AIDV: The Right to Take the Wrong Turn

Mercedes-Benz's vision of a hyper-personalised customer experience: a car that understands a driver's habits and context to proactively offer the features and services they need.

A car has begun to understand its driver more precisely — and to act on that understanding. But knowing what someone wants and knowing how far you're allowed to act on their behalf are two different problems. The more choices and actions AI takes on, the more it matters that it can also ask, wait, and sometimes do nothing at all. This is the seventh and closing installment of the “Experience of the Automobile” series, gathering the threads that ran through Memory Lane, Top Gear, Systems Over Cylinders, Audi F1, Whose Experience, and Begin Again.

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기






It's Monday morning. You drink your coffee and check the day's schedule. When it's time to leave, the car is already waiting outside the garage. The cabin temperature is set, the radio is tuned to its usual station at its usual volume. You get in, and the MBUX Virtual Assistant suggests a destination. Work. It skips the drive-through you usually stop at — because it already knows you had coffee at home.
This isn't a hypothetical. It's a scene from the hyper-personalised customer experience Mercedes-Benz is building toward — a future in which the car understands a driver's habits, schedule, and needs, and acts without waiting to be asked. It prepares based on what it already knows, instead of asking each time. This is, in large part, what it means for a car to understand you.
But what if, today, you don't want to go to work. You want to go to the sea.
It's a small thing. You type a beach into the navigation, and that's that. The car suggesting work doesn't mean you have to go there. If you don't like the music, you change it. If the cabin's too warm, you adjust it. Most of these things are easily reversed. So the real question isn't whether the car blocks your choice.
The question changes once the car's ambitions go beyond guessing a destination or recommending a song. There's a real gap between the car suggesting a charging station and the car choosing one for you. Booking it is another step further. Paying moves money. And if the outcome ripples into your calendar or other connected services, the reach of that one action grows wider. An AI agent entering the car doesn't just mean the car knows you better — it means the car has started turning what it knows into action. It might correctly predict that you're heading to work. But being right about that isn't the same as having permission to drive you there on your behalf.




HARMAN's Ready Care detects a driver's cognitive and visual load and state, intervening when it matters. Safety-driven action and proactive convenience call for different standards.



The car that moves first

Cars have been moving ahead of people for a long time. When a wheel is about to lock, ABS engages. When the car begins to lose its line, ESC steps in. When a collision is imminent and the driver hasn't reacted, AEB applies the brakes. In these moments, the car doesn't ask.
It cannot afford to ask, “Should I brake now?”
The auto industry has spent decades refining this kind of judgment — detecting danger faster, reading it more precisely, intervening more quickly when it counts. What's changed is that AI moving into the cockpit has vastly expanded the range of things a car can decide to do first. But not every action deserves the logic of AEB. Set HARMAN's Ready Care and Ready Engage side by side, and the difference becomes visible.
Ready Care watches the driver's state. Using AI and sensing technology, it monitors cognitive and visual load along with breathing, detects distraction or drowsiness, and steps in with a tailored intervention when needed. When a driver is in a dangerous state, the reason to act first is obvious. Ready Engage, through an AI avatar named Luna, offers personalized interaction using voice and visual cues — anticipating what a passenger might want and responding naturally, leaning toward a more customized in-cabin experience.
Both read the person and can respond ahead of being asked. But warning a drowsy driver and proactively suggesting content you might enjoy cannot be judged by the same standard. That's not to say the second is dangerous — if you don't like the suggestion, you simply choose something else. One exists so you don't have to wait, for the sake of safety. The other exists so you don't have to wait, for the sake of convenience. Prediction accuracy alone doesn't settle how much authority to act should follow from it.




The BMW Intelligent Personal Assistant surfaces features based on a driver's past behavior. Part of personalization is knowing when to stop offering a suggestion that's been repeatedly declined.



Refusal is data too

BMW offers one example. The BMW Intelligent Personal Assistant can make suggestions based on a driver's past behavior — if you've used Sport Mode before, it might offer it again on a suitable stretch of road.
But if the driver repeatedly ignores or dismisses that suggestion, the system learns from that too. And it stops offering it. This isn't processing more data or unveiling some flashy new capability, so it may not look like impressive AI. But it's interesting precisely because it does nothing. Personalization is usually described as a technology for figuring out what a user wants — remembering favorite places, favorite music, frequently used features — so it can prepare them in advance next time. BMW's case runs the other way.
Refusal is data too. If saying yes repeatedly reveals a preference, so does saying no repeatedly. If a system is meant to learn what you want, it also has to learn what you don't want, and when to stop offering it. That doesn't mean every unanswered prompt is a rejection — a driver might have simply missed the suggestion mid-drive, or it just didn't fit the moment. A single silence should weigh less than an explicit no, a pattern of non-response, and what the driver does afterward, taken together.


