AI is moving from infotainment into steering, braking, and other core driving functions, raising safety questions that matter far beyond one market. Traditional testing rules are no longer enough, Irene Chuang, senior business manager for Industrial Services and Information Security at TÜV Rheinland Taiwan, said at a Taiwan industry forum, as regulators and automakers worldwide confront black-box AI risks in vehicles.
AI takes the wheel, and testing struggles to keep up
The global automotive industry is entering a major shift as AI becomes embedded in autonomous-driving decisions, rather than being confined to voice assistants or dashboard features. That shift is forcing companies to rethink how vehicles are verified for safety — particularly, Chuang said, as road conditions, sensor performance, and model behavior interact in unpredictable ways.
Much of that verification work still rests on two established frameworks. Chuang said the sector has long depended on ISO 26262 for functional safety and ISO 21448, also known as SOTIF, for safety of the intended functionality. ISO 26262 focuses on software and hardware faults in electrical and electronic systems, while ISO 21448 addresses risks that arise even when no component has failed, such as poor performance in difficult weather or unusual traffic scenes.
Where existing standards fall short
Neither framework, however, fully covers the decision-making logic of AI models, which can be hard to interpret and may generate unpredictable outputs, Chuang said. To address that gap, ISO PAS 8800 has emerged as a standard aimed specifically at AI safety in road vehicles.
That gap is also why, as vehicle systems grow more complex, Chuang argued automakers cannot rely on any single standard — ISO 26262, ISO 21448, and ISO PAS 8800 need to be applied together in a coordinated way. As an example, she said an autonomous car that fails to brake in time could reflect a sensor defect, a recognition problem caused by rain or fog, or an error in the AI model itself.
More data isn't always better
Standards aside, Chuang also warned against a common belief that more training data automatically improves AI performance. Instead, data quality, diversity, and labeling accuracy matter more than sheer volume, she said, because repetitive or narrow datasets can lead to overfitting and weaker performance in rare or extreme conditions.
Taiwan's stake in the standards race
These pressures are mounting as global AI rules tighten, including the EU's AI Act, which imposes stricter data governance requirements on high-risk applications. Against that backdrop, Taiwan, a key part of the automotive electronics supply chain, stands to benefit from early adoption of verification systems aligned with international standards, which could help local companies compete in global markets.
Article translated by Jingyue Hsiao and edited by Jerry Chen
