Last week ended with news that grabbed the attention of pretty much every car lover: Tesla officially launched the Cybercab into its Robotaxi service in Austin — a two-seater car with no pedals and no steering wheel.

Cars that help us drive still feel almost magical to me, although they are quickly becoming the new normal. By 2025, Level 2 automation — cars capable of controlling both steering and speed while the driver remains responsible — was already fitted to well over a third of new cars sold worldwide. Even more: here in Dubai, we have spotted robotaxis gliding along with the wheel turning gently on its own, as if some invisible, very polite ghost were steering. But the Cybercab goes one step further — no wheel, no pedals, nothing. Just you, a seat and a car quietly taking care of everything.

It is the kind of news that makes you smile and ask questions at the same time. Some are almost fun: what happens if you desperately want it to break a rule? Others are rather less amusing: if a crash becomes unavoidable, how does it decide whom to save?

I probably won’t be introducing you to anything new with the trolley problem. You know the one: three people are lying on one railway track, one person on another, and a train is coming. Do you pull the lever and save three people at the expense of one? What if that one person is a child? What if it is someone you love?

These are questions most of us would happily leave to philosophers. But when you build a self-driving car, at some point they become rather less hypothetical. Because while the car may be designed to drive perfectly, we live in a decidedly imperfect world. A pedestrian can run into the road at the last second. A bus can lose control.

Take a deliberately horrible example. A collision is now unavoidable. The car has two options:

First — to continue on its path and put several pedestrians at risk.

Second — swerve and put its own driver (plus passengers if any) at risk.

With a human driver, the decision belongs to one person. Whatever happens in that split second — instinct, panic, experience or pure reflex — it is ultimately their call.

An autonomous car has no instinct to fall back on. It doesn’t panic, hesitate or make a gut decision. So when every available option carries some risk, what exactly has it been taught to prioritise — and by whom?

You might assume the obvious principle would be: protect the person inside the car. Otherwise, why would you buy a car that doesn’t prioritise your safety? That idea has certainly existed. In 2016, a Mercedes-Benz executive suggested that, if a car could definitively save at least one life, it would prioritise its occupants. But it never became an industry-wide rule, largely because it creates an uncomfortable problem of its own: prioritising the passenger at all costs would effectively mean programming a car to treat the life inside it as more valuable than the lives outside.

So what about simply saving the greatest number of people? Even that gets messy quickly. What if one person is a child and the others are adults? Should age matter at all? What about who broke the rules and caused the dangerous situation in the first place? Germany’s Ethics Commission guidelines, for example, explicitly prohibit distinguishing between people based on personal characteristics such as age or gender.

And suddenly, even “save the most people” isn’t quite as simple as it sounds.

So, no wonderfully neat answer. But until now, there has been a fairly simple escape from this philosophical mess: you. Whether you are driving a Tesla Model 3 or a BYD Seal, there is still a steering wheel in front of you. If the system does something you disagree with, you can intervene. Whatever the car has been programmed to do, the final decision can still sit, quite literally, in your hands.

And then comes the Cybercab. No steering wheel. No pedals. No way for you to take over.

So who will decide what happens when something goes wrong?

The answer is less dramatic than the trolley problem suggests. There is no universal rule telling an autonomous car to save three pedestrians over one passenger — and no single person sitting at Tesla deciding whose life matters more. Regulators set the safety requirements, manufacturers build their systems within them, and the car is designed to choose the safest available manoeuvre — one calculated to avoid a crash or, when that is no longer possible, reduce the risk of harm.

That sounds considerably more reassuring. Until you ask what exactly we mean by “safest”.

MIT’s Moral Machine experiment put versions of that question to millions of people across 233 countries and territories. Some preferences were remarkably consistent: people generally chose to save humans over animals, more lives over fewer, and younger people over older ones. But geography mattered too — for example, the preference for saving younger people was much weaker in many East Asian countries than elsewhere.

And that is where the problem becomes less about whether a car can make the calculation and more about the humans who taught it what to calculate. We may be able to programme a car to minimise harm. Agreeing on what that means is considerably harder.

Anyone who drives knows that you can learn the rules, pass the test, get behind the wheel — and still spend half your time making judgement calls. A police officer waves you through a red light. Roadworks send you across a line you would normally never cross. An ambulance appears behind you and suddenly staying exactly where the rules tell you to stay is the wrong thing to do.

With a human behind the wheel, that judgement is yours. With an autonomous car, things get rather more complicated.

And this isn’t hypothetical. In July 2026, NHTSA reported a “clear pattern” of autonomous cars struggling around emergency scenes — blocking ambulances and fire crews, or misreading flashing lights, smoke and even traffic cones.

That last one says a lot. A few cones can suddenly change how a road works. Humans read the context and adapt; a car has to be taught to. And sometimes it gets it wrong: this summer, Waymo recalled nearly 3,900 robotaxis over the risk of entering closed freeway construction zones.

Sounds too abstract? Here is what it can look like in real life. During a wildfire near Redwood City, police closed part of the I-280 and told drivers to turn around and head the wrong way down the freeway to escape. Humans understood the exception. A Waymo didn’t. It stopped, remote assistance couldn’t resolve the situation, and the car remained stuck for around 30 minutes. Eventually, a police officer had to get inside, take manual control and move it himself.

That doesn’t mean humans disappear completely. Waymo, for example, can call on Remote Assistance when it needs help making sense of an unusual situation. Cybercab passengers can hit STOP or contact Support. But here is the catch: none of those people gets the wheel. They can provide context or ask the car to stop, but they don’t suddenly become the driver.

So we haven’t quite removed human judgement from autonomous driving. We have just moved it out of the driver’s seat.

It is easy to look at all of this and think: maybe cars should just keep their steering wheels, thank you very much. But that isn’t really the point — and I am certainly not here to deliver some grand verdict on autonomous driving. New technology rarely arrives with every awkward “what if” already solved. Some problems only appear once it meets the real world. What matters is how we deal with them when they do.

And we don’t have to look far for an example. You have probably already heard about Dubai’s driverless taxis. What is more interesting is how cautiously they are being introduced: not all at once, but step by step. Dubai’s approach essentially works like this:

And one more thing: after all the trolley problems, confused traffic cones and emergency scenarios above, autonomous cars may be starting to sound slightly terrifying. But none of this means they are necessarily worse drivers. Quite the opposite: Waymo’s latest data suggest that across more than 220 million fully autonomous miles, its cars were involved in 82% fewer injury-causing crashes and 94% fewer serious or fatal ones than human drivers covering comparable distances in the same areas.

But safer doesn’t mean crash-proof. And it certainly doesn’t mean every strange situation has already been solved. Perhaps that is where we are with autonomous driving right now: good enough to be on the road, new enough to still surprise us.