Agenda

Background.

Artificial intelligence is no longer confined to screens and data centres. It now flies drones over inhabited areas, drives ground robots through hospitals, airports and city streets, and operates assistive and surgical systems — taking decisions that act directly on the physical world, often in the presence of civilians who never consented to be near them. Yet while aviation, nuclear power and rail each rest on mature, independent safety-certification regimes, no equivalent body certifies the safety of AI systems acting in open physical environments. Consider a simple question: who certifies that an autonomous delivery robot, or an inspection drone flying above a crowd, is safe for the people around it? Today, there is no clear answer. The European AI Act addresses governance and risk classification; ISO/IEC standards are emerging; EASA has begun work for aviation. But the general case — an AI making consequential physical decisions among people — falls between existing mandates. This is a structural gap in the global digital architecture, and one the WSIS framework itself does not yet name.

Relevant practice and precedents.

Rather than claim premature successes, this session grounds its case in proven adjacent practice. The maritime industry separated insurer from certifier in 1834, recognising that one cannot impartially judge the safety of what one has an interest in deploying — the founding principle of independent third-party assurance that later built modern aviation and safety-critical software certification. More concretely, EASA's SORA methodology already shows how to assess the operational risk a drone poses to people on the ground. The methods to evaluate the safety of an autonomous system operating among people therefore exist; what is missing is the institutional home to apply them across the general cyber-physical case. The session draws on decades of safety-critical software certification to make this point precise and credible.

Vision towards WSIS+20 and beyond.

As the WSIS+20 review takes stock of twenty years of building trust in ICTs, AI acting in the physical world is the frontier where "confidence and security" must now be earned. This session offers policymakers, regulators and civil society a shared vocabulary for a gap most have sensed but few have named, and calls for an independent, pre-normative effort — based in Geneva, at the heart of the multilateral system — to define safety criteria for real-world AI. It invites governments, standards bodies, industry and civil society to help build it.