Independent researcher

A primary barrier in current AI governance lies in the attempt to universalize "correct" value judgments across states with divergent cultural and legal frameworks. This proposal draws on the principles of pharmacovigilance in the pharmaceutical domain and advances a lightweight yet effective framework based on Execution-Boundary Design at the OS layer. The key success factor of this approach lies in its ability to mechanize only the enforceable components of governance while leaving value judgments to individual states and institutions. Specifically, provenance and rights information are embedded as non-semantic "tags" at the OS level, and the AI model is required only to verify their consistency. This architecture eliminates the need to impose computationally intensive and inherently subjective ethical reasoning on the model itself, while enabling immediate alignment with jurisdiction-specific legal systems and cultural boundaries (Hamecohming–Institution). In addition, the introduction of a specialized post hoc auditing function—AI-PV (AI Safety Vigilance)—ensures transparency through pre-incident logging. This "non-judgmental architecture" offers a viable pathway toward a global standard that simultaneously respects national sovereignty, mitigates systemic risks, and reduces inference costs.

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

A key limitation of current AI governance lies in the heavy reliance on semantic censorship, which requires AI models to make subjective ethical judgments. This approach imposes significant computational burdens and fails to accommodate global diversity. To address this, this proposal introduces Execution-Boundary Design at the OS layer, inspired by pharmacovigilance in the pharmaceutical sector. Among the available options, four domains are particularly central: (1) interoperability of governance approaches, (2) transparency, accountability, and human oversight, (3) safe, secure, and trustworthy AI, and (4) protection and promotion of human rights. The core principle of this framework is the ability to mechanize only the enforceable components of governance while leaving value judgments to individual states and institutions. At the OS level, provenance and rights information are embedded as non-semantic "tags," which the AI model verifies mechanically. This enables unified execution control while preserving jurisdictional differences, thereby providing a foundation for interoperability. In addition, the introduction of a specialized post hoc auditing function-AI-PV (AI Safety Vigilance)-ensures verifiability through pre-incident logging (Phase-0 Logs). This structure institutionalizes transparency, accountability, and human oversight, while reinforcing safety and human rights protection through auditability. This "non-judgmental architecture" offers a viable foundation for an interoperable global standard by enabling effective control without imposing uniform value systems.

An underexplored cross-cutting issue in existing AI governance discussions is the sustainability of governance itself. This includes (1) the reduction of computational and energy burdens, and (2) the development of specialized human expertise required to sustain governance systems. First, current semantic censorship approaches impose significant computational overhead by requiring ethical evaluation at each inference step. This leads to increasing energy consumption across AI systems. The proposed Execution-Boundary Design addresses this by replacing semantic judgment with non-semantic tag verification at the OS level, thereby reducing inference costs. This shift aligns governance with the requirements of green IT, enabling both environmental efficiency and operational effectiveness. Second, governance cannot be sustained by technical systems alone. It requires dedicated human roles responsible for continuous monitoring and feedback. This proposal introduces AI-PV (AI Safety Vigilance) as a specialized function inspired by pharmacovigilance. Positioned between technical systems and institutional frameworks, AI-PV professionals detect early signals of incidents and ensure adaptive responses. Institutionalizing such human-centered oversight is essential to prevent governance frameworks from becoming formalistic or ineffective. By integrating both efficiency and human capacity into governance design, this approach addresses a critical gap in current frameworks and contributes to the long-term reliability and scalability of AI governance systems.

In this thematic area, a key governance gap manifests in Asia as a sharp increase in electricity demand driven by AI systems. Current AI governance frameworks emphasize principles such as ethics, safety, and human rights, but insufficiently address the management of computational resources and energy consumption. As a result, prevailing approaches—particularly those relying on semantic control within models—incur high energy costs at each inference step, placing growing pressure on power infrastructure. This impact is reflected in rising electricity costs and potential supply constraints. In regions where energy supply is already tight, the expansion of AI itself becomes a structural cost driver. Moreover, energy-inefficient governance methods pose long-term constraints from a sustainability perspective. The core gap lies in the fact that governance discussions focus primarily on what should be regulated, while overlooking the energy cost required to implement such regulation. Therefore, future AI governance must incorporate energy efficiency and implementation cost optimization alongside ethical and safety considerations. Without this shift, the simultaneous expansion of AI deployment and electricity demand will undermine the sustainability of AI systems at scale.

In this thematic area, a key governance gap manifests in Asia as a sharp increase in electricity demand driven by AI systems. Current AI governance frameworks emphasize principles such as ethics, safety, and human rights, but insufficiently address the management of computational resources and energy consumption. As a result, prevailing approaches—particularly those relying on semantic control within models—incur high energy costs at each inference step, placing growing pressure on power infrastructure. This impact is reflected in rising electricity costs and potential supply constraints. In regions where energy supply is already tight, the expansion of AI itself becomes a structural cost driver. Moreover, energy-inefficient governance methods pose long-term constraints from a sustainability perspective. The core gap lies in the fact that governance discussions focus primarily on what should be regulated, while overlooking the energy cost required to implement such regulation. Therefore, future AI governance must incorporate energy efficiency and implementation cost optimization alongside ethical and safety considerations. Without this shift, the simultaneous expansion of AI deployment and electricity demand will undermine the sustainability of AI systems at scale.

