"What is a true AI car? It is a question asked every time, yet a standard answer remains elusive."
Recently, at the 4th AI-Defined Vehicle Forum 2026, Meng Chao, Senior Director of the Intelligent Software Center of SAIC Passenger Vehicle and CTO of Z-ONE Technology, raised this question.
However, he is not without his own answer.
In Meng Chao's view, a true AI car is not just about adding a voice assistant or a specific autonomous driving feature, but rather the ability to understand user goals, break down tasks, and then invoke the vehicle's overall capabilities to accomplish them. Judging in-car AI should also shift from "what model is used" to "what complete tasks can it accomplish for the user."
He provided a very specific example.
It is raining at night, and the child is already asleep in the back seat. The user gets in the car and says only one sentence: "Keep it quiet, find a pharmacy that is still open along the way, and remind me when we are nearby."
In the past, this might have meant turning down the volume, searching for a pharmacy, altering the navigation route, and setting a reminder, with each function operating independently.
What the AI car, as understood by Meng Chao, needs to do is first figure out: the user wants to buy medicine but does not want to wake the child or take a long detour. Then, it invokes navigation, the cockpit, vehicle control, and even external services to complete the entire task.
The difference is not just that the car infotainment system is better at chatting, but that the car begins to try to understand: what exactly you want to do.
Questions also arise from this. If this qualifies as an AI car, does being good at chatting qualify? Does being able to drive qualify? Does AI controlling the chassis qualify? If AI has already entered vehicle R&D, testing, and even after-sales service, what conditions must a car meet to deserve being redefined as an "AI car"?
In 2026, this question has not become any simpler.
Instead, the answers are multiplying.
"AI Cars" Are Suddenly Everywhere
XPeng P7+ was once crowned the "world's first AI car"; Roewe Jiayue 07 proposed the "world's first AI-native car"; AIVA shouted "AI defines cars, AI first, then the car"; IM Motors started talking about Physical AI, Geely talked about the cockpit-driving fusion super intelligent agent, Great Wall Motor continued to push AI to the bottom layer of the operating system, and FAW Group defined AI Car as a data-driven, continuously evolving vehicle intelligent agent.
At first glance, everyone seems to be telling the same story.
However, when examined closely, it is not the same thing at all.
XPeng focuses more on how AI drives. End-to-end, VLA, and world models are all core to enabling the car to navigate complex roads without relying solely on pre-programmed rules.
The "AI-native" concept emphasized by Roewe focuses more on whether AI can truly invoke the vehicle's capabilities. The Jiayue 07 exposes a massive number of SOA (Service-Oriented Architecture) service interfaces, allowing seats, lighting, audio, and even autonomous driving capabilities to be uniformly orchestrated.
Geely has taken a step further, aiming for AI to be more than just a standalone model within the cockpit or autonomous driving system. Instead, through a world behavior model and multiple agents, it seeks to connect the cockpit, driving, and other vehicle domains.
For Great Wall Motor, the focus has shifted further down to the operating system. Coffee AI OS 4 no longer emphasizes "adding AI to the OS," but rather embedding AI capabilities into the system foundation to invoke vehicle capabilities from the bottom up.
This is why Chen Xiaofeng, Vice President and Chief Scientist of Great Wall Motor Technology Center, said: "Large models can be creative, but the operating system must never hallucinate."
FAW Group's answer is somewhat different.
Xu Baojun, Chief Engineer of the EE Architecture at the FAW Group R&D General Institute, described the AI Car as a vehicle intelligent agent capable of perception, decision-making, execution, and continuous evolution. At the same time, he emphasized the model matrix, agents, permission sandboxes, and traceable mechanisms.
If these various approaches must be compressed into a single sentence:
XPeng is answering how AI drives, Roewe is answering how AI invokes a car, Geely is answering how AI connects the cockpit and driving, Great Wall Motor is answering how AI enters the bottom layer of the car, and FAW is trying to answer how AI finally forms a complete vehicle system.
This is also why there is still no standard answer for "AI cars" today.
It is not that no one has provided an answer, but rather that the yardsticks everyone is holding are inherently different.
Why has there been such a concentrated explosion in just the past two years?
Because several previously relatively independent technological paths have finally begun to converge.
