For most of functional safety’s history, the methods assumed a system you could fully specify and exhaustively analyze. AI components break that assumption: their behavior is learned, not written, and the conditions under which they fail are not always knowable in advance. ISO/PAS 8800 is the standards community’s first published attempt to bridge that gap for road vehicles.
01What ISO/PAS 8800 actually covers
The document sits alongside ISO 26262 and ISO 21448 (SOTIF) rather than replacing them. It addresses the safety of AI and machine-learning components used within a vehicle’s safety-related functions — how to specify them, how to argue about their performance, and what evidence a credible safety case needs when part of the system was trained rather than designed.
02Where independent judgment still matters
A published specification does not make an AI component safe; it gives teams a shared vocabulary for the argument. The hard questions — what counts as sufficient performance, how to bound the operational domain, when residual risk is acceptable — still demand experienced safety engineers willing to challenge a model’s assumptions rather than rubber-stamp them.
As this class of standard matures, the teams that fare best will be the ones treating AI safety as an engineering discipline with evidence and traceability, not a compliance afterthought.