AI dialogue should build upon and interoperate with existing frameworks such as the International Council for Harmonisation (ICH) in the pharmaceutical sector, ISO/IEC standardization processes, and multilateral initiatives like the UN Global Digital Compact. Among these, the international cooperation mechanisms developed in pharmacovigilance (PV) represent one of the most mature and operationally effective models for cross-border risk management. The distinctive added value of AI dialogue lies in its ability to function as a bridge between normative frameworks and system-level technical enforcement. First, it can translate existing high-level ethical principles into operational specifications, such as common formats for non-semantic "tags" at the OS layer. This enables governance requirements to be implemented consistently across systems, thereby enhancing interoperability without imposing uniform value systems. Second, by adapting pharmacovigilance practices, AI dialogue can support the development of AI-PV (AI Safety Vigilance) as a professional and institutional function. Establishing shared standards for monitoring, logging, and incident response would help close the practical gap between regulation and implementation. Through these functions, AI governance can evolve from fragmented, model-specific moderation practices into a lightweight and robust infrastructure that operates at the system level. By allowing value judgments to remain locally defined while standardizing execution protocols, this approach reduces the cost of international coordination and facilitates scalable, cooperative governance.

Effective AI dialogue requires a structured allocation of stakeholder roles across different layers of governance infrastructure. This proposal adopts a three-layer model to define both contributions and dialogue formats. First, governments and international organizations (Hame layer) should focus on defining legal and ethical boundaries. In practice, this involves leading discussions on boundary conditions, including culturally and jurisdictionally specific requirements translated into tag-based specifications. Second, developers and OS-level vendors (Umecohming layer) are responsible for implementing these boundaries through Execution-Boundary Design. Their contribution to AI dialogue lies in establishing technical consensus on interoperable protocols, such as non-semantic tag verification mechanisms at the system level. Third, specialized professionals and independent oversight bodies (AI-PV layer) provide continuous monitoring and feedback based on audit logs. Drawing on models such as international pharmacovigilance (e.g., ICH), this layer contributes through practice-oriented dialogue formats, including shared standards for incident detection, logging, and response. Rather than aiming for uniform agreement across all domains, AI dialogue should be structured as a layered governance design process, where each group contributes within its domain of expertise. This approach enables distributed responsibility while maintaining systemic coherence. By organizing dialogue in this way, diverse stakeholders can make concrete, operational contributions, transforming AI governance from abstract discussion into implementable infrastructure.

One of the most underrepresented perspectives in current global AI governance discussions is that of individual creators and small-scale expressive communities, who provide essential training resources yet remain excluded from decision-making processes and value distribution mechanisms. A practical way to include these actors is to shift governance from output-based moderation to input-based consent and value redistribution. First, this proposal introduces OS-level tagging (Umecohming), which embeds usage conditions—such as "non-training consent" and licensing terms—directly into individual works as non-semantic, machine-readable tags. This enables even unaffiliated individuals to exert technical control over how their content is used by large-scale AI systems, effectively granting them a form of digital veto power. Second, these tags can be directly linked to an economic layer, enabling automated micropayments based on actual usage. This mechanism transforms creators from passive subjects of protection into active participants in the AI ecosystem, with enforceable rights and economic agency. Importantly, inclusive governance does not require that all stakeholders participate equally in political deliberation. Rather, it requires that fundamental rights—specifically, the ability to refuse and to receive fair compensation—are guaranteed at the infrastructure level. By embedding these rights into the OS layer, AI governance can incorporate marginalized voices in a scalable and enforceable manner, ensuring that those who contribute to AI systems are no longer structurally excluded from their benefits.

The most effective innovative engagement format is an evidence-based, dynamic feedback loop inspired by pharmacovigilance (PV) in the pharmaceutical industry. Conventional engagement formats often rely on static guidelines or one-time dialogues. In contrast, this proposal introduces a continuous, practice-oriented model centered on a specialized function—AI-PV (AI Safety Vigilance). AI-PV professionals monitor system-level logs (Phase-0 Logs) and compliance with execution boundaries at the OS layer on an ongoing basis. Based on this operational data, stakeholders—including developers, regulators, and users—engage in a real-time, iterative dialogue to evaluate which boundary conditions (e.g., tag specifications) function effectively and where adjustments are needed. This creates a structured mechanism for continuous calibration rather than episodic consultation. The key added value of this format is the shift from ex ante restriction to ex post continuous improvement. By grounding dialogue in verifiable evidence (logs), governance becomes adaptive and responsive to technological change, addressing the persistent gap between the speed of AI development and the pace of regulatory adjustment. Furthermore, this process enables the accumulation of a "safety history", where trust is built not through abstract assurances but through documented performance over time. By integrating monitoring, dialogue, and system updates into a single loop, this dynamic engagement model provides a scalable and resilient governance mechanism capable of balancing safety and innovation in AI systems.

The most effective innovative engagement format for AI dialogue is an evidence-based, dynamic feedback loop inspired by pharmacovigilance (PV) in the pharmaceutical sector. Conventional engagement models tend to rely on static guideline-setting or one-time consultations. In contrast, this approach introduces a continuous, operational model centered on a specialized function-AI-PV (AI Safety Vigilance). These professionals continuously monitor system-level logs (Phase-0 Logs) and compliance with execution boundaries at the OS layer. Using this empirical monitoring data, stakeholders-including developers, regulators, and users-engage in a real-time, iterative dialogue to evaluate which boundary conditions (e.g., tag specifications) are functioning effectively and where adjustments are required. This creates an ongoing calibration process grounded in observed system behavior rather than abstract principles. The key added value of this format lies in transforming governance from ex ante prohibition to ex post continuous improvement. By anchoring dialogue in verifiable evidence (logs), it enables governance mechanisms to adapt to the rapid pace of technological change, addressing the persistent mismatch between innovation and regulation. Moreover, this process allows for the accumulation of a "safety history," where trust is built through documented system performance over time rather than static compliance claims. By integrating monitoring, dialogue, and iterative updates into a unified loop, this dynamic engagement model provides a scalable and resilient governance mechanism capable of balancing safety and innovation in AI systems.