Indonesia National AI Taskforce / Data Protection Center of Excellence

1. Consensus on technical requirement for AI. We no longer need principles or soft laws, governance MUST be measurable and one of the item is the technical requirement (eg: parity testing, opacity level, etc). 2. The use of dual-purpose AI models in case of Anthrophic must be regulated to ensure that AI is not developed to harm humanity. 3. Government-level and international-level procurement of AI system must be transparent.

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

Fragmented regulatory regimes create governance mismatches that undermine global AI safety, particularly as nations pursue AI sovereignty through increasingly closed models Model distillation attacks exemplify the urgency. Adversaries extract proprietary safeguards from open APIs (e.g., OpenAI, Anthropic), creating "shadow models" that retain capabilities but strip safety mechanisms. Anthropic documented how distillation bypasses constitutional AI protections, enabling unrestricted deployment of high-risk systems. Without interoperable standards for model provenance and audit trails, these attacks proliferate unchecked, posing national security risks as distilled models undercut safety-conscious providers AI sovereignty trends exacerbate this vulnerability. By 2026, 93% of executives view sovereignty as mission-critical, driving localization of compute and data. Governments increasingly restrict cross-border model access (e.g., Chinese models banned in Western critical infrastructure; US export controls creating a two-tier AI world). This "splinternet" fragments safety oversight: a model safe in one jurisdiction becomes dangerous when distilled and redeployed elsewhere without traceability Safe, secure, trustworthy AI requires interoperable baselines: mutual recognition of safety audits, standardized risk thresholds, and distillation-resistant provenance protocols. Open source software, open data, and open AI models enable collaborative benchmarking but must integrate with sovereign frameworks to prevent distillation abuse.

A critical gap exists: democratic legitimacy in frontier AI alignment. Current governance debates focus on risk mitigation and technical standards but overlook the foundational question: who decides what values AI systems encode? Professor Gilad Abiri's Public Constitutional AI framework offers a pathway. Constitutional AI, pioneered by Anthropic, hardcodes explicit principles into model training (e.g., "respect human rights," "reject misinformation"). However, Anthropic's constitution was drafted by employees, not publics. This creates an opacity deficit (technical principles inaccessible to democratic contestation) and a political community deficit (universal principles divorced from situated social contexts). The UN should lead a global Public Constitutional AI initiative. Rather than jurisdiction-by-jurisdiction fragmentation, the UN General Assembly could convene a Global AI Constitutional Assembly, representative deliberation producing a foundational constitution for transnational AI systems. This would: 1.Bridge sovereignty and interoperability: Nations adopt shared baseline principles while localizing context-specific rules 2.Democratize AI governance: Shift from expert-driven to participatory legitimacy, addressing the "AI legitimacy crisis" 3.Operationalize UN values: Translate Universal Declaration of Human Rights into auditable model behavior 4.Counter corporate monopoly: Prevent unilateral norm-setting by frontier labs

Challenge 1: Ministerial Coordination Breakdown Multiple ministries are competing to "own" AI governance, creating conflicting mandates. The Ministry of Communication and Digital Affairs leads the National AI Roadmap (delayed to 2026), but sectoral implementation falls to individual ministries as each developing parallel frameworks without coordination. The result: everyone wants to jump on the AI bandwagon, but no entity has clear authority. For example, the Ministry of Industry pushes AI-driven manufacturing, the Ministry of Education develops AI literacy programs, and the financial regulator (OJK) drafts AI-in-finance rules—all operating in silos. UNESCO's Indonesia assessment explicitly recommended creating a National Agency for AI to resolve this fragmentation, but political resistance from ministries unwilling to cede jurisdiction has stalled progress. This coordination failure prevents Indonesia from implementing interoperable governance approaches globally. We cannot adopt mutual recognition agreements when domestic agencies cannot agree on baseline standards internally. Challenge 2: Research Ecosystem Gatekeeping AI policy development is dominated by KORIKA (Collaborative Research and Industrial Innovation in AI), a small consortium of selected academics and industry leaders. While KORIKA drafted the 2020 National AI Strategy, its closed structure blocks broader participation from civil society, regional universities, and emerging startups. This creates an echo chamber where governance priorities reflect incumbent interests rather than diverse stakeholder needs. Opportunity: Public Constitutional AI These challenges make Indonesia an ideal pilot for UN-convened Public Constitutional AI. Our governance gaps stem from legitimacy deficits—exactly what participatory constitutional design resolves. A UN-backed process would: 1.Break ministerial deadlock: External legitimacy forces coordination 2.Open research ecosystems: Global standards require inclusive deliberation 3.Align domestic-global frameworks: Constitutional principles bridge sovereignty and interoperability

As mentioned in the previous answer, The AI Dialogue must bridge the legitimacy gap between technical standards and democratic governance by institutionalizing Public Constitutional AI processes. It should convene a Global AI Constitutional Assembly, translating the Universal Declaration of Human Rights into auditable model behaviors through representative deliberation. This prevents corporate monopoly over value alignment while enabling interoperability—nations adopt shared baseline principles while localizing context-specific rules. The UN Scientific Panel provides technical rigor; the Dialogue provides democratic legitimacy. Without participatory input mechanisms, AI governance defaults to whoever controls compute—the Dialogue must ensure Global South voices shape norms, not just comply with them.

Governments should organize national AI constitutional assemblies using citizen sortition, feeding democratic input into global norm-setting rather than expert position papers. The Dialogue should establish tiered deliberation: Global Constitutional Assembly (representative citizens), Regional Working Groups (localization), and Technical Implementation Committee (Scientific Panel translating principles into auditable metrics)

Global South implementation practitioners, displaced workers, regional universities outside elite consortia, indigenous communities, and youth (16-25) are critically absent. Include them through sortition-based selection for constitutional assemblies, preventing elite capture while ensuring statistical representativeness across geographies, sectors, and generations.

Industry Centres of Excellence (CoEs) offer a proven model for closing AI governance knowledge gaps. A CoE is a consortium where multiple companies pool expertise to develop shared standards, policies, and technical protocols by avoiding duplicated effort while establishing sector-wide governance baselines. CoEs operate through cross-functional working groups translating abstract principles into auditable processes: approval gates for model deployment, review boards for ethical oversight, and documentation protocols for transparency. Microsoft's AI CoE defines technology stack standards, preventing tool sprawl while enabling reusable governance templates across divisions. Governments must adopt this model urgently. Indonesia's ministerial fragmentation stems from treating AI governance as jurisdictional competition rather than collaborative problem-solving. A National AI Governance CoE-convening ministries, startups, civil society, and academia-would: 1.Develop shared standards (e.g., algorithmic transparency requirements) applicable across sectors, ending siloed frameworks 2.Pool scarce expertise through knowledge repositories, addressing capacity constraints 3.Run pilot projects testing governance approaches before codifying regulations, reducing implementation failures 4.Create decision-making frameworks clarifying which entity approves what, resolving ministerial deadloc