Where things really stand
Fully driverless cars are not everywhere. A small number of automated taxi and delivery services run in selected areas of some cities, typically in mapped zones, with conditions the system is built to handle. Most cars sold to ordinary buyers still need an attentive human driver at all times, even when they offer advanced assistance. Marketing names sometimes blur this line, which is why understanding the basics matters.
The levels of driving automation
A widely used framework from the engineering body SAE International describes six levels, numbered 0 to 5.
| Level | Name | Who is responsible |
|---|---|---|
| 0 | No automation | The human does all driving, though warnings may be provided |
| 1 | Driver assistance | The car helps with steering or speed, but not both together; the human drives |
| 2 | Partial automation | The car handles steering and speed together; the human must supervise constantly |
| 3 | Conditional automation | The car drives in limited conditions; the human must take over when asked |
| 4 | High automation | The car drives itself within a defined area or conditions, without needing a human to take over |
| 5 | Full automation | The car drives itself everywhere a human could |
The key divide is between levels 2 and 3. At level 2 and below, the person is driving, even if hands are briefly off the wheel, and is legally and practically responsible. From level 3 upward, the system is doing the driving part of the time. Level 3 is hard because handing control back safely to a person who has stopped paying attention is difficult. Level 4 systems are the ones found in current automated taxis, while Level 5 does not exist in public use.
Terms such as “self-driving” and “autopilot” are not formal levels, and the wording used in advertising does not always match what a system can do. Checking the owner’s manual for the system’s level and conditions is more informative than trusting a product name.
How an automated vehicle perceives the world
Automated systems combine several kinds of sensors:
- Cameras, which read lane markings, signs, lights and objects, using image recognition based on neural networks.
- Radar, which measures distance and speed and works reasonably well in rain or fog.
- Lidar, which uses laser pulses to build a three-dimensional map of surroundings. Some developers rely on it; others aim to manage with cameras alone, which is a continuing engineering debate.
- GPS and detailed maps, which help the vehicle know where it is.
- Ultrasonic sensors, mainly for close-range parking.
Software then combines these inputs, tries to work out what each object is and where it is likely to go, plans a path and controls steering, braking and acceleration. Much of this relies on machine learning, trained on large amounts of driving data. As in other systems, what the software has seen in training shapes what it handles well, which is the same reason training data matters for any AI system.
What makes safety hard to establish
Humans are not perfect drivers. Crashes are often caused by distraction, tiredness, impairment or speeding, and automation could in principle reduce those. Whether it actually does is a question of evidence, and the evidence has several complications.
Rare events. Serious crashes are rare per mile, so very large amounts of driving are needed to compare accurately. A system that looks safe over a modest distance might still have flaws that appear only occasionally.
Different conditions. Automated services mostly run in favourