SAFE

A successful first Global Dialogue on AI Governance would move beyond principles into practical alignment on how safety is defined, measured, and verified in real-world systems. At present, there is broad agreement on the importance of responsible AI, but no consistent mechanism to independently validate how systems actually behave once deployed. This creates a gap between policy intent and lived reality, particularly in high-risk environments involving children and vulnerable populations. Success would mean establishing a shared understanding that governance cannot rely solely on self-declaration or retrospective harm reporting. It requires the ability to detect, interpret, and respond to risk as it emerges. Key outcomes should include: * Recognition of the need for independent validation layers that sit between policy and product. * Agreement on minimum viable safety signals that can be detected across systems without over-reliance on content access * Commitment to developing interoperable frameworks that allow safety to be consistently assessed across jurisdictions * Inclusion of frontline perspectives (educators, legal practitioners, child safety experts) alongside technical and policy voices Importantly, success is not a fully formed global standard at this stage, but a clear pathway toward one — grounded in evidence, collaboration, and real-world testing. If this dialogue can shift the focus from high-level intent to verifiable safety in practice, it will create the foundation for meaningful, scalable governance.

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

The selected priorities reflect a core gap in current AI governance: while there is strong global focus on responsible AI principles, there is still no consistent way to independently verify how systems behave in practice. Safe, secure and trustworthy AI is the outcome we are all working toward, but trust cannot be assumed, it must be demonstrated. This directly connects to transparency, accountability, and human oversight, where the current reliance on self-reporting by platforms limits meaningful assurance. Protection and promotion of human rights is central, particularly in environments involving children and vulnerable users, where the consequences of delayed intervention are significant and often irreversible. Interoperability of governance approaches is critical because AI systems operate across jurisdictions, yet safety expectations and enforcement mechanisms remain fragmented. Without shared, verifiable frameworks, protection becomes inconsistent and difficult to scale globally. From our perspective, urgent action is needed to bridge the gap between policy intent and real-world outcomes. This includes developing mechanisms that can detect, interpret, and respond to risk as it emerges, rather than relying solely on retrospective analysis. Together, these priorities support a shift from high-level principles toward practical, measurable, and globally applicable approaches to AI safety and governance.

One critical cross-cutting issue is not explicitly captured: the absence of independent, real-time validation of AI system behaviour in live environments. Current governance approaches focus heavily on principles, model development, and post-incident accountability. However, there is still no consistent mechanism to detect and verify how AI-mediated systems behave once deployed, particularly as they interact dynamically with users over time. This creates a structural gap between policy intent and lived outcomes. Safety is often defined by the same entities that design and operate the systems, and assessed retrospectively after harm has occurred. As a result, early risk signals, especially behavioural patterns that precede harm are frequently missed. An emerging priority is the ability to identify and respond to risk as it forms, without over-reliance on intrusive data access. This includes developing privacy-preserving methods to detect behavioural shifts, escalation patterns, and other indicators of emerging harm in real time. This issue cuts across all existing themes: * It underpins trust and safety * It strengthens accountability and oversight * It directly impacts the protection of human rights * It requires interoperable approaches across systems and jurisdictions Without this layer, governance remains reactive and difficult to enforce consistently. Addressing this gap would enable a shift from retrospective response to proactive, verifiable safety which is essential as AI systems become more adaptive, autonomous, and embedded in everyday life.

Governance gaps in AI are most visible at the point where systems interact with people in real time, particularly in environments involving children. In practice, there is no independent way to verify how platforms and AI-mediated systems behave once deployed. Safety is largely defined and reported by the same entities that design and operate these systems, creating a reliance on self-assessment and retrospective response. This limits the ability of regulators, educators, and families to identify risk early or intervene before harm escalates. This challenge is not confined to one country. While Australia has taken a proactive regulatory approach, similar gaps exist globally across jurisdictions with varying policy maturity but shared dependence on platform self-reporting and post-incident enforcement. The most significant challenge is the disconnect between policy intent and real-world outcomes. Without mechanisms to independently detect and validate system behavior, accountability is difficult to operationalise in a consistent and scalable way. At the same time, this presents a global opportunity. There is increasing alignment across governments, institutions, and communities on the need for safer digital environments, particularly for children. Advances in privacy-preserving technologies and behavioral signal detection create the potential to move from reactive models toward earlier identification of risk. For our sector, this is an opportunity to develop approaches that enable safety to be measured, tested, and validated in practice, supporting more consistent, interoperable, and globally applicable governance outcomes.

