Ethica Group Ltd

A successful first Global Dialogue on AI Governance would move from defining principles to demonstrating control in practice. Current frameworks emphasise risk management, transparency, and human oversight. However, the presence of oversight does not reliably ensure the ability to intervene in system behaviour once deployed. This gap limits the effectiveness of existing governance approaches. The Dialogue should recognise intervention capacity as a core dimension of AI governance: the practical ability to pause, override, or redirect AI-enabled decisions in real time, supported by clear authority and executable escalation. To advance this, the Dialogue could: promote simple, comparable metrics (e.g. time to intervention, clarity of decision authority); encourage testing of override mechanisms under realistic conditions; highlight governance designs that enable timely intervention across contexts. Focusing on whether systems can be stopped or redirected when needed would strengthen the credibility of global AI governance and bridge the gap between stated oversight and actual control.

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

These priorities reflect the need to move from high-level principles to operationally effective AI governance. Transparency, accountability, and human oversight are central, but in practice often focus on monitoring rather than the ability to act. Ensuring that oversight translates into effective intervention is critical to making these mechanisms meaningful. Safe, secure and trustworthy AI depends not only on system design, but on whether organisations can maintain control once systems are deployed. This requires governance approaches that can operate under real-world conditions, including speed, scale, and complexity. Interoperability of governance approaches is important to enable consistency across jurisdictions and sectors. A shared understanding of core operational capabilities, such as the ability to intervene, can support alignment without requiring identical regulatory frameworks. Finally, the broader social, economic, and ethical implications of AI are directly shaped by how systems behave in practice. Where intervention is delayed or ineffective, impacts can scale quickly. Strengthening governance therefore requires attention to how oversight functions under real conditions, not only how it is specified.

A key cross-cutting issue not fully captured across the listed themes is the gap between the design of oversight and the ability to intervene in practice. Existing discussions on transparency, accountability, and human oversight largely focus on the presence of monitoring, reporting, and governance structures. Less attention is given to whether these mechanisms enable timely and effective intervention once AI systems are deployed. This raises a broader question of operational control: whether organisations can pause, override, or redirect system behaviour under real-world conditions, including speed, scale, and organisational complexity. Without this capability, governance risks remaining procedural rather than effective. This issue cuts across all thematic areas, including safety, human rights, and societal impacts, as the consequences of delayed or ineffective intervention can propagate rapidly. Addressing this gap would benefit from greater focus on decision authority, escalation pathways, and the practical conditions required to enable intervention. Strengthening these elements can help ensure that governance frameworks translate into meaningful control in practice.

Across regulated sectors, AI governance has advanced in structure but remains limited in operational effectiveness. Organisations increasingly meet formal expectations on transparency, documentation, and oversight. Roles are defined, controls are described, and assurance processes are in place. However, these structures do not reliably ensure control once systems are deployed. A critical challenge is the gap between oversight and intervention. In practice, organisations may detect issues but are not positioned to act with sufficient speed or authority. Decision rights can be fragmented, escalation pathways conditional, and intervention mechanisms untested under real operating conditions. As a result, systems can continue to produce harmful or unintended outcomes despite being formally compliant. This challenge is amplified by the scale and velocity of AI-enabled decision-making. Issues can propagate rapidly across thousands of outcomes before effective intervention occurs, with direct implications for customers, markets, and institutional trust. At the same time, this exposes a clear opportunity. Embedding measurable intervention capability-such as clarity of authority, time to intervention, and effectiveness of override-would strengthen governance from procedural assurance to demonstrable control in practice.

The AI Dialogue can play a critical role by shifting international cooperation from alignment on principles to alignment on operational capability. Global consensus already exists around core concepts such as safety, transparency, accountability, and human oversight. However, cooperation remains limited where it matters most: how these principles function in practice across different systems, sectors, and jurisdictions. The Dialogue can advance cooperation by establishing a shared focus on what effective governance looks like operationally. This includes developing common reference points for capabilities such as intervention, whether organisations can pause, override, or redirect AI-enabled decisions in real time, supported by clear authority and executable escalation. Rather than seeking uniform regulation, the Dialogue can enable interoperability through shared expectations of outcomes. Agreeing on a small number of comparable operational metrics - such as time to intervention or clarity of decision authority-would allow different governance approaches to be assessed on a consistent basis. By focusing on demonstrable control rather than formal alignment, the Dialogue can strengthen trust, improve comparability across jurisdictions, and support more effective cooperation in the governance of AI systems.

