes, and READ-COOP, a European cooperative governing AI infrastructure for historical document recognition with democratic member ownership across 35 countries. The added value the Dialogue could bring is to connect these dispersed initiatives, draw out the governance principles they embody, and build the case — with evidence — for redirecting public support toward models that prioritize participation, context-sensitivity, and community ownership over scale and speed.