Central computing and SOA enable vehicle capabilities previously scattered across different controllers to be uniformly invoked; end-to-end, VLA, and world models reduce intelligent driving's reliance on manually coded rules; while large models and agents introduce intent understanding, memory, task decomposition, and tool invocation.
In the past, so-called smart cars were more about continuously adding smart features to a vehicle.
The change happening today is that AI is beginning to participate in how these capabilities are generated, how they are combined, and how they continue to evolve on their own.
Meng Chao summarized the difference between software-defined vehicles and AI-defined vehicles very directly: in the software era, capabilities were mostly predefined by engineers; in the AI era, some capabilities begin to be generated by models through data, learning, and reasoning within safety boundaries.
Great Wall Motor provided a more vivid analogy as early as 2024: in the past, it was about "teaching machines how to drive," but in the future, it may shift to "teaching AI to drive."
The technological logic has indeed changed. However, before the technological paths have even converged, commercial competition has already begun.
"AI car" happens to be a label that automakers find hard to abandon. Whoever proposes a definition first is also more likely to turn their technological advantages into a new industry narrative.
This is why terms like "AI car," "AI-native," "Physical AI," and "Super Intelligent Agent" keep emerging.
Technology is still searching for answers, while the market has already begun competing for the right of definition.
This may well be the most authentic state of "AI cars" today.
What Consumers Want May Not Be "AI Cars" at All
The industry's enthusiasm for "AI cars" does not mean that consumers are also waiting for a new vehicle category.
Users certainly demand intelligence.
McKinsey's China Auto Consumer Survey this year shows that 69% of respondents now consider advanced autonomous driving features like urban NOA as standard requirements for car purchases, and 84% expect in-car intelligent agents to proactively provide services rather than just passively following instructions.
However, the other side of the coin is equally worth examining.
J.D. Power's research on the intelligent experience in China's auto market shows that while smart configurations are still rapidly proliferating, the functional performance index has seen its first decline in four years, with "unstable/inaccurate operation" becoming one of the most prominent issues with advanced configurations.
Putting these two results together is actually quite interesting.
Consumers want cars to become increasingly smart, but they are also increasingly unwilling to pay for "immature smartness."
For the vast majority of ordinary users, their needs may be quite basic: navigation should not take unnecessary detours, voice commands should be understood on the first try, parking assist should work consistently rather than failing the next day, and proactive services should not turn into proactive harassment.
As for whether it is powered by VLA, world models, or multiple agents behind the scenes, it mostly does not matter.
Users Buy Results, Not Tech Stacks.
This is also where the "AI car" label becomes somewhat awkward.
In the cockpit, the more obvious the AI, the easier it is to become a selling point. It can chat, remember the user, recommend restaurants, and understand vague prompts. Even if a recommendation is occasionally inaccurate, the consequences are limited.
Recommending Restaurant A today and Restaurant B tomorrow will not cause most people to feel safety anxiety about the car.
However, AI integration on the industry side has already moved beyond the consumers' line of sight.
If consumers are told that AI is now involved in the R&D, testing, and fault diagnosis of their car, or even partial driving and vehicle control, their first reaction might not be "how advanced," but rather is it reliable?
Over the past few years, generative AI has already undergone a unique round of "consumer education."
Everyone knows AI is smart, and everyone also knows it can make mistakes with absolute confidence.
This leads to a somewhat counterintuitive result: the closer AI is to the consumer, the more it needs to make its presence felt; the closer AI is to the core of the vehicle, the more it needs to make people forget it exists.
Users See AI, While Automakers Have Already Started "Using" AI
The proliferation of AI in the automotive industry is actually no longer just about "putting AI in the car."
SERES proposed building an "AI-native enterprise" this year, further integrating AI into R&D, manufacturing, and services.
FAW Group has already applied its proprietary large models and AIGC to creative vehicle design.
Great Wall Motor has also deployed enterprise agents for engineering R&D.
Fang Yu, Co-founder and CTO of Sonatus, has pushed AI further into testing, diagnostics, and after-sales service.
He mentioned a case where combining R&D information with vehicle test data, and using AI to assist in analyzing the root cause of issues, reduced the investigation time for complex problems from about two weeks to around two days.
Of course, the practices of these leading companies do not mean the entire automotive industry has reached the same stage.
Different automakers vary significantly in their data accumulation, software capabilities, organizational structures, and even their willingness to invest costs.
