ANTROPOLOGIC

A successful outcome of the first Global Dialogue on AI Governance would not be defined only by the breadth of participation or the diversity of perspectives, but by the introduction of operationally meaningful criteria that can be applied across real systems. At present, most governance frameworks focus on compliance, transparency, and the formal presence of human oversight. These are necessary, but not sufficient. A key gap remains largely unaddressed: whether the human retains a real and effective capacity to exercise command within increasingly automated systems. In many high-risk environments, the human remains formally responsible, yet the architecture of the system progressively constrains the ability to perceive, interpret, and intervene in time. This creates a structural mismatch between responsibility and actual control. A successful Dialogue would therefore achieve three concrete outcomes: First, the recognition that "human oversight" is not a binary condition, but a variable that depends on operational factors such as timing, cognitive load, system transparency, and real intervention capacity. Second, the development of shared criteria to assess the operational viability of human command within AI-enabled systems, moving beyond formal compliance toward verifiable conditions of control. Third, the integration of these criteria into governance discussions, standards, and future regulatory approaches, ensuring that responsibility is aligned with actual human capability. Without this shift, there is a risk that governance frameworks may preserve the appearance of control while overlooking its effective erosion.

• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights

The selected priorities reflect a common underlying concern: the growing gap between formal responsibility and actual human control in AI-enabled systems. "Transparency, accountability, and human oversight" is central, but current approaches often assume that the presence of a human in the loop ensures meaningful control. In practice, this is not always the case. In many operational environments, increasing system speed, complexity, and abstraction constrain the human capacity to perceive, interpret, and intervene in time. This directly connects with "Safe, secure and trustworthy AI." Safety cannot be reduced to technical reliability alone; it also depends on whether human operators can effectively exercise judgment under real conditions. A system may function correctly while still degrading the conditions required for human command. The inclusion of "Protection and promotion of human rights" reflects the implications of this gap. When responsibility is assigned to humans who cannot exercise real control, issues of accountability, liability, and fairness emerge in a structurally unresolved way. Finally, the broader category of "Social, economic, ethical, cultural, linguistic and technical implications of AI" is relevant because these dynamics are not purely technical. They involve organizational structures, decision-making practices, and the evolving role of human agency in mediated environments. Across these areas, there is a shared need to move beyond formal definitions of oversight toward criteria that can assess the operational viability of human command in AI systems. Without this shift, governance risks preserving the appearance of control while overlooking its effective erosion.

Yes. A key cross-cutting issue not fully captured by the listed themes is the structural gap between formal human responsibility and the actual operational capacity to exercise control in AI-enabled systems. Current governance approaches tend to address safety, transparency, and oversight as distinct dimensions. However, in practice, these dimensions converge around a deeper question: whether the human can still meaningfully perceive, understand, and intervene within the system under real conditions. As systems increase in speed, complexity, and autonomy, a structural condition can emerge in which human responsibility is preserved, while the capacity to exercise judgment and timely intervention is progressively constrained. This is not necessarily visible at the level of formal design or compliance, and therefore remains largely unmeasured. This issue cuts across safety, accountability, and human rights. A system may be technically reliable, compliant with governance frameworks, and formally supervised, yet still create conditions in which human control is only nominal. In such cases, responsibility is maintained without a corresponding capacity to act. An important area for further attention is therefore the development of criteria and methodologies to assess the operational viability of human command within AI systems. This includes factors such as timing, cognitive load, system transparency, and the real possibility of intervention. Without addressing this cross-cutting issue, there is a risk that governance frameworks will continue to focus on the presence of oversight, rather than on its effective functionality.

