MIT CSAIL, EU+EIT projects

Our focus at MIT CSAIL and within our partner ecosystems and TTOs in Italy, Spain, and France includes social and emotional robotics, assistive robotics - autism, dyslexia, elders, disabilities, humanoid systems, industrial and urban robotics - including industrial, inspection, exctraction workers safety, as well as high-stakes scenarios and uncertain environments). Areas which are important for us: Developing runtime assurance frameworks for adaptive systems (beyond pre-deployment certification) Integrating edge AI constraints (latency, privacy, offline safety) into certification models Advancing sensor fusion and spatial data standards as a foundation for reliable navigation Defining human-robot interaction metrics (predictability, intent communication, safety perception) Structuring multi-stakeholder governance and auditability for shared robotic systems Leveraging simulation, digital twins, and regulatory sandboxes for validation prior to deployment A successful Global Dialogue on AI Governance should move beyond principles toward operational alignment and implementation pathways across jurisdictions. First, it should deliver shared baseline frameworks for safety, reliability, and accountability, including guidance on high-risk AI systems operating in real-world environments (e.g., healthcare, public infrastructure, and robotics). These frameworks should be actionable, measurable, and compatible with existing standards efforts (ISO, OECD, regional regulations). Second, the Dialogue should advance interoperability across governance approaches, enabling cross-border collaboration on data, models, and systems. This includes alignment on terminology, risk classification, evaluation methods, and auditability mechanisms. Third, it should prioritize capacity-building and equitable access, particularly for low- and middle-income countries, ensuring that AI deployment is not limited by infrastructure, compute, or data asymmetries. Fourth, the Dialogue should support practical validation mechanisms, such as regulatory sandboxes, shared testbeds, and real-world pilot environments, allowing safe experimentation and iterative governance. Finally, it should foster multi-stakeholder collaboration, bridging policymakers, technical experts, industry, and civil society, with a focus on translating scientific advances into trusted, deployable systems. Our input is also reflected here: OECD Sovereign AI - shorturl.at/OCUNH OECD Assistive Repo. - shorturl.at/wspGv OECD Hiroshima AI - shorturl.at/JePAl OECD AI Act and accessibility - https://oecd.ai/en/wonk/eu-ai-act-disabilities

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• AI capacity-building
• Transparency, accountability, and human oversight

1) Safe, secure and trustworthy AI, particularly as AI systems increasingly operate in high-stakes and physical environments such as healthcare, infrastructure, and robotics. Safety must extend beyond model performance to include sensor reliability (e.g., LiDAR, vision, tactile sensing), sensor fusion, and spatial intelligence (e.g., SLAM and 3D mapping), as well as system-level robustness in uncertain and dynamic environments. This includes physical AI constraints such as energy resilience, battery safety, and real-time edge operation, where compact and distributed models (e.g., VLMs combined with small language models on-device, such as in smart glasses or assistive systems) must operate reliably under latency, compute, and power limitations. 2) Interoperability of governance approaches is critical due to the fragmented global landscape. AI systems, data infrastructures, and deployment environments span jurisdictions, requiring alignment in standards, risk classification, evaluation methods, and auditability. This is particularly important for multi-sensor and multi-agent systems, where interoperability must extend to spatial data, mapping standards, and shared representations of environments, enabling consistent operation across platforms, vendors, and regions. 3) Transparency, accountability, and human oversight is needed to ensure that increasingly autonomous systems remain understandable, auditable, and aligned with human intent. This includes explainability of perception, planning, and actuation decisions, traceability of data and model updates, and clear allocation of responsibility across developers, deployers, and operators, particularly in systems that adapt over time and interact physically with humans. 4) It also requires strong user agency in the context of agentic and long-term AI systems, ensuring individuals can control their data assets, inspect and manage system memory, and exercise rights such as deletion, reversibility, particularly for companion and persistent AI systems.

1) A key cross-cutting issue is the governance of AI systems operating in physical, adaptive, and uncertain environments, including robotics, IoT, and edge AI. These systems rely on sensor fusion (e.g., LiDAR, vision, tactile sensing), spatial intelligence (e.g., SLAM, 3D mapping), and real-time decision-making, and often interact directly with humans. This challenges traditional governance models based on static evaluation, requiring a shift toward runtime assurance, continuous monitoring, and lifecycle-based certification. 2) A second emerging issue is the role of compute, energy, and hardware constraints in AI governance. Many real-world systems operate on-device using compact and distributed models (e.g., combinations of VLMs and small language models) under strict latency, power, and connectivity limitations. Governance frameworks should include edge deployment, energy resilience, battery safety, and hardware dependencies 3) Third, the governance of adaptive and learning systems. Systems that evolve over time through data feedback loops including validation, auditability, and long-term behavior, particularly in multi-agent and continuously learning environments.

The most visible in the deployment of AI in real-world, high-stakes environments, where current frameworks lag behind adaptive, sensor-driven, and edge-based systems

It should support convergence across jurisdictions by aligning risk frameworks, evaluation methods, and certification approaches, particularly for AI systems operating in physical and high-stakes environments such as healthcare, infrastructure, and robotics.

The AI Dialogue should build on and connect existing initiatives across policy, standards, and implementation layers. This includes frameworks such as the OECD AI Principles, G7 Hiroshima AI Process, and UNESCO AI recommendations, as well as regulatory and standards efforts including the EU AI Act, ISO/IEC standards, and regional data space initiatives (e.g., EHDS).

Working groups linking policy, standards, and deployment, supported by testbeds, regulatory sandboxes, and cross-sector expert exchanges.

Interactive formats such as simulation-based policy labs, real-world case demonstrations, and cross-disciplinary "implementation tracks"

Effective approaches include risk-based frameworks (e.g., EU AI Act), standards-driven certification (ISO/NIST), and regulatory sandboxes/testbeds to validate systems in real-world conditions. Additionally, data spaces and edge AI (privacy-by-design) enable interoperable, secure, and scalable deployment across sectors and regions.