However, at least one thing is becoming increasingly clear: the automotive industry's understanding of AI is expanding from "what AI features to offer consumers" to "how to use AI to build cars."
One path is inside the vehicle. The cockpit, autonomous driving, intelligent agents, and proactive services all ultimately require firsthand experience from the consumer.
The other path is hidden behind the scenes. Design, R&D, testing, production, fault diagnosis, and after-sales service—users may never see most of the AI involved in these processes.
Both paths are indispensable. The cockpit and autonomous driving determine whether consumers are willing to believe in "AI cars," while industry-side AI determines whether this belief is backed by engineering capabilities.
What is truly worth noting about industry-side AI is not that it is "more important" than the cockpit or autonomous driving.
Rather, it has changed something rarely discussed in the context of AI cars: AI is not only becoming a capability of the vehicle but is also becoming a tool to build it.
As this path goes deeper, the boundaries of what constitutes an "AI car" will become increasingly blurred.
AI participates in product definition, design and development, as well as testing and manufacturing; after the vehicle is delivered, it continues to exist in the cockpit, driving, diagnostics, and after-sales service.
Then does the so-called "AI car" refer to a product that uses AI, or an automotive industry system that is being reshaped by AI?
This may be far more worth discussing than arguing over who has the "world's first AI car."
The Real Challenge Is: How Much Decision-Making Power to Hand Over to AI
When the issue reaches this point, it truly touches the bottom line of the automotive industry.
Fang Yu gave a very simple example at the forum: "When asking for restaurant recommendations in the cockpit, it doesn't matter if the answers vary; but during vehicle diagnostics, AI must guarantee deterministic accuracy."
A single sentence separates the two worlds.
What AI truly excels at is navigating complex, open environments where rules cannot be entirely predefined, finding patterns from massive amounts of data, and making judgments under conditions of uncertainty.
What the automotive industry has long pursued, however, is something else: verifiability, repeatability, and traceability.
There can be ten reasonable answers for a restaurant recommendation. But for why a car triggers a fault light, the answer cannot be one thing today and another tomorrow.
If AI generates an unappealing styling image, the designer can simply delete it. If an incorrect hypothesis is proposed during the R&D phase, engineers are still there to review it.
But once AI begins to influence fault diagnosis, autonomous driving decisions, chassis control, or even vehicle execution, "let's just try it and see" is no longer an acceptable approach.
Therefore, while the definition of "what qualifies as an AI car" is still being debated, another set of rules can no longer wait.
That is safety.
What the product is called can continue to be debated.
However, once AI enters safety-critical processes, it must answer much more specific questions.
What tasks can be delegated to the model, and what must be verified by traditional rules? Which outcomes allow for probabilistic judgments, and which must meet deterministic safety conditions?
These are boundary issues.
And beyond the boundaries, there is the fallback issue: if the AI makes a mistake, can the system detect it? Who provides the fallback? And how should accountability be determined?
ISO/PAS 8800 has already begun establishing a dedicated framework for safety-related AI in road vehicles, and domestic standards for AI model testing, risk governance, functional safety, and data and information security are also advancing.
This is also why companies that appear to be on completely different technological paths are ultimately starting to say similar things.
Meng Chao emphasizes safety boundaries; FAW sets permission sandboxes for agents; Great Wall Motor demands that the operating system must not hallucinate; Fang Yu insists that diagnostic results must be reliable.
The truly difficult question has never been whether cars can use AI, but rather how much decision-making power the automotive industry is actually willing to hand over to AI.
Returning to the very beginning.
"What is a true AI car?" Today, it is not that there is no answer; it is just that the answers have not yet converged.
And the question consumers truly care about may never have been that complex: what exactly has AI made better about this car?
The automotive industry must also answer another question: can this "improvement" be achieved in a stable, reliable, and verifiable manner?
If these two questions can be answered, what "AI cars" are called may truly not matter that much.
In fact, one day, when large models, agents, AI autonomous driving, and AI-assisted R&D all become default capabilities in the automotive industry, the "AI car" label itself may quietly fade into the background.
What truly determines whether AI can redefine the automobile is not which press conference it appears at.
It is whether AI can help the vehicle make better judgments when rules cannot fully cover a scenario—and ensure that this judgment can withstand the scrutiny of safety, reliability, and verifiability.
—END—