 

A more recent example comes from Korea. Socar's parking platform Modu-ui-Juchajang launched an AI voice-based parking-ticket purchase service on September 17. Previously, buying a ticket meant up to twelve steps — searching a destination, comparing lots, checking details, selecting ticket type and duration, confirming the vehicle, and paying. Now, saying something like “find me parking near Gangnam Station for two hours” lets the AI understand the location, time, price range, and parking type, recommend candidates, and carry the user through to the payment step. Most of the steps disappeared — but the very last one, “confirm purchase,” wasn't removed. For privacy and payment security, Socar limited the AI's role to recommending options and reaching the payment screen, leaving the final tap to the person. What matters isn't that a human always has to press the last button — it's what's been handed off, and whether that handoff can always be taken back. Pick the wrong song, and you skip to the next one. Get the wrong destination, and you type in another. But once a payment has gone through, undoing it means a cancellation or a refund. And if it's tied to a calendar entry or another linked service, there's even more to unwind.
As cars increasingly move first, the question isn't only how often they're wrong — it's what has to be undone when they are. That's where the friction worth removing and the friction worth keeping part ways. No matter how accurate the prediction, how far to let it act still has to be decided separately.




Context-aware HMI can surface what's needed before being asked. But what to show and whether to rearrange an environment the user already knows are different questions.



The car that asks less

Automotive HMI has spent a long time evolving toward asking fewer questions. Personalization doesn't just cut down on repetition — it can also surface choices you never went looking for. You shouldn't have to reset your seat, temperature, music, and destination from scratch every time. It's better not to juggle multiple apps and re-enter your account details every time you charge.
HARMAN recently described this shift in automotive HMI as “From Buttons to Behavior.” As features multiply, so does the amount of information a driver has to process, and HMI is moving away from fixed menus toward reading context and behavior. Personalization becomes a way to surface what's needed from accumulated patterns and situational cues, cutting down on interaction. Behavior-based systems don't just look at a single input — they learn long-term patterns and respond ahead of them.
That's a good direction. None of us wants to talk to our car all the time. If the car knows you go to work every Monday, it would be strange for it to ask “Where to?” each time. If it senses you're on a trip, it might surface the information that trip calls for, unprompted.
But surfacing needed information and rearranging a dashboard you've grown used to are not the same act. As the car's AI agent takes on more, the ability to figure out what a user wants stops being enough on its own — it also has to recognize when you haven't decided yet. Sometimes it needs to ask again. Sometimes it needs to wait for an answer. And sometimes, after being turned down enough times, it needs to quietly step back. Removing every last bit of friction isn't the whole point.
At the 2026 Future Mobility Week in COEX, Seoul this past August, LG Electronics' principal researcher Hong Hyun-taek said that in the age of AIDV, travel time itself becomes part of the experience's value, and described an in-cabin AI under development that reads a passenger's state through voice, gaze, emotion, and biosignals to deliver tailored services. Suggesting content suited to someone who looks tired doesn't erase their choice — if they don't like it, they can change it. The real question is what happens to human choice as the car understands us with ever-greater precision, continually narrowing the moments where we'd otherwise have to choose again. We don't know yet. What's clear is that cars are moving toward reading behavior and anticipating intent, then acting on it first.
Last summer, in Begin Again, I wrote that no matter how well a car remembers me, it can't remember a version of me I haven't yet become. This time, the question sits closer: once a car has predicted me accurately, how far should it be allowed to carry that judgment into action?




MB.DRIVE ASSIST PRO's cooperative steering lets a driver correct the wheel even while the system is engaged — without reading that correction as a rejection of the system as a whole.



When a hand turns the wheel

Driver assistance offers a clue.
MB.DRIVE ASSIST PRO, which Mercedes-Benz unveiled at CES 2026, is an SAE Level 2 driving-assistance system that helps navigate urban routes to a destination. Thanks to what it calls cooperative steering, the driver can correct the steering at any time while the system is active, and that input alone doesn't disengage the whole system. What matters isn't the feature itself, but the fact that it doesn't read driver intervention as an outright rejection of the system, or as a full handover of control. Neither the driver nor the system has to own everything. Of course, this principle can't simply be transplanted onto music or scheduling — safety-critical driving assistance and everyday life choices each carry a different weight of responsibility and intervention.
The question is what the system should do once a human steps back in. The aliveness we saw in Top Gear didn't come from three people reading each other perfectly. It came from the fact that they diverged, made different choices, and that difference is what built the relationship. A car getting better at understanding a person doesn't mean it can substitute for that aliveness. Perhaps the job of good AI isn't to persuade more effectively, but to leave room for a person to choose differently. If a hand comes in on a Monday morning and types in the sea instead of the office, the car should treat that not as an error, but as a new choice.



