The Current State and Future Trends of Automotive AI Safety Governance

Engineering notes

Key topics: autonomous driving, end-to-end driving, end-to-end, waymo, tesla fsd, tesla, cruise, adas, large language model, deployment, perception, prediction. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The Current State and Future Trends of Automotive AI Safety Governance Author: [Yongshou Ma] Affiliation: [Mercedes-Benz Group China] Date: August 2026 Keywords: automotive AI, AI governance, EU AI Act, ISO/PAS 8800, SOTIF, responsible AI, end-to-end driving, agentic AI, in-cabin LLM, ADAS, autonomous driving Abstract The rapid embedding of artificial intelligence into road vehicles — from large language model (LLM)-powered in-cabin assistants to end-to-end neural networks that perceive, plan, and act on the road — has outpaced the governance frameworks designed to keep it safe. This paper maps the global regulatory landscape through a heat map that distinguishes strongly-regulated from development-priority markets, surveys the AI governance practices of major Western, Chinese, and Japanese/Korean original equipment manufacturers (OEMs), analyzes the "agent-ization" of three automotive AI domains (human–vehicle interaction, in-vehicle functions, and intelligent driving), and proposes a forward-looking framework for responsible, controllable, unbiased, and safe automotive AI. Drawing on the EU AI Act (Regulation 2024/1689), UNECE Regulations R155/R156/R157, ISO 26262, ISO 21448 (SOTIF), ISO/PAS 8800:2025, the UNECE–WHO "12 Principles for AI in Road Traffic," the NIST AI Risk Management Framework, and concrete OEM disclosures from Mercedes-Benz, Volkswagen, Tesla, BYD, NIO, XPeng, and others, this article argues that the next phase of automotive AI safety will depend less on a single prescriptive rulebook and more on the convergence of sectoral standards, internal AI management systems (e.g., ISO/IEC 42001), and demonstrable post-market AI assurance. 1. Introduction Between 2024 and 2026, the automotive industry crossed three thresholds simultaneously. First, LLM-based agents entered the cabin at scale: Mercedes-Benz reported more than one million vehicles running ChatGPT-enabled MBUX voice interactions, Volkswagen integrated ChatGPT into its IDA assistant across multiple model lines, and Chinese OEMs (NIO NOMI, XPeng XOS 5.0, Li Auto Mind GPT) deployed in-house multimodal models with function-calling capabilities that allow the car to take actions, not merely answer questions. Second, end-to-end neural driving stacks — in which perception, prediction, and planning are subsumed by a single learned model — moved from research demonstrations (Wayve LINGO/GAIA, Tesla FSD V12) to consumer-grade deployments in mass-market vehicles (XPeng XNGP, Huawei ADS 3.3, NIO's NWM world model). Third, cockpit-driving integration ("舱驾一体") became a stated product strategy, collapsing the historical boundary between the entertainment/ADAS domains on a single SoC (Qualcomm Snapdragon Ride Flex, NVIDIA DRIVE Thor), enabling a unified "agentic" loop that spans cabin and road. Each of these transitions undermines assumptions embedded in the existing safety architecture. Classical automotive functional safety (ISO 26262) was designed for deterministic E/E systems; it has had to be supplemented by ISO 21448 (SOTIF) for hazards arising from intended-function insufficiency, and now by ISO/PAS 8800:2025 for hazards arising specifically from machine learning [1]. Regulators, meanwhile, have moved from voluntary guidance to binding horizontal rules: the EU AI Act (Regulation 2024/1689) entered into force on 1 August 2024 and will impose high-risk obligations on automotive AI systems that are safety components subject to type-approval under Regulation (EU) 2018/858 by 2 August 2026 [2,3]. In parallel, the UNECE–WHO "12 Principles for AI in Road Traffic" (April 2024, with subsequent 2025 amendments through the WP.29/GRVA framework) restate a normative baseline — most pointedly that "decisions that affect life and death must never be delegated to machines" [4]. At the same time, a sequence of high-profile incidents — the December 2023 recall of roughly two million Tesla vehicles over Autopilot, the October 2023 Cruise pedestrian-drag event in San Francisco, the May 2025 fatal Waymo crash in Los Angeles, and the cumulative ~47 AV-related fatalities logged by NHTSA since 2019 — have made accountability, transparency, and bias testing board-level concerns [5,6,7]. This paper is organized around the four questions that the industry, regulators, and academic observers are now forced to answer together. Section 2 maps the global legal landscape, distinguishing strongly-regulated from development-priority markets, and presents a heat map of regulatory intensity across twelve jurisdictions. Section 3 surveys how the major OEM groups — Western (Tesla, Mercedes-Benz, BMW, Volkswagen, Ford, GM, Stellantis), Chinese (BYD, NIO, XPeng, Li Auto, Geely, Xiaomi, Huawei), and Japanese/Korean (Toyota, Honda, Hyundai) — have publicly organized their internal AI governance. Section 4 decomposes the agent-ization trend into its three operational domains and identifies the cross-cutting safety, validation, privacy, and bias challenges. Section 5 proposes a practical forward framework for responsible automotive AI, drawing on technical safeguards (explainability, formal verification, simulation, red-teaming), organizational safeguards (AI ethics committees, ISO/IEC 42001 management systems), and regulatory-convergence levers. Section 6 concludes. 