Aqta

A successful first Global Dialogue on AI Governance would move from general principles to a small number of concrete, widely supported next steps that states, industry and civil society can actually implement. It should produce: (1) a shared, high‑level threat and opportunity map for frontier and deployed AI systems, aligned with existing efforts such as the G7 Hiroshima Principles and the UN "Governing AI for Humanity" work. (2) a short menu of interoperable governance tools for example, model and incident reporting templates, evaluation and red‑teaming norms, and baseline transparency expectations for high‑risk AI systems that different jurisdictions can adapt without fragmenting the ecosystem. It should also result in a clear mandate and workplan for the Independent International Scientific Panel on AI, including how its evidence will feed into subsequent Dialogues and national policymaking. Finally, success would mean meaningful participation from low‑ and middle‑income countries and marginalised communities, not only well‑resourced actors so that Global South capacity‑building and local data/compute needs are reflected in priorities from the beginning rather than retrofitted later.

• Safe, secure and trustworthy AI
• Transparency, accountability, and human oversight
• Interoperability of governance approaches
• Protection and promotion of human rights

Our work focuses on making AI systems safe, governable and usable in high-stakes, real-world contexts so our priorities mirror that. We prioritise safe, secure and trustworthy AI because many of the most concerning failure modes arise from deployed systems that behave unpredictably under adversarial use, distributional shift, or emergent capabilities. Governance needs to be linked to technical realities: robust evaluations, incident reporting, and secure deployment practices for both frontier and widely deployed models. We emphasise transparency, accountability, and human oversight because oversight is only meaningful when practitioners, regulators and affected communities can understand how systems are being used, what risks were foreseen, and how they are being mitigated in practice. That includes documentation of model purpose and limitations, clear lines of responsibility, and mechanisms for contestability and redress. The protection and promotion of human rights must remain a non-negotiable foundation, especially as AI is increasingly embedded in public services, health, finance and welfare. AI governance should reinforce existing human rights obligations, ensuring that AI deployment does not exacerbate discrimination, surveillance or exclusion, particularly for already marginalised groups. Finally, we see interoperability of governance approaches as essential to avoid a fragmented landscape where different standards in different regions create both gaps and unnecessary friction. Interoperability does not mean uniformity, but rather common reference points for example, shared taxonomies of risk, compatible reporting formats, and mutual recognition of certain assurance mechanisms so that safeguards can scale across borders and supply chains

Two cross-cutting issues we see as under-emphasised are runtime governance of AI systems and the security of AI infrastructure and supply chains. Today's discussions focus on model development and high-level principles, but many harms occur at runtime: when models are composed into tools, agentic workflows, or socio-technical systems that behave unpredictably. We need more attention on continuous monitoring, policy-enforcement and kill-switch capabilities at the point of use, supported by technical standards for audit logging, safety-policy enforcement layers, and human-in-the-loop escalation in high-risk operations. AI security is now system security, spanning data pipelines, model weights, deployment infrastructure, and downstream integrations. Emerging risks include model theft and exfiltration, training-data poisoning, supply-chain compromise of widely reused components, and targeted attacks on safety tooling itself. Governance discussions should explicitly integrate AI system-security baselines, secure development lifecycles for AI, incident-response mechanisms, and norms for cross-border vulnerability disclosure rather than treating security as a separate track. Both issues cut across the listed themes, linking safety, human rights and interoperability with practical questions of implementation: who is responsible for what, what technical and organisational capabilities are required, and how safeguards can keep pace with deployment speed.

The AI governance infrastructure sector sits at the intersection of current gaps. We build the tools that make governance operational, and we feel the consequences of incoherence directly. The main challenge is fragmentation. The EU AI Act sets comprehensive obligations for high‑risk AI, but organisations operating across jurisdictions have no clear path to unified compliance. Engineering teams building governance tooling must design for incompatible audit formats, different definitions of "high‑risk", and divergent oversight requirements at the same time. This raises compliance costs, concentrates capacity in large organisations that can absorb legal and technical overhead, and leaves smaller teams in research and civil society effectively ungoverned by default. The second challenge is the speed gap. Agentic AI systems where models plan, execute and spawn sub‑tasks autonomously are already deployed. Existing frameworks were not designed for systems that act across long sequences without per‑step human review, leading to runaway loops, compounding errors and unaudited resource use. The opportunity is that many solutions already exist. Cryptographic audit trails, zero‑knowledge compliance verification, real‑time policy enforcement at the inference layer and automated AI‑Act article mapping are deployable today. The gap is not technical capacity but governance frameworks that recognise and require these mechanisms. The Dialogue can close this by grounding outputs in what is technically achievable now, not only in aspirational futures.

