APrIGF Document Platform

Recommendation: Integrate Child Safety and AI Literacy into AI Governance

AI governance should place greater emphasis on child safety and age-appropriate AI literacy. Governments and education stakeholders should consider integrating AI education into school curricula from an early stage, helping children understand how AI and large language models (LLMs) work, their potential benefits, their limitations, and the risks associated with their misuse.

This education should promote responsible AI use, critical thinking, privacy protection, recognition of AI-generated misinformation and deepfakes, and awareness of online harms. Alongside education, stronger child-centred safeguards, age-appropriate AI tools, and accountability for technology providers are needed to protect children in AI-enabled digital environments.

AI governance should not focus solely on regulating technology; it should also empower children with the knowledge and skills to use AI safely, critically, and responsibly.

A major conceptual point here is that Internet resilience is not necessarily Internet-only resilience. In emergencies, disasters or connectivity disruptions, trusted local communication channels can complement online systems. Community radio is therefore relevant to Internet governance because resilience concerns the broader information ecosystem through which people receive reliable and actionable information. Second, the session highlights the continuing importance of local and community-based media even as communication becomes increasingly digital. Community radio can serve populations whose Internet access may be intermittent, unaffordable or limited by geography, digital literacy or infrastructure. Third, trust is strongly connected to proximity and local knowledge. Community broadcasters often operate within the communities they serve, use local languages and understand local circumstances. In a crisis, that can make them valuable intermediaries between official information and the people who need to act on it. Fourth, the “digital era” component suggests that the relationship between radio and the Internet is no longer simply competitive. Community broadcasters can combine terrestrial broadcasting with streaming, social media, messaging platforms, podcasts or other digital channels. The issue therefore becomes how traditional community communication networks can be incorporated into a more resilient hybrid information infrastructure. A fifth important point concerns meaningful connectivity. A person may technically be “connected” but still lack reliable access to information because of cost, language, literacy, disability, network disruption or lack of trustworthy sources. The inclusion of community radio broadens the definition of connectivity from mere network access to the ability to receive information that is usable, relevant and trusted.

The main concern is the transition from technical debt – such as outdated systems, deferred security updates, unsupported software, and fragmented infrastructure  – to risks affecting institutional security, public trust, and wider Internet resilience. AI can compound these risks by increasing system complexity, expanding attack surfaces, and introducing new governance and accountability challenges.

Accummulated technical debt can develop into broader strategic vulnerabilities affecting cybersecurity, institutional continuity, and public trust. The increasing integration of AI into digital infrastructure introduces additional complexities, including new attack surfaces, risks associated with AI-enabled threats, and questions concerning accountability.

Strengthening cyber resilience requires moving beyond reactive incident management toward preventive, risk-based governance. Possible measures include regular infrastructure assessments, lifecycle security management, coordinated vulnerability disclosure, incident response preparedness, and sustained investment in cybersecurity expertise.

AI systems used in critical infrastructure and public services should be subject to appropriate security assessments, human oversight, and clearly defined institutional accountability. Regional cooperation through information sharing, capacity building, and multistakeholder engagement can help address disparities in cybersecurity readiness across Asia-Pacific economies. Cyber resilience should be treated as a continuing governance responsibility rather than solely a technical compliance exercise.

Internet connectivity and energy security are closely interdependent. Disruptions to electricity supplies can compromise telecommunications infrastructure, interrupt essential digital services, and disproportionately affect remote and underserved communities. Digital resilience strategies should therefore address the reliability of energy supplies alongside network availability, affordability, and accessibility.

Potential measures include decentralized renewable energy systems, appropriate backup power arrangements, energy-efficient network infrastructure, and coordinated contingency planning between telecommunications operators, energy providers, governments, and local communities. Public investment and financing mechanisms should also consider the long-term maintenance and sustainability of connectivity infrastructure, particularly in vulnerable regions.

Internet resilience should be assessed not only by network coverage under normal conditions, but also by the ability to maintain or restore meaningful connectivity during energy shortages, disasters, and other crises.

“Strongly agree with the focus on digital public infrastructure (DPI) and data protection in South Asia. To make this section more actionable, the final document should explicitly highlight the need for independent civil society oversight to ensure that DPI rollouts do not compromise fundamental privacy rights or enable mass profiling under national security pretexts

Thank you for the comprehensive draft. To strengthen the disaster resilience sections, it would be valuable to explicitly highlight regional peer-learning and cross-border solidarity among technical communities during infrastructure crises

Just as clean water is essential for life, demanding recognition as a fundamental right to prevent systemic exclusion, representative data is the digital ecosystem’s vital resource.

When AI foundation models omit Asia-Pacific’s rich linguistic diversity and indigenous knowledge, entire communities face digital starvation and erasure. Yet, once absorbed, this data becomes nearly impossible to “unlearn,” rendering the traditional Right to be Forgotten obsolete.

We urgently need robust governance, moving beyond individual data deletion toward collective consent, digital data sovereignty, and inclusive AI training that preserves cultural heritage without exploitation.

Keeping in view the permanent mandate of UN-IGF, there is dire need to overcome massive linguistic diversity in Asia Pacific on high priority.

This session’s call for a binding regime and meaningful human control came closest to governing AI itself rather than its uses. I suggest the synthesis note that the same principles should apply to general-purpose frontier models, since those models can be adapted for military and other high-risk uses after release. Binding rules for weapons alone will leave a gap if the underlying models are not governed.

The sessions under this track looked at how AI is used: in elections, in military targeting, in training data, and in content that spreads bias. I think the synthesis should also record a gap. We did not discuss the development of frontier AI models themselves, or whether that development should be slowed until governance catches up. I suggest adding a line here, or a point under Responsible AI Governance, on governing frontier model development.
We cannot reliably measure what these models can do. Static benchmarks stop working once they are public, because the questions end up in training data. Research published by AI labs has also shown that models can sometimes recognise when they are being evaluated and behave differently. When companies publish their safety tests to build trust, that information goes online and later models can learn from it. Honest evaluation gets harder every year.
AI has also lowered the skill needed to cause harm. Someone no longer needs expert knowledge in a dangerous field. Intent and a basic jailbreak can be enough.
The EU AI Act is the most complete framework we have, but it covers what is placed on the EU market. It does not stop a more capable model from being trained somewhere else, and most APAC countries have nothing comparable. I believe frontier development should be slowed, for something like 6 to 12 months, while governments agree on evaluation standards and on penalties for companies and developers who release unsafe systems. (Note: the duration is dependent upon the time to catch up with policies).
 
The common objection is the US-China race: if one side slows down, the other gets ahead. The Cold War is often used to argue that rivals cannot cooperate. The record shows the opposite. The US and the Soviet Union signed the Partial Test Ban Treaty in 1963 and the Non-Proliferation Treaty in 1968 because the risk was shared. Asia Pacific has a direct stake, since China is in this region and much of the world’s advanced chip production happens in East Asia. I would recommend that APrIGF 2027 hold a dedicated session on frontier AI safety and regional cooperation.