Shenzhen Institute of Artificial Intelligence and Robotics for Society

Success would mean the first Global Dialogue moves beyond broad principles and begins to identify the key AI risks that require global coordination. These should include not only frontier-model risks, but also physical harm, mass surveillance, cybersecurity vulnerabilities, misinformation, labor disruption, inequality, and the emerging risks of embodied AI systems acting in the real world. It should also produce a shared agenda for action: common safety-testing methods, incident reporting, transparency mechanisms, liability principles, and capacity-building support for countries that currently lack AI governance infrastructure. Most importantly, it should establish an inclusive, recurring process where major powers, developing countries, industry, academia, standards bodies, and civil society can jointly shape AI governance. A successful outcome would be a practical roadmap that reduces fragmentation, builds trust, and ensures AI development serves shared human benefit rather than deepening global inequality.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Safe, secure and trustworthy AI

We selected these priorities because AI governance must address both virtual AI systems and physical AI systems. For virtual AI, urgent risks include misinformation, cyber abuse, privacy violations, algorithmic bias, labor disruption, unequal access to AI capabilities, and the concentration of power among a small number of technology actors. These risks directly affect social trust, economic opportunity, cultural diversity, linguistic inclusion, and human rights. For physical AI, especially embodied AI systems such as robots, autonomous vehicles, drones, and intelligent machines, the risks extend beyond information space into the real world. These systems can cause physical harm, safety failures, surveillance expansion, liability gaps, labor displacement, and new forms of human-machine dependency. Because embodied AI acts in open environments, governance must consider not only model behavior, but also sensing, decision-making, control, deployment context, and human oversight. Therefore, safe, secure, and trustworthy AI should be the foundation of the Dialogue. At the same time, the social, economic, ethical, cultural, linguistic, and technical implications of AI must be treated as urgent priorities, because AI governance is not only about preventing harm, but also about ensuring that AI development benefits all societies, including developing countries and underrepresented communities.

Yes. One cross-cutting issue is the convergence of virtual AI and physical AI. Most current governance discussions focus on digital risks such as misinformation, bias, privacy, cyber misuse, and market concentration. These remain urgent. However, AI is increasingly moving from screens into the physical world through robots, autonomous vehicles, drones, smart homes, industrial systems, medical devices, and embodied agents. This creates a new risk layer: AI systems can now sense, decide, and act in open environments, potentially causing physical harm, safety failures, surveillance expansion, labor displacement, and unclear liability. This convergence also creates governance gaps. A model may appear safe in a digital benchmark but behave unpredictably when connected to sensors, actuators, users, and real-world infrastructure. Therefore, AI governance should not only evaluate models, but also deployment contexts, human oversight, system integration, data flows, cybersecurity, and accountability across the full AI stack. Another emerging issue is unequal access to AI infrastructure. If compute, data, talent, and deployment capacity remain concentrated in a few countries and companies, AI may deepen global inequality. The Dialogue should therefore address both risk reduction and inclusive capacity-building, so that AI becomes a shared development infrastructure rather than a source of dependency.

China is already one of the most advanced countries in physical AI, including robotics, autonomous vehicles, drones, smart manufacturing, logistics automation, and embodied AI systems. It is also likely to be among the first countries to see large-scale integration of physical AI into daily life, public services, industrial production, and urban infrastructure. This creates both major opportunities and governance challenges. The opportunity is that China can use physical AI to improve productivity, elderly care, healthcare delivery, manufacturing efficiency, public safety, transportation, and services in underserved regions. Physical AI can become a new layer of social and economic infrastructure. The challenge is that existing AI governance is still largely focused on virtual AI, such as large language models, online content, privacy, and algorithmic bias. Physical AI introduces additional risks because systems can sense, move, manipulate objects, interact with people, and make decisions in open environments. Failures may lead not only to information harm, but also to physical injury, infrastructure disruption, surveillance concerns, cybersecurity vulnerabilities, labor displacement, and unclear liability. Therefore, China's key governance gap is the need to move from model-level governance to system-level governance. This includes safety testing, real-world deployment standards, human oversight, incident reporting, cybersecurity, data protection, liability rules, and international technical standards for embodied AI. If addressed well, China can help define a practical governance model for the global deployment of physical AI.

The AI Dialogue can advance international cooperation by creating a concrete workstream on embodied AI and its societal impacts. My participation would focus on how AI systems that sense, decide, and act in the physical world, robots, autonomous vehicles, drones, smart devices, industrial systems, and future humanoid robots, should be governed before they are deployed at scale. The Dialogue can help define shared risk categories for embodied AI, including physical safety, cybersecurity, surveillance, labor displacement, liability, human oversight, and public trust. It can also promote practical cooperation on testing methods, deployment standards, incident reporting, data governance, and accountability across the full AI system stack, not only the AI model. Given my existing roles across IEEE, ACM, ISO, and the World Economic Forum, I see the Dialogue as a bridge between policy discussions and technical standardization. It can translate global governance concerns into actionable standards, benchmarks, best practices, and capacity-building programs, especially for countries preparing for large-scale physical AI deployment.