The AI Dialogue can play a critical role in moving international cooperation from shared principles toward practical alignment on how AI systems are governed in real-world environments. At present, there is broad global consensus on the importance of safe, trustworthy, and human-centered AI. However, cooperation is limited by fragmentation in how these principles are interpreted, implemented, and assessed across jurisdictions. Without a shared way to evaluate system behaviour in practice, it is difficult to achieve consistent accountability or build mutual trust between countries. The Dialogue can help address this by focusing on interoperability and not only at the level of policy, but at the level of verification. This includes advancing common approaches to how safety is defined, what constitutes meaningful risk signals, and how these can be assessed in ways that are both privacy-preserving and applicable across different regulatory environments. It also provides an opportunity to bring together perspectives that are often siloed, including policymakers, technical experts, and frontline practitioners to ensure that governance frameworks are informed by both system design and lived experience. Importantly, the goal is not to impose a single global standard, but to create sufficient alignment so that safety expectations are consistent, measurable, and transferable across borders. By anchoring cooperation in verifiable outcomes rather than abstract commitments, the Dialogue can help build the foundation for more effective, scalable, and trusted AI governance internationally.

There are several important initiatives already shaping the global AI governance landscape, including the OECD AI Principles, UNESCO's Recommendation on the Ethics of AI, the G7 Hiroshima AI Process, and emerging regulatory frameworks such as the EU AI Act. In parallel, national bodies such as online safety regulators, data protection authorities, and organisations working on child rights and digital safety are contributing valuable domain-specific expertise. These efforts have established critical foundations, particularly in defining principles, rights, and policy directions. However, they remain largely focused on guidance, regulation, and high-level alignment, with limited mechanisms to independently assess how AI systems behave once deployed in real-world environments. The AI Dialogue can add value by acting as a bridge between these existing frameworks and practical implementation. Specifically, it can support the development of interoperable approaches to verification — enabling different jurisdictions and sectors to move toward shared, evidence-based methods of assessing safety, accountability, and risk. It can also create a space where policy, technical, and frontline perspectives are integrated more effectively. Many current initiatives operate in parallel, with limited feedback loops from those directly experiencing harm or implementing safeguards on the ground. By connecting these efforts and focusing on verifiable outcomes, the Dialogue can help move the ecosystem from fragmented governance toward more cohesive, actionable, and globally relevant approaches.

Effective participation in the AI Dialogue requires moving beyond representation toward meaningful contribution from different stakeholder groups. Governments can provide policy direction and regulatory alignment, but also benefit from clearer insight into how systems perform in practice. Technical stakeholders can contribute to defining what is measurable and feasible, particularly in areas such as risk detection, system behaviour, and privacy-preserving methods. Civil society, educators, and legal practitioners bring critical frontline perspectives, identifying how risks emerge and where current protections fail in real-world settings. Industry plays a key role in Implementation, but should be engaged alongside independent perspectives to ensure balanced input. To be effective, the Dialogue should be structured as a working process rather than a series of high-level discussions. This could include: * Focused working groups aligned to specific challenges (e.g. risk detection, verification methods, child safety, interoperability) * Evidence-based sessions where real-world case studies and system behaviours are examined, not just policy positions * Cross-sector collaboration loops that connect policy, technical, and frontline insights, ensuring continuous feedback between design and impact * Clear outputs and milestones, such as draft frameworks, shared definitions, or pilot initiatives that can be tested across jurisdictions * Importantly, participation should prioritise those actively working within these challenges, rather than being limited to formal representation. A structured, iterative approach grounded in real-world testing and measurable outcomes will enable the Dialogue to move from discussion to practical progress in AI governance.

