{"id":26767,"date":"2026-08-04T10:00:56","date_gmt":"2026-08-04T09:00:56","guid":{"rendered":"https:\/\/neocoeurintelligence.com\/?p=26767"},"modified":"2026-07-23T23:05:17","modified_gmt":"2026-07-23T22:05:17","slug":"ai-autonomous-driving-sae-xai-and-simulation-at-the-heart-of-safety","status":"publish","type":"post","link":"https:\/\/neocoeurintelligence.com\/fr\/ai-autonomous-driving-sae-xai-and-simulation-at-the-heart-of-safety\/","title":{"rendered":"IA & Conduite Autonome : SAE, XAI et Simulation au c\u0153ur de la s\u00e9curit\u00e9"},"content":{"rendered":"\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t
\n\n
\n\n
\n
ART-13<\/strong> \u00b7 Neo Coeur Intelligence \u00b7 Published: July 2026 \u00b7 Mobility vertical \u00b7 Based on SAE J3016 \u00b7 ISO 21448 (SOTIF) \u00b7 ISO\/PAS 8800<\/div>\n<\/div>\n\n
\n

\ud83d\udccb TABLE OF CONTENTS -<\/h4>\n
00.<\/span>Introduction<\/div>\n
I.<\/span>General Overview: Autonomous Driving and the Safety Equation<\/div>\n
1.1<\/span>The 6 SAE levels of driving automation<\/div>\n
1.2<\/span>Why AI safety is a new discipline (ISO\/PAS 8800)<\/div>\n
1.3<\/span>The EU AI Act and high-risk ADAS<\/div>\n
II.<\/span>The Three Pillars of Autonomous Driving Safety<\/div>\n
2.1<\/span>Explainable AI (XAI) for critical decisions<\/div>\n
2.2<\/span>Simulation: CARLA, SUMO and the billion-mile problem<\/div>\n
2.3<\/span>Adversarial robustness: the NCI angle<\/div>\n
III.<\/span>A 3-Step Safety Assurance Method<\/div>\n
IV.<\/span>Conclusion<\/div>\n
V.<\/span>FAQ<\/div>\n
VI.<\/span>References<\/div>\n<\/div>\n\n
\n
\n\n

Introduction<\/h2>\n
<\/div>\n\n

At SAE Level 3, liability shifts from the driver to the machine. And no one has fully settled who pays when a Level 3 vehicle causes an accident. That single sentence captures why autonomous driving is not just an engineering challenge. It is a safety, legal and cybersecurity problem rolled into one, and it lands squarely on the desk of every OEM, Tier-1 supplier and mobility operator in Europe.<\/p>\n\n

The promise is real. Around 94% of road accidents involve human error, so removing the human from the loop should, in theory, save lives. But an autonomous vehicle does not \"see\" the road. It runs a stack of statistical models that infer, from camera, lidar and radar inputs, what to brake for and what to ignore. When those models are wrong, opaque or manipulable, the safety gain evaporates. This guide sets out how sovereign, explainable and resilient AI actually secures autonomous driving in 2026, and how it connects to the broader discipline of AI risk assessment and adversarial AI<\/a>.<\/p>\n\n

We will cover the six SAE levels and where responsibility breaks, the regulatory stack that now governs in-vehicle AI (ISO\/PAS 8800 and the EU AI Act included), the three technical pillars that make an autonomous stack defensible, and a practical method to prove safety before a regulator or an insurer asks you to.<\/p>\n\n

\n

\u26a1 WHERE WE STAND, 2026<\/h4>\n

Mercedes-Benz and Honda have both homologated SAE Level 3 systems for conditional automated driving in defined operational domains. That milestone matters because it is the first point where the vehicle, not the human, is legally the driver during activation. Every OEM chasing that certification now has to prove not only that the AI works, but that its decisions are explainable, auditable and robust against manipulation. This is the same evidence burden imposed by UN-R155, ISO\/SAE 21434 and the new ISO\/PAS 8800, extended to the machine-learning layer.<\/p>\n<\/div>\n\n

I. General Overview: Autonomous Driving and the Safety Equation<\/h2>\n
<\/div>\n\n

1. The 6 SAE Levels of Driving Automation<\/h3>\n\n

The SAE J3016 standard defines six levels of driving automation, from 0 to 5. The distinction that matters most for cybersecurity and liability is the jump from Level 2 to Level 3, because that is where the human stops being the fallback and the system becomes responsible for the driving task within its operational design domain.<\/p>\n\n