2. Global AI Governance Landscape: A Heat Map of Regulatory Intensity 2.1 Methodology Mapping "regulatory intensity" for a sector as heterogeneous as automotive AI requires three orthogonal dimensions: (i) whether binding horizontal AI law exists and what risk tiers it imposes; (ii) the maturity of vertical automotive AI/type-approval regulation (UNECE, GB/T, FMVSS); and (iii) the operational governance infrastructure (AI safety institutes, conformity-assessment bodies, mandatory disclosure regimes). Markets that score high on all three are classified as strongly regulated; those that combine voluntary frameworks with light-touch horizontal law and aggressive industrial policy are classified as development-priority. The classification is necessarily simplified — no jurisdiction is monolithic — but it is useful for OEM compliance planning. 2.2 Strongly regulated markets European Union. The most prescriptive jurisdiction. The AI Act (Regulation 2024/1689), published in the Official Journal of the European Union on 12 July 2024 and in force since 1 August 2024, classifies an AI system as high-risk when it is a safety component of a product covered by EU harmonization legislation listed in Annex I — explicitly including Regulation (EU) 2018/858 on motor-vehicle type-approval [2,3]. For automotive OEMs this means that, by 2 August 2026, AI components underpinning type-approved safety functions (automated driving, emergency braking, lane-keeping, steering, DMS) must satisfy a binding catalogue of requirements: risk management, data governance, technical documentation, logging, transparency, human oversight, and conformity assessment. Prohibited-practice provisions took effect on 2 February 2025, and general-purpose AI (GPAI) obligations on 2 August 2025. Sectors that already have a sectoral safety regulator — including automotive — were deliberately chosen as the first tranche of high-risk obligations, on the theory that they would face the least disruption [2,3]. The European automotive industry, through ACEA, has nonetheless pressed for coherence with existing UNECE type-approval to avoid duplicative assessments [3]. United Kingdom. The UK has chosen a pro-innovation, sectoral approach: the government has explicitly declined to adopt horizontal AI legislation in the near term, but has established the AI Safety Institute (now the AI Security Institute) as a state-backed evaluator of frontier model risk, and has begun a phased rollout of binding regulation for the most impactful sectors. For automotive AI, the operative instruments remain UNECE-derived type-approval and the existing product-safety regime, but the UK AI Security Institute has begun engaging with OEMs and Tier-1s on third-party testing of in-cabin LLMs and driver-monitoring systems (the UK is not formally a contracting party to UNECE 1958, but accepts equivalent approvals). China. China's regime combines vertical sectoral rules with a fast-evolving horizontal AI layer. The Interim Measures for the Management of Generative AI Services (effective August 2023) and the Measures for the Management of Synthetic Content Generated by AI (September 2025) require algorithm filing, content labeling, and security assessments for generative AI deployed to the public — including in-cabin LLM assistants. On 9 July 2024 the Cyberspace Administration of China, jointly with the Ministry of Industry and Information Technology and others, released the Global AI Governance Initiative and, in 2024–2025, a series of Safety Governance Frameworks for generative AI, embodied intelligence, and industry-specific AI [authoritative English texts of the latest framework updates were being finalized in mid-2026]. Crucially, China's automotive regulators (MIIT, MoT) operate the most permissive L3/L4 testing regime in the world: by 2025, MIIT had approved intelligent-connected vehicle (ICV) road-testing pilots in seven major cities covering more than 32,000 km of designated roads, and pilot commercial L3/L4 deployments were under way in Beijing, Shanghai, Shenzhen, Chongqing, Wuhan, and Hangzhou. China is therefore a hybrid: it regulates outputs (algorithmic content, transparency) firmly, while leaving deployment latitude wide. United States (federal, with state-level pressure). At the federal level, the U.S. has oscillated between two postures. President Biden's Executive Order 14110 (October 2023, "Safe, Secure, and Trustworthy Development and Use of AI") was rescinded by President Trump's Executive Order 14179 (January 2025, "Removing Barriers to American Leadership in AI"), which directed OSTP and OMB to develop an AI Action Plan emphasizing innovation, infrastructure, and

Links and sources

Comments