In the European context, and particularly in sectors operating under the AI Act and related regulatory frameworks, governance gaps are increasingly visible at the operational level. While significant progress has been made in defining requirements around risk management, transparency, and human oversight, a key challenge remains: the translation of these principles into real conditions of human control within complex, automated systems. In high-impact sectors such as finance, critical infrastructure, and security, AI systems are often integrated into decision processes characterized by high speed, large volumes of data, and increasing system opacity. Under these conditions, human operators frequently retain formal responsibility, but face growing constraints in their ability to fully understand system outputs, exercise independent judgment, or intervene effectively in time. This creates a structural tension between regulatory expectations and operational reality. Compliance can be demonstrated, while the effective capacity for human control may be significantly reduced. At the same time, this gap represents an opportunity. As regulatory frameworks mature, there is increasing demand for approaches that move beyond formal oversight toward the verification of its operational viability. This includes the need for methodologies that assess factors such as cognitive load, timing constraints, system interpretability, and real intervention capacity. Addressing this gap could strengthen the credibility and effectiveness of governance frameworks, ensuring that accountability is aligned with actual human capability. It also opens a space for new forms of independent assessment focused on the relationship between human agency and system behavior in real-world conditions.

The AI Dialogue can play a critical role by shifting international cooperation from alignment on principles to convergence on operational criteria. Current efforts have made important progress in establishing shared values, such as safety, transparency, and human oversight. However, these concepts are often interpreted differently across jurisdictions, and their implementation varies significantly in practice. This creates a gap between formal agreement and operational reality. The Dialogue can help address this by fostering a common understanding of what these principles mean under real system conditions. In particular, it can support the development of shared criteria to assess whether governance mechanisms—especially human oversight—remain effective as systems increase in complexity, speed, and autonomy. A key contribution would be to move beyond the assumption that the presence of oversight ensures control, and instead promote methods to evaluate its actual functionality. This includes examining factors such as timing constraints, cognitive demands on human operators, system interpretability, and the real capacity for intervention. By facilitating this shift, the Dialogue can help reduce fragmentation between regulatory approaches and improve the comparability of governance practices across countries and sectors. Ultimately, international cooperation will be strengthened not only by agreeing on what should be achieved, but by developing a shared capacity to verify whether it is truly happening in practice.

The AI Dialogue should build upon existing regulatory, standardization, and multi-stakeholder initiatives that have already advanced key aspects of AI governance. These include regulatory frameworks such as the EU AI Act, international principles developed by the OECD and UNESCO, and technical standardization efforts led by organizations such as ISO/IEC and IEEE. In addition, multi-stakeholder platforms and research initiatives have contributed to the development of best practices around risk management, transparency, and accountability. These initiatives provide an essential foundation: they define shared principles, establish risk-based approaches, and promote interoperability across jurisdictions. However, they also tend to focus on formal requirements and high-level guidelines, while leaving a gap at the level of operational verification. The added value of the AI Dialogue would be to connect these efforts by addressing how governance principles translate into real-world system conditions. In particular, it can serve as a platform to align not only on what should be implemented, but on how to assess whether it is effectively working in practice. This includes advancing shared approaches to evaluate the functionality of governance mechanisms—especially human oversight—under conditions of increasing system complexity, speed, and opacity. By fostering convergence around operational criteria, the Dialogue can strengthen coherence between regulatory frameworks, technical standards, and organizational practices. In doing so, it can help ensure that existing initiatives are not only aligned in principle, but also mutually reinforcing in their practical impact.

Different stakeholders can contribute most effectively to the AI Dialogue by engaging at complementary levels: principles, implementation, and operational validation. Governments and international organizations can provide regulatory direction, define common frameworks, and ensure alignment with public interest objectives. The private sector and technical community can contribute practical insights on system design, deployment, and performance under real conditions. Academia and civil society can offer critical analysis, interdisciplinary perspectives, and independent scrutiny. To maximize impact, the Dialogue should be structured to move beyond general discussion toward progressively more operational layers. First, plenary sessions can focus on high-level alignment around principles and shared challenges. Second, thematic working groups should address specific governance areas (e.g., safety, oversight, accountability), incorporating cross-sector expertise. Third, and critically, the Dialogue should include structured case-based analysis, where real or simulated high-risk systems are examined to understand how governance mechanisms function in practice. This case-based layer is essential. It allows stakeholders to move from abstract agreement to the evaluation of concrete conditions, including timing constraints, human cognitive load, system transparency, and the feasibility of intervention. In addition, the Dialogue could benefit from independent expert inputs focused on assessing the operational effectiveness of governance mechanisms, ensuring that formal requirements correspond to real capabilities. By combining multi-level participation with structured, practice-oriented formats, the AI Dialogue can strengthen both conceptual alignment and practical coherence across stakeholders.