The AI Dialogue can build on a growing set of regulatory, standard-setting, and multi-stakeholder initiatives that have established important foundations for AI governance. These include regional regulatory frameworks, international principles on trustworthy AI, and emerging technical standards that address risk management, transparency, and oversight. While these efforts provide structure, they remain fragmented across jurisdictions and often focus on compliance design rather than how governance performs in practice. A key gap is the absence of shared, operational reference points for assessing whether organisations can effectively control AI systems once deployed. The added value of the AI Dialogue would be to connect these initiatives through a focus on practical governance capability. This includes creating a common language for outcomes such as intervention - whether systems can be paused, overridden, or redirected under real conditions - and how this capability can be demonstrated and compared across contexts. By complementing existing frameworks with a small number of shared operational expectations and metrics, the Dialogue can enable interoperability without requiring uniform regulation. This would strengthen coherence across initiatives, improve comparability, and support more effective international cooperation grounded in how governance works in practice.

Effective participation requires moving beyond representation towards contribution that is structured and outcome-oriented. Different stakeholders bring distinct value: governments provide regulatory direction, the private sector contributes implementation experience, academia and the technical community offer analytical and design expertise, and civil society highlights societal impact. The Dialogue should structure contributions around these roles rather than treat all inputs as equivalent. To enable this, the Dialogue could be organised around focused thematic tracks with defined outputs, where stakeholders are invited to contribute specific evidence, use cases, or frameworks. Contributions should be anchored in practical questions, such as how governance functions under real conditions, rather than general principles. A useful structure would combine: short, curated interventions to surface key perspectives; moderated working sessions to develop concrete outputs; synthesis mechanisms that translate inputs into reusable frameworks or recommendations. Clear framing of expected outputs, such as operational definitions, metrics, or governance models, would help ensure contributions are actionable and comparable. This approach would strengthen the Dialogue's ability to move from broad participation to meaningful, structured collaboration.

Global AI governance discussions often include diverse stakeholder categories, but underrepresent those with direct responsibility for operating and intervening in AI-enabled systems. These include senior operational decision-makers, risk owners, and control functions within organisations who are accountable for real-time outcomes. Their perspective is critical to understanding how governance frameworks function under pressure, including constraints related to authority, escalation, and intervention. In addition, perspectives from organisations operating in high-impact environments, such as financial services, healthcare, and public sector delivery, are not always sufficiently reflected in a way that captures operational realities rather than policy positions. There is also a gap in representing experiences from contexts where governance capacity is still developing, particularly in relation to the practical challenges of implementing oversight at scale. To address this, the Dialogue could: prioritise participation from individuals with direct accountability for system outcomes; invite structured case-based contributions focused on real-world governance challenges; create space for operational insights alongside policy and technical perspectives. Including these voices would strengthen the Dialogue's ability to ground governance discussions in how systems behave in practice.

To foster meaningful engagement, the Dialogue should incorporate formats that test how governance works in practice, not only how it is described. One approach would be scenario-based sessions, where participants are presented with realistic situations involving AI system behaviour and asked to respond in real time. This can surface gaps in decision authority, escalation, and intervention capability that are not visible in abstract discussion. Another format is cross-stakeholder working labs, focused on producing specific outputs such as operational definitions, metrics, or governance patterns. These sessions should be time-bound and outcome-driven, with clear expectations for deliverables. The Dialogue could also include comparative case reviews, where different organisations or jurisdictions present how similar governance challenges are addressed in practice. This would support learning across contexts and highlight differences in operational capability. Finally, incorporating feedback loops, where insights from these sessions are synthesised and tested in subsequent discussions, would help build continuity and depth. These formats would enable the Dialogue to move from exchange of views to development of practical, usable governance approaches.

A number of existing initiatives provide important foundations for effective AI governance, particularly where they move beyond principles to operational practice. Regulatory frameworks such as the EU AI Act establish structured approaches to risk classification, lifecycle governance, and human oversight. Similarly, international standards, including ISO/IEC 23894 (AI risk management) and the OECD AI Principles, have helped define expectations around transparency, accountability, and system reliability. In practice, more advanced approaches are emerging within regulated sectors. For example, in financial services, model risk management and validation frameworks are increasingly extended to AI systems, incorporating pre-deployment testing, ongoing monitoring, and defined escalation processes. These approaches demonstrate how AI governance can be embedded within existing control environments. However, a consistent gap remains between governance design and operational effectiveness. Emerging good practice is therefore shifting towards testing governance under realistic conditions, including scenario-based exercises to assess whether issues can be detected, escalated, and acted upon in a timely manner. Further progress is seen where organisations define clear decision authority for intervention and align technical controls with organisational accountability. Taken together, these approaches point to a necessary evolution: from governance as documentation and oversight, towards governance as demonstrable control in practice. Strengthening this shift across jurisdictions remains a key opportunity for the Dialogue.