\n
\u25cf<\/span>\u25cf<\/span>\u25cf<\/span><\/span>FIGURE 01 \u00b7 SAE J3016 \u00b7 SIX LEVELS OF DRIVING AUTOMATION<\/span>SAE J3016<\/span><\/div>\n
\n
0<\/div>
No Automation<\/div>
The human does everything. Warnings and momentary assistance only.<\/div><\/div><\/div>\n
1<\/div>
Driver Assistance<\/div>
One function automated (steering OR speed). Human monitors permanently.<\/div><\/div><\/div>\n
2<\/div>
Partial Automation<\/div>
Steering AND speed automated. Human must supervise and take over instantly. Most \"self-driving\" features sold today stop here.<\/div><\/div><\/div>\n
3<\/div>
Conditional Automation \u00b7 liability shift<\/div>
The system drives within its domain. The human can disengage but must resume when requested. Mercedes and Honda are homologated here.<\/div><\/div><\/div>\n
4<\/div>
High Automation<\/div>
No human fallback needed inside the operational domain (robotaxi zones, defined routes).<\/div><\/div><\/div>\n
5<\/div>
Full Automation<\/div>
No steering wheel required. Any road, any condition. Not commercially available.<\/div><\/div><\/div>\n<\/div>\n
Source: SAE International<\/a> J3016 \u00b7 levels of driving automation for on-road motor vehicles<\/div>\n<\/div>\n\n

Below Level 3, an accident is the driver's problem. At Level 3 and above, it becomes the manufacturer's problem, and by extension a question about the integrity of the AI stack. This is the precise moment where cybersecurity stops being an IT concern and becomes a product safety obligation.<\/p>\n\n

2. Why AI Safety Is a New Discipline (ISO\/PAS 8800)<\/h3>\n\n

Classical automotive safety, governed by ISO 26262, assumes deterministic systems: a brake controller behaves the same way every time, and you can test it exhaustively. A perception model does not work like that. It is a probabilistic function trained on data, and it can fail on inputs no engineer anticipated. The industry created a dedicated standard for exactly this gap, ISO 21448 (SOTIF, Safety Of The Intended Functionality), to address hazards that arise even when nothing is technically broken.<\/p>\n\n

In 2023 the industry went one step further with ISO\/PAS 8800 (Road vehicles, Safety and artificial intelligence)<\/strong>, now the reference standard for machine-learning systems in vehicles. Where ISO 26262 covers functional safety and ISO 21448 covers the intended-functionality gap, ISO\/PAS 8800 addresses the AI itself: the safety lifecycle of an in-vehicle ML model, its data requirements, its performance and robustness targets, and the evidence needed to argue it is safe. For any OEM deploying AI-driven perception or decision-making, ISO\/PAS 8800 is no longer optional reading. It is the framework a technical service will expect you to have applied. Together, these four texts form the regulatory stack for autonomous AI.<\/p>\n\n

\n
\u25cf<\/span>\u25cf<\/span>\u25cf<\/span><\/span>FIGURE 02 \u00b7 THE REGULATORY STACK FOR IN-VEHICLE AI<\/span>STANDARDS<\/span><\/div>\n
\n
26262<\/div>
ISO 26262<\/div>
Functional safety of deterministic electronic systems<\/div><\/div>\n
21448<\/div>
ISO 21448 (SOTIF)<\/div>
Safety of the intended functionality, even when nothing is broken<\/div><\/div>\n
8800<\/div>
ISO\/PAS 8800<\/div>
Safety and AI: the ML model lifecycle, data and robustness targets<\/div><\/div>\n<\/div>\n
Plus UN-R155 (cyber management) and the EU AI Act (high-risk classification) \u00b7 the full compliance stack for autonomous AI<\/div>\n<\/div>\n\n

Three properties separate a trustworthy autonomous stack from a dangerous one. It must be explainable<\/strong>, so a human or an auditor can understand why it braked or swerved. It must be validated at scale<\/strong>, because you cannot drive a billion physical miles to prove an edge case is handled. And it must be robust<\/strong>, because an adversary who can nudge a sensor input can flip a decision. These map directly to the three pillars in Part II.<\/p>\n\n

3. The EU AI Act and High-Risk ADAS<\/h3>\n\n

The regulatory stack does not stop at automotive standards. The EU AI Act (Regulation 2024\/1689) now sits on top of them, and it treats AI systems used as safety components of vehicles as high-risk<\/strong>. An ADAS or autonomous-driving perception system is not a peripheral feature in the eyes of the Act. It is precisely the kind of system the regulation was written to govern.<\/p>\n\n

\n

\u2696 WHAT THE EU AI ACT MEANS FOR ADAS AND AUTONOMOUS SYSTEMS<\/h4>\n

AI systems that act as safety components of regulated products, which includes ADAS and automated-driving functions, fall under the high-risk regime of the AI Act (Article 6 and Annex III \/ Annex I product logic). That classification carries three consequences OEMs and suppliers must plan for:<\/p>\n