One of the most underrepresented perspectives in global AI governance discussions is the operational viewpoint of those directly interacting with AI systems in real-world, high-stakes environments. Current debates are largely shaped by policymakers, technical experts, and high-level ethical frameworks. While these perspectives are essential, they often overlook how AI systems are actually experienced and managed by human operators under real conditions—particularly in sectors such as finance, healthcare, security, and critical infrastructure. These actors are not merely "users." They are formally responsible for decisions mediated by AI systems, yet their practical ability to understand, challenge, or intervene in system behavior is often constrained by factors such as time pressure, system complexity, interface design, and organizational incentives. As a result, an important dimension of governance remains underrepresented: the gap between formal responsibility and effective human control in operational contexts. In addition, there is limited representation of independent, cross-disciplinary perspectives that examine AI systems not only from a technical or regulatory standpoint, but from the standpoint of human agency within system architectures. To address this, the AI Dialogue could incorporate structured mechanisms to include operational actors, such as practitioners from high-risk sectors, through case-based sessions, scenario analysis, and targeted consultations. It could also encourage contributions from independent researchers working at the intersection of human factors, organizational dynamics, and system design. Including these perspectives would help ensure that governance frameworks reflect not only how AI systems are intended to function, but how they actually operate in practice.

Innovative engagement formats should move beyond panel-based discussions toward structured, practice-oriented interaction that reflects how AI governance challenges emerge in real systems. One effective approach would be the inclusion of case-based labs, where stakeholders collaboratively examine real or realistically simulated high-risk AI use cases. These sessions would allow participants to analyze how governance mechanisms—such as oversight, accountability, and control—function under operational conditions, including time constraints, system complexity, and limited interpretability. Another valuable format would be cross-functional simulation exercises, bringing together policymakers, engineers, operators, and legal experts to work through decision scenarios in real time. This can reveal gaps between formal governance assumptions and practical system behavior, fostering a shared understanding across disciplines. Structured "challenge sessions" could also be introduced, where specific governance assumptions (for example, the effectiveness of human oversight) are tested against concrete scenarios. This would encourage critical examination and reduce reliance on abstract or unverified assumptions. In addition, smaller expert roundtables focused on operational verification could help develop more precise criteria for assessing whether governance mechanisms are functioning as intended in practice. Finally, incorporating feedback loops—where insights from these formats are synthesized into actionable recommendations—would ensure that discussions translate into practical outcomes. By prioritizing formats that expose real-world dynamics, the AI Dialogue can foster more meaningful engagement and support the development of governance approaches that are not only principled, but operationally robust.

Several existing approaches contribute meaningfully to effective AI governance, particularly when they move from high-level principles to enforceable or testable conditions. Regulatory frameworks such as the EU AI Act provide an important foundation by establishing risk-based classifications, documentation requirements, and obligations around human oversight. Similarly, standards developed by organizations such as ISO/IEC and IEEE help translate governance principles into technical and procedural practices, supporting consistency and interoperability. On the implementation side, emerging practices in "governance-as-code" and real-time policy enforcement are increasingly relevant. By embedding constraints directly into system architectures-through rule engines, access controls, and automated checks-these approaches reduce reliance on purely procedural oversight and help ensure that certain conditions are enforced before actions are executed. In addition, audit and assurance mechanisms, including third-party assessments and internal validation processes, play a key role in evaluating compliance and system performance. These practices contribute to transparency and accountability, particularly in high-risk sectors. However, a common limitation across many of these approaches is that they often verify the presence of governance mechanisms, rather than their effectiveness under real operational conditions. An important complementary direction is therefore the development of approaches that assess the functionality of these mechanisms in practice. This includes evaluating whether human oversight is not only formally defined, but also operationally viable-taking into account factors such as timing, cognitive load, system interpretability, and the feasibility of intervention. By combining enforceable controls with methods to verify their real-world effectiveness, AI governance can move toward more robust and reliable outcomes.