The AI Dialogue should connect directly with existing technical and policy mechanisms, including ISO work on robotics and AI standards, IEEE initiatives on autonomous and intelligent systems, ACM technology policy discussions, and World Economic Forum work on emerging technology governance. These communities already contain technical expertise, standards processes, and policy experience that can make the Dialogue more actionable. The added value of the AI Dialogue would be to integrate these efforts around emerging real-world AI deployment challenges. For embodied AI, governance cannot be handled by model regulation alone. It requires coordination among AI safety, robotics safety, cybersecurity, data protection, labor policy, product liability, and international standards. The Dialogue could therefore support a dedicated embodied AI governance track, bringing together governments, standards bodies, industry, academia, and civil society to develop shared terminology, risk taxonomies, testing protocols, and deployment guidelines. This would help ensure that physical AI is adopted safely, inclusively, and responsibly across different societies.

The AI Dialogue should be structured around concrete workstreams rather than only plenary discussions. One important workstream should focus on embodied AI and physical AI systems, covering robots, autonomous vehicles, drones, smart infrastructure, and AI-enabled machines. Governments can contribute policy priorities, regulatory experience, and capacity-building needs. Standards bodies such as ISO, IEEE, and IEC can translate governance concerns into technical standards, safety requirements, testing protocols, and certification pathways. Industry can provide deployment evidence, failure cases, risk data, and implementation constraints. Academia can support risk analysis, benchmarks, and long-term societal impact assessment. Civil society can represent affected communities, human rights concerns, labor impacts, and public trust. The Dialogue should combine annual high-level meetings with smaller expert working groups, written consultations, technical workshops, and pilot projects. Each workstream should produce practical outputs: risk taxonomies, governance toolkits, incident-reporting templates, interoperability recommendations, and capacity-building programs.

Several perspectives remain underrepresented in global AI governance. These include developing countries, small and medium-sized economies, non-English-speaking communities, workers affected by automation, elderly people, persons with disabilities, children, local governments, and communities directly exposed to AI deployment. For embodied AI, additional voices are especially important: factory workers, caregivers, patients, transport users, urban residents, robotics engineers, safety experts, and communities living with robots, autonomous vehicles, drones, and smart surveillance systems. These groups experience the real-world consequences of physical AI but are rarely present when governance frameworks are designed. They could be included through regional consultations, multilingual submissions, funded participation, civil-society panels, worker and community hearings, and pilot projects in both developed and developing countries. The Dialogue should not only invite global AI leaders, but also those who will live with AI systems in daily life. This would make governance more legitimate, practical, and socially grounded.

The AI Dialogue should use engagement formats that move from speeches to problem-solving. First, it could organize thematic expert labs on concrete issues such as embodied AI safety, AI incident reporting, open-source governance, AI and labor, and AI capacity-building. Each lab should produce short practical outputs, such as risk taxonomies, policy options, testing methods, or pilot proposals. Second, it could use scenario-based exercises. For example, participants could examine a real-world case involving an autonomous vehicle accident, a robotics deployment failure, a cyberattack on AI-enabled infrastructure, or misuse of generative AI. This would help governments, industry, technical experts, and civil society understand governance gaps in practice. Third, the Dialogue could create regional and multilingual consultations before the main meeting, so that developing countries, workers, local communities, and non-English-speaking groups can shape the agenda rather than only respond to it. Fourth, it should include standards-policy roundtables that bring together policymakers and bodies such as ISO, IEEE, IEC, ACM, and ITU. This would help translate high-level governance concerns into technical standards, testing protocols, certification pathways, and implementation guidance. Finally, the Dialogue should support small pilot projects between meetings, especially on emerging topics such as embodied AI governance. This would make the process dynamic, evidence-based, and action-oriented rather than purely diplomatic.

First, risk-based governance frameworks can classify AI systems by potential harm and apply stronger requirements to higher-risk uses, especially in healthcare, transportation, education, employment, public services, and physical AI systems such as robots and autonomous vehicles. Second, safety testing and certification should become standard practice. For embodied AI, this means not only testing models, but also testing the full system: sensors, decision-making, control, cybersecurity, human override, failure modes, and real-world deployment conditions. Third, AI incident reporting platforms should be developed internationally. Governments, companies, researchers, and users should be able to report safety failures, misuse, near misses, bias incidents, cyberattacks, and physical AI accidents. This would allow governance to learn from real deployment evidence. Fourth, international technical standards are critical. ISO, IEEE, IEC, ITU, and related bodies can translate governance goals into concrete standards for safety, data quality, transparency, interoperability, robotics, and human oversight. Fifth, open-source platforms, open datasets, benchmarks, and shared evaluation tools can support more inclusive governance, especially for developing countries and smaller organizations that lack access to proprietary AI infrastructure. Finally, regulatory sandboxes and pilot zones can help test governance approaches before large-scale deployment. This is particularly important for embodied AI, where the impact of AI systems depends heavily on the physical environment, users, and social context.