The Dialogue's main value should be coordination, not inventing new standards. Standards bodies already exist; what is missing is a mechanism that makes them work together. First, mutual‑recognition brokering. The Dialogue can support bilateral and multilateral agreements so that compliance with one jurisdiction's AI framework is legible in another. This reduces duplication without forcing full harmonisation. Second, gap identification. It can publish an annual map of what existing frameworks cover and where gaps remain. Agentic systems, multi‑model pipelines and privacy‑preserving compliance methods are examples of areas that are still poorly covered. Third, capacity transfer with accountability. The Dialogue can connect organisations in lower‑income countries directly to technical infrastructure, not just advice, and track whether support leads to deployed governance capability. It fails if it produces another declaration that bodies may voluntarily adopt. It succeeds if it creates durable coordination mechanisms with named owners and review cycles.

Four initiatives deserve direct connection. EU AI Act implementation is the most detailed operational governance framework in use. The Dialogue should treat it as a reference implementation and seek interoperability with it, not design around it. The NIST AI Risk Management Framework is complementary: more flexible and less prescriptive. A mutual‑recognition arrangement between these two would already cover a large share of current AI deployment. The UNESCO Recommendation on the Ethics of AI has been adopted by almost all member states. The Dialogue should help convert its principles into measurable technical and organisational requirements instead of drafting a parallel ethical text. Finally, ITU's AI for Good platform already brings together technical experts and Global South actors. Routing capacity‑building commitments through such existing channels is more efficient than creating new administrative structures. The added value the Dialogue can bring is binding coordination between these initiatives: shared definitions, cross‑referenced requirements and joint review cycles that reduce fragmentation without replacing existing frameworks.

The Dialogue should be structured in three tracks running in parallel, not sequentially. A technical track for engineers and infrastructure builders – the people actually deploying AI systems. Governance forums are dominated by legal and policy voices, producing frameworks that are coherent on paper and unimplementable in practice. Technical practitioners should co‑author outputs, not just present to them. A civil‑society track with binding input mechanisms, not observer status. Affected communities should shape agenda items before the Dialogue convenes, not comment after positions are set. A regulatory track for coordination between existing bodies – the EU AI Act, US executive frameworks, and emerging national legislation. Outputs should include a mutual‑recognition roadmap, not another principles document. Format recommendation: publish all submissions and draft outputs publicly before the Dialogue opens, with a simple way to comment. Closed consultations produce outcomes that only participants can interrogate; an open record makes it easier for practitioners and affected communities to hold decision‑makers to their own standards.

Three groups are consistently absent. First, engineers and product teams deploying AI in production. Policy is written about their systems without their input, yet they understand failure modes and operational constraints that governance bodies have not yet named. Second, organisations in the Global South, not as recipients of capacity‑building, but as co‑authors of governance frameworks. AI governance shaped entirely by the EU and US will encode those regions' assumptions as defaults. Third, small and mid‑sized organisations. Current compliance frameworks are calibrated for enterprises with legal teams, but most AI deployment happens in entities that cannot absorb that overhead. Inclusion mechanism: funded participation, not just open invitations. Travel, time and translation support should be budgeted as core governance infrastructure. Access without resources is not access.

Live red‑teaming sessions: have technical teams attempt to break proposed governance frameworks and safety mechanisms in real time, then publish the results. This surfaces gaps faster than written consultation. Structured disagreement panels: instead of consensus presentations, explicitly seat opposing technical and policy views and document where they diverge and why. Rolling public comment with version control: treat output documents like open‑source software, public drafts, tracked changes, and attributed contributions over time. Finally, simulation exercises where regulators, engineers and civil‑society actors role‑play incident response to concrete AI failure scenarios. This forces governance proposals to be tested against realistic time pressure and institutional constraints.

Inference-layer governance is the most effective pattern. Governance that sits between the application and the model, with policy enforcement, loop detection, budget controls and automatic audit-trail generation, can block violations in real time and create tamper-evident records without relying on periodic manual checks. The EU AI Act's risk-tiered conformity-assessment model is currently the most operationally coherent framework, distinguishing prohibited, high-risk and lower-risk systems. Its main limitation is that it was designed around large organisations. A simplified pathway for smaller deployers, supported by open-source tooling and templates would extend its reach without lowering standards. Privacy-preserving audit mechanisms address the tension between logging and data protection. Zero-knowledge proof systems make it possible to verify that specific policies were enforced without exposing underlying prompts or personal data. These techniques are deployable today and should be explicitly recognised as valid compliance-verification methods. Across all three, the common lesson is that effective governance must be implemented as technical infrastructure, not only as process. Approaches that can be deployed as a layer around existing systems will spread faster and deliver more consistent protection than ones that depend on changing every team's internal practices.