The Centre for Artificial Intelligence in Government at University of Birmingham

Success would require the Dialogue to move beyond reaffirming shared principles, which existing frameworks such as the UNESCO Recommendation and the OECD AI Principles have already achieved, and instead produce concrete governance infrastructure. Three outcomes would mark genuine progress: A shared diagnostic baseline. Participating states currently lack comparable data on the AI systems operating within their borders. A successful Dialogue would commit to developing common methodologies for auditing AI systems, not only for technical safety, but for the governance values and political assumptions they express. Without this, international coordination rests on incompatible national assessments of the same systems. An asymmetry mechanism. The Dialogue risks reproducing existing power imbalances if its outputs are designed around the capabilities of states that already have advanced AI industries. Success requires a concrete mechanism through which technically advanced states share auditing tools and capacity with countries that are primarily AI adopters rather than developers. A living institutional process. A successful one establishes a review process with defined timelines, reporting obligations, and a secretariat function that can track implementation. The 2026 session should end with a clear answer to the question: who is responsible for following up, and by when? Underlying all three is a more fundamental test: whether the Dialogue can hold together states with genuinely different views of what trustworthy AI governance means. That tension should be named and worked through, not papered over with language broad enough that every delegation can claim victory. A Dialogue that surfaces real disagreements and creates institutional space to work through them will have done more for long-term governance than one that produces consensus at the cost of clarity.

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

My selections reflect an underlying argument: that effective AI governance requires tools to see inside AI systems, not only rules applied to their outputs. Safe, secure and trustworthy AI is the foundational priority. Trustworthiness is currently defined largely in technical terms, robustness, reliability, absence of harmful outputs. My research suggests this is incomplete. AI systems may systematically express different governance values depending on their developer origin, meaning that a technically safe system can still embed political assumptions that undermine the governance goals of deploying states. Expanding the definition of trustworthiness to include value transparency is urgent. Transparency, accountability, and human oversight follow directly. Meaningful accountability requires that deploying states, regulators, and affected communities can identify what values and assumptions are built into the AI systems they use. Current transparency frameworks focus on data provenance and model architecture; they do not yet address the political and governance values that emerge from training choices. This gap needs closing. Interoperability of governance approaches is where the geopolitical stakes are highest. My comparative work on US- and Chinese-developed models suggests that AI systems from different national contexts express systematically different governance values. Any interoperable governance framework must reckon with this: convergence on principles is insufficient if the underlying systems reflect divergent assumptions about the relationship between individuals, states, and markets. Social, economic, ethical, cultural and technical implications provide the broader frame. The value differences I observe across AI systems are not merely technical artefacts; instead, they reflect and potentially reinforce different political cultures. As AI systems are adopted globally, these embedded assumptions travel with them, with implications for political pluralism and cultural autonomy that governance frameworks have not yet adequately addressed.

One cross-cutting issue is conspicuously absent: the governance values embedded in AI systems themselves. The listed themes address how AI should be governed through transparency, accountability, interoperable frameworks, and human oversight. But they do not address a prior question: what governance values do AI systems already express, and do these vary systematically by developer origin? This is not a technical safety question. It concerns the political and ideological assumptions built into AI systems through training data, fine-tuning procedures, and reinforcement learning from human feedback. This matters for the Dialogue in at least two ways. First, if AI systems from different national contexts embed different assumptions about the appropriate relationship between individuals, states, and markets, then deploying those systems in public administration, education, or social services is not a neutral act. The governance values of the system interact with and may quietly reshape the governance values of the adopting context. Second, interoperability of governance frameworks is much harder to achieve if the AI systems those frameworks are meant to govern are themselves carriers of divergent political assumptions. Governance frameworks applied to AI outputs cannot resolve value disagreements that are embedded in the systems themselves. A related emerging issue is the absence of agreed methodologies for value auditing. Technical safety evaluation is relatively mature; tools for systematically identifying the political and governance values expressed by AI systems are not. Without such tools, transparency commitments remain aspirational, what cannot be measured cannot be disclosed. The Dialogue would benefit from explicitly naming value transparency, distinct from technical transparency, as a governance priority, and from commissioning work on common methodological standards for its assessment.

The governance gaps identified above manifest differently depending on a country's position in the global AI supply chain, but the underlying problem is shared. For AI-importing advanced economies including the United Kingdom, the challenge is one of hidden assumptions. The UK has invested significantly in AI safety evaluation, yet existing frameworks assess whether AI systems cause harm, not what governance values they express. Public bodies, regulators, and educators increasingly deploy AI systems built elsewhere, with no established means of auditing whether those systems embody assumptions consistent with UK democratic norms and legal frameworks. The opportunity is that advanced economies have the technical capacity to develop value auditing methodologies and a strong interest in doing so as AI penetrates public administration. For developing countries, the asymmetry is sharper. Most nations in the Global South are primarily AI adopters, not developers. They deploy systems whose training data, fine-tuning procedures, and reward signals were determined by a small number of actors in the US and China. Without auditing tools or the technical capacity to use them, these countries cannot know whether the AI systems shaping their public services reflect their own governance values or those of the countries that develop the systems. This is not a hypothetical concern; it is a structural feature of the current AI supply chain. The opportunity the Dialogue presents is to reframe this as a collective action problem rather than a bilateral one. No single importing country, advanced or developing, has sufficient leverage to demand value transparency from frontier AI developers alone. A multilateral mechanism that pools auditing capacity and establishes common disclosure standards would shift that balance, and the Dialogue is well placed to initiate it.

The Dialogue's most valuable role is one that no bilateral or regional forum can play: creating a genuinely universal space in which states with fundamentally different AI governance philosophies are nonetheless obliged to engage with the same questions. Convened under the UN, the Dialogue carries a legitimacy that regional and bloc-specific initiatives, however useful, cannot claim. That legitimacy can be used in three ways. First, as a bridge between governance blocs: the Dialogue can create structured engagement between states operating under fundamentally different assumptions about AI including assumptions about the appropriate role of the state, individual rights, and market governance without requiring premature convergence. Naming disagreements clearly is more productive than obscuring them behind consensus language. Second, as a standard-setter for process rather than substance: where agreement on AI governance principles remains elusive, the Dialogue can establish shared methodological standards for auditing, for transparency reporting, for impact assessment that all parties can adopt without requiring ideological alignment. Third, as an accountability anchor: the Dialogue can establish expectations against which national and regional AI governance frameworks are measured, creating reputational incentives for compliance even in the absence of enforcement mechanisms. The Dialogue has a unique opportunity to become the place where genuine international coordination across blocs, beyond existing frameworks, and on terms that all states can own finally begins.

One concrete methodological approach worth highlighting is systematic value auditing of AI systems using a battery of policy vignettes grounded in established cross-national survey instruments. This approach, developed in my research, is designed to elicit the governance values expressed by AI systems across key dimensions including output legitimacy versus procedural legitimacy, individual versus collective rights, state versus market, and equality versus efficiency in a way that is both empirically grounded and comparable across national contexts. Unlike technical safety evaluations that assess whether AI systems cause harm, value auditing assesses what political and governance assumptions AI systems express by default. This distinction matters for policy: a system can pass all standard safety benchmarks while still systematically favouring one governance philosophy over another. The approach is transferable. It can be applied to any AI system, adapted to different governance contexts, and scaled to comparative cross-national analysis. It offers regulators, procuring governments, and international bodies a practical tool for the kind of value transparency that current frameworks call for but do not yet know how to measure.