Global discussions on AI governance are still disproportionately shaped by those designing, regulating, or commercialising the systems rather than those encountering their consequences in real time. One of the most significant gaps is the absence of frontline voices: educators, child safety practitioners, legal advocates, and those working directly with individuals and families experiencing harm. These groups have early visibility into how risk emerges, escalates, and is often missed — yet their insights are rarely embedded into the design of governance frameworks. Young people, as primary users of many AI-mediated environments, are also underrepresented. Their lived experience is critical to understanding behavioural influence, vulnerability, and system interaction in ways that cannot be inferred from policy or technical design alone. In addition, contributions from the Global South remain uneven. While these regions are increasingly central to AI adoption and innovation, their perspectives are not consistently reflected in how governance approaches are defined, tested, or scaled. A deeper structural issue is the fragmentation between disciplines. Policy, technical, legal, and social perspectives often operate in parallel, with limited integration. This results in governance that is well-intentioned but difficult to operationalise in practice. Inclusion, therefore, must be designed into the process. This means creating structured roles for frontline expertise, enabling continuous rather than one-off engagement, and establishing feedback loops between real-world outcomes and governance design. Without this, AI governance risks remaining abstract and disconnected from the environments it is intended to protect.

To foster meaningful and dynamic engagement, the AI Dialogue should move beyond traditional panel discussions toward formats that enable real problem-solving, cross-sector interaction, and evidence-based collaboration. One effective approach is scenario-based working sessions, where participants engage with realistic use cases (e.g. child safety risks, misinformation spread, or automated decision-making impacts) and work together to identify how governance mechanisms would function in practice. This grounds discussions in real-world complexity rather than abstract principles. Live system walkthroughs or "governance simulations" could also be valuable. These would examine how an AI system behaves over time, highlighting where risks emerge, how they are currently handled, and where governance gaps exist. This allows participants to collectively explore what effective oversight and intervention would require. Cross-disciplinary design sprints can bring together policymakers, technical experts, and frontline practitioners to co-develop practical approaches to challenges such as risk detection, accountability, and interoperability. These sessions should be structured to produce tangible outputs, not just dialogue. In addition, evidence-sharing forums where practitioners present anonymised real-world cases or observed patterns would help ensure that governance discussions are informed by lived experience, not only theoretical models. Finally, the Dialogue should incorporate iterative working groups that continue beyond the event itself, enabling sustained collaboration, testing, and refinement of proposed frameworks across different jurisdictions. By prioritising interactive, evidence-driven, and outcome-oriented formats, the AI Dialogue can shift from discussion to practical progress in AI governance.

Several initiatives are advancing effective AI governance across policy, technical practice, and operational implementation. On the policy side, the EU AI Act establishes a risk-based framework that aligns regulatory obligations with potential harm, particularly for high-risk systems. The OECD AI Principles and UNESCO's Recommendation on the Ethics of AI have also created widely adopted baselines around transparency, accountability, and human oversight. At the implementation level, practices such as algorithmic impact assessments (AIAs), model evaluations, and adversarial testing (red-teaming) are increasingly being used to identify risks prior to deployment. Leading AI developers are beginning to publish safety reports and evaluation benchmarks, contributing to greater transparency. Standards bodies such as ISO/IEC are also playing an important role in formalizing governance processes, particularly around risk management, quality assurance, and lifecycle controls. However, most of these approaches remain focused on pre-deployment assessment or internally conducted evaluations. A critical gap is the lack of independent, ongoing verification of how AI systems behave once deployed in real-world, dynamic environments. This is especially important in contexts involving vulnerable populations, where risk evolves over time and cannot be fully predicted in advance. Strengthening governance will require moving beyond static compliance toward continuous validation - including the ability to detect emerging behavioural risks, assess system responses in real time, and provide independent assurance that safety claims hold in practice. Bridging this gap is key to ensuring AI governance is not only well-defined, but effective under real conditions.