The capability beyond execution

“Execution is what wins.”
Christian Sobottka, CEO of HARMAN Automotive, said earlier this year that technologies like AI and machine learning are no longer, on their own, a point of differentiation in cars. What matters more than having AI is the ability to make it actually work inside a real vehicle. As automotive AI moves past demos and into production cars, the range of things it can execute keeps growing. But before execution comes another question.
When to act. How far to act. When to ask again.
None of these can be answered by model accuracy alone. What also has to be weighed: how easily a mistake can be undone; whether money or a schedule or another service gets pulled into motion; whether safety means there's no time to wait; and whether the user's intent is really settled.
Cinemo's Ivan Dimkovic used trip planning as an example. AI can find hotels along a route, compare options, and prepare a booking. But should it go ahead and confirm the reservation and charge the credit card without asking? Some people would welcome that. Others would rather make that final confirmation themselves. The authority needed to pick a song with 95% confidence and the authority needed to complete a payment aren't the same thing. If a task has already been explicitly delegated, there may be no need to ask again. If it's still something the user is mulling over, it's better to wait. Higher AI accuracy doesn't automatically widen the scope of what it's allowed to do.
Last June, in Whose Experience, I wrote that as AI agents enter the cockpit, authority is shifting — but the rules for where that authority lands still aren't finished. This time, the question moves one step further: once a car has read a user's intent, when is it allowed to turn that into action? Act immediately, ask first, merely suggest, back off after being refused, or simply never act on something it's technically capable of doing.
How good an automotive AI is won't be measurable only by how much it does. The ability not to act, when it shouldn't, matters just as much. In an earlier piece in this series, I described the Seamless Experience that DiConium's Guang Yang spoke of as the new “ceiling” for cars in the SDV era. That direction still holds — but a seamless experience isn't one where nothing is ever asked. We don't want to open a different app every time we charge. We don't want to re-explain who we are every time we get in the car. But like the “confirm purchase” step Socar chose to keep, some friction might be better left in place.




The old Chevrolet Suburban that set this series in motion. It never predicted what the family wanted, never suggested the next destination. It simply went along with their time.



Detour, again

This series of seven pieces began, at the end of last year, with an old Chevrolet Suburban.
An elderly couple looked back on the years that car had carried them through — past a Christmas tree farm and an ice rink, pulling over on the roadside, driving down dirt roads, into an old barn. Children who once ran into that barn came out of it, at some point, as adults. The car has empty seats now. A dog that once rode along is gone too. The couple said: “It wasn't easy.” “Would you want it to have been?” “No. I'd just want it to stay the same.”
That car never learned what the family liked. It never suggested the next destination. It never made a reservation or a payment. It didn't know how anyone was feeling that day. It simply went along.
At the time, I wrote that memory sometimes comes not from choosing, but from taking the long way around. That doesn't mean the car of the future should know as little as the past did. Cars can now know far more — where we're headed, what we're listening to, where we stop, what we repeat. And they're beginning to learn, bit by bit, what we're trying to do. That's a capability that can make for a better car.
In the Audi F1 chapter, Benedikt Steil said, “We're starting from the position of a challenger, so mistakes are allowed — but stopping isn't.” A system's mistake is data for the next improvement. A person's detour can't be handled the same way. Cars have long evolved toward reducing the dangerous mistakes people make. Going forward, they'll take on more of life's tedious choices too. In front of a driver who fails to brake before a collision, the car shouldn't hesitate to step in. But choosing music you don't like, going to a restaurant no one recommended, taking a slightly longer route, or suddenly heading for the sea instead of the office one Monday — the car doesn't need to correct any of that into a better answer. The choice doesn't even have to be a good one. The trip to the sea might turn out to be a letdown. The restaurant might be terrible. The detour might really have been a waste of time. Even so, it was mine to choose.
What we'll need, in an age when cars know us better and better, isn't only the ability to eliminate every wrong choice. It's a car smart enough to know when not to move. Maybe that, too, is a way of staying with someone.

The day I take a different road — how far will that car follow me?

AEM(오토모티브일렉트로닉스매거진)

---