AI Duty

A successful Dialogue must move beyond high-level "principles" to actionable interoperability. Success would be defined by three key outcomes: 1- A Unified Governance Taxonomy: Establishing a shared technical language for risk classification so that a "high-risk" system in one jurisdiction is recognized and audited similarly in another. 2- The "Global AI Commons": A concrete commitment to pooled resources—compute, datasets, and pre-trained open-source models—ensuring the Global South is not merely a consumer of AI but a co-creator. 3- The ADS-Interoperability Roadmap: A formal recognition that technical architecture (like the ADS-2026 Framework which I am currently working on) must be the "enforcement layer" for policy. The Dialogue should result in a roadmap for integrating automated compliance and "Safety-by-Design" into the global supply chain.

• Safe, secure and trustworthy AI
• Open-source software, open data and open AI models
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

These four priorities represent the "Technical Bedrock" of the ADS-2026 Framework. Safety & Trust are the end goals; without them, public pushback will stifle innovation. Interoperability is the most urgent pragmatic need. As a Lead Architect, I see how fragmented regulations create "compliance silos" that hinder small-to-medium researchers. Transparency & Accountability must be baked into the architecture, not added as an afterthought. Open-source/Open data is the only way to ensure that the "AI Divide" doesn't become a permanent digital caste system. By prioritizing these, we ensure that governance is both technically feasible and globally equitable.

Yes: Compute Governance and Environmental Sustainability. The current themes focus on the "software" and "ethics," but AI is a physical reality. The concentration of specialized hardware (GPUs/TPUs) creates a geopolitical bottleneck that ethics alone cannot solve. Furthermore, the carbon and water footprint of large-scale model training is an emerging "AI Divide" issue-developing nations may suffer the environmental costs of AI training without reaping the economic benefits. We need a framework for Sustainable Compute Equity that governs the physical infrastructure of AI.

In the research and architectural sector, the primary gap is Regulatory Arbitrage. Because there is no global baseline, we see "Safety Laundering," where models are trained in jurisdictions with zero oversight and then deployed globally. This puts ethical organizations like AI Duty at a competitive disadvantage. Challenges: The "black box" nature of proprietary models makes it impossible to verify claims of "safety" or "human oversight" without a standardized auditing protocol. Opportunities: There is a massive opportunity to lead the world in Architectural Accountability. By filling these gaps, we can create a global marketplace for "Verified AI," where the ADS-2026 Framework serves as a gold standard for trust.

The Dialogue should act as the "Registry of Record." While the G7 or OECD provide guidance for wealthy nations, the UN Dialogue is the only forum with the legitimacy to host a Global AI Risk Registry and a Best Practices Clearinghouse. It should serve as the "bridge" that translates the technical standards developed in the Global North into capacity-building programs for the Global South, ensuring that "International Cooperation" isn't just a euphemism for "Technology Transfer" but is instead true collaborative R&D.

It should build upon the OECD AI Principles (for policy definitions), the ISO/IEC 42001 (for management systems), and the G7 Hiroshima AI Process. Added Value: The Dialogue's unique value is Universal Inclusivity. Existing initiatives are often "club-based" (Western-centric). The AI Dialogue can add value by creating a "Global Sandbox"—a multi-stakeholder environment where researchers from emerging economies can test models against the ADS-2026 Framework using UN-facilitated compute resources.

The Dialogue should adopt a "Hub and Spoke" structure. Technical Community: Provide "Red-teaming" reports and architectural templates. Civil Society: Act as "Ethics Auditors" to ensure the ADS Framework protects marginalized communities. Format Recommendation: I propose "Policy Hackathons." Instead of just speeches, stakeholders should spend sessions "stress-testing" a hypothetical AI deployment scenario against proposed regulations to see where the architecture breaks.

Indigenous Knowledge Keepers and SME Developers from the Global South. Most AI is trained on Western, "WEIRD" (Western, Educated, Industrialized, Rich, and Democratic) data. This leads to linguistic and cultural erasure. To include them, the Dialogue must provide Digital Participation Grants and host regional satellite events in Nairobi, Jakarta, and La Paz, rather than requiring all input to happen in Geneva or New York.

Digital Twin Policy Simulations. Use AI to simulate the impact of a proposed governance rule on a specific economy or sector. This allows stakeholders to see the "downstream" effects of their decisions in real-time. Additionally, a "Reverse Pitch" format, where governments describe a problem they face (e.g., "how to audit local language LLMs") and the technical community pitches architectural solutions like ADS-2026.

The ADS-2026 Framework (Architecture-Driven Safety) is our primary contribution. It moves away from "Post-hoc" auditing and instead implements "Continuous Verification." Modular Guardrails: Using "wrapper" architectures that monitor AI inputs and outputs in real-time against a set of ethical constraints. - Immutable Audit Logs: Utilizing distributed ledger technology to create a permanent, tamper-proof record of how an AI reached a decision, facilitating the "Transparency and Accountability" mentioned in GA Resolution 79/325. - The "Duty of Care" Protocol: A technical standard that requires AI systems to "self-report" when they encounter data or prompts that fall outside their "Safe Operating Envelope." These practices ensure that governance is not just a document on a shelf, but a core component of the software's binary code.