At the 2026 AI-Defined Vehicle Forum, Meng Chao, senior director of SAIC's intelligent software center and CTO of Z-ONE Technology, posed a question that has no easy answer: "What is a true AI car?" According to the Chinese outlet D1EV, Meng argued that an AI car is not merely one with a voice assistant or an added driver-assistance feature, but a vehicle that can understand a user's goal, break it down into tasks, and coordinate the entire vehicle to accomplish it.
Meng illustrated his point with a scenario: a user gets into the car at night in the rain, with a child asleep in the back seat, and says, "Keep it quiet, find a pharmacy that's still open on the way, and remind me when we get close." In a conventional car, this would require manually lowering the volume, searching for a pharmacy, changing the navigation destination, and setting a reminder. An AI car, in Meng's view, would first understand that the user wants to buy medicine without waking the child or going too far out of the way, and then coordinate navigation, cabin controls, vehicle controls, and even external services to complete the task seamlessly.
The forum highlighted how different automakers are approaching the concept. Xpeng, which has called its P7+ the "world's first AI car," focuses on how AI drives, using end-to-end models, VLA (vision-language-action), and world models to handle complex road situations without relying on pre-programmed rules. Roewe, which has touted its JY07 as the "world's first AI-native car," emphasizes AI's ability to call upon vehicle functions through open SOA service interfaces, allowing unified control of seats, lighting, audio, and even intelligent driving. Geely is working on a "cockpit-driving fusion super intelligent agent" that connects cabin and driving domains through world behavior models and multiple agents. GWM is embedding AI into the operating system itself, with its Coffee AI OS 4, rather than adding AI on top. Chen Xiaofeng, vice president of GWM and chief scientist at its technology center, said, "Large models can be creative, but operating systems must never have hallucinations." FAW, meanwhile, describes AI Car as a vehicle-level intelligent agent that perceives, decides, executes, and continuously evolves, with mechanisms for model matrices, agents, permission sandboxes, and traceability.
Why no consensus?
The lack of a standard definition stems from the convergence of several previously separate technology tracks. Central computing and SOA enable unified control of capabilities that were once scattered across different controllers. End-to-end models, VLA, and world models reduce reliance on manually written rules for intelligent driving. Large models and agents bring intent understanding, memory, task decomposition, and tool invocation. The shift, as Meng put it, is from capabilities pre-defined by engineers to capabilities generated by models through data, learning, and reasoning, within safety boundaries.
But the technical logic has changed before the technology has converged, and commercial competition has already begun. The "AI car" label is hard for automakers to abandon, as whoever defines it first can turn their technical advantages into a new industry narrative. Hence the proliferation of terms like "AI car," "AI-native," "physical AI," and "super intelligent agent."
Consumer enthusiasm, however, does not necessarily mean they are waiting for a new vehicle category. According to a McKinsey survey cited in the article, 69% of Chinese car buyers already consider high-level driver assistance like urban NOA (city advanced driver assistance) a standard feature, and 84% want in-car agents to proactively offer services rather than passively follow commands. Yet J.D. Power's China intelligent vehicle experience study shows that while intelligent features are spreading rapidly, the functional performance index has declined for the first time in four years, with "unstable or inaccurate" being a top complaint. Consumers want smarter cars, but they are increasingly unwilling to pay for immature intelligence.
For most ordinary users, the needs are simple: navigation that doesn't detour unnecessarily, voice recognition that works on the first try, parking assistance that works consistently, and proactive services that don't become intrusive. Whether the underlying technology is VLA, world models, or agents rarely matters to them. This is the awkward position of the "AI car" label. In the cabin, AI is a clear selling point—it can chat, remember the user, recommend restaurants, and understand vague commands. Even if a recommendation is off, the consequences are limited. But when AI begins to participate in vehicle development, testing, fault diagnosis, and even driving and control, the consumer's first reaction may not be "advanced" but "is it reliable?"
