The automotive industry

The primary risks are accidents, injuries, or fatalities caused by AI errors. For example, AI may fail to recognize a pedestrian or obstacle in poor visibility conditions or in an unusual traffic situation that was not in its training dataset. As a result, it may fail to warn the driver of a collision risk or stop the vehicle in time.

On one hand, this raises risks of legal liability for AI errors. If an autonomous vehicle is involved in an accident, who bears responsibility: the manufacturer, the AI system developer, or the vehicle owner? Such situations may result in lawsuits, creating distrust and skepticism among consumers about handing over vehicle control to AI.

On the other hand, automating decision-making processes raises a number of serious ethical questions. How should AI act in critical situations where human lives are at risk? If a collision is unavoidable, should AI prioritize the safety of the HAV’s driver, or rather aim to protect the largest possible number of people [6]? At present, there is no consensus on this issue. In terms of technology, this increases the importance of research into the interpretability of AI. AI decision-making logic in various traffic scenarios must become more transparent to humans and subject to greater human oversight.

Finally, there are traditional cyberthreats, exacerbated by AI-specific vulnerabilities. Researchers have already demonstrated various attacks on AI in autonomous vehicles, such as using stickers and special adversarial patches applied to road signs[7]. In these cases, the attacked AI system confidently classified a stop sign as a speed limit sign or completely ignored certain types of signs. Adversarial attacks have been carried out not only with images but also using LiDAR data by generating false point clouds[8].

To minimize risks and threats, the development and deployment of AI in automobiles takes place within regulatory frameworks and in compliance with strict safety standards adopted by the automotive industry.

Special attention is given to the multi-stage testing of AI systems in HAVs.

Initial testing stages include evaluating the AI's performance under controlled conditions using computer simulations and hardware-in-the-loop simulations, as well as on closed test tracks. Simulation platforms such as CARLA [9] and NVIDIA Drive Sim [10], are capable of realistically reproducing a wide range of traffic scenarios.

The final stage is testing the AI on public roads in real operation, including different climatic conditions and traffic intensities, to confirm the AI’s performance in real-world use. In all countries, this type of testing is subject to government regulation in one form or another (see, for example, [11], [12]).

Consistency of requirements and technical solutions is also ensured by a system of standards. In Russia, 25 GOST standards currently govern AI for automotive transport. These standards include requirements and methodologies for testing detection and recognition algorithms for obstacles [13], traffic light signals 14], road markings [15], and traffic signs [16]. A series of GOST standards has been developed for intelligent transportation infrastructure management systems (see, for example, [17]-[19]), tools for monitoring human behavior and predicting intentions [22], and big data in automotive transport (the Autodata platform) [20, 21].

The operation of HAVs on public roads is currently allowed only within designated zones and specific scenarios, and remains under regulatory supervision. In Russia, HAV operation on public roads is governed by an experimental legal regime introduced by Government Resolution of the Russian Federation No. 309 of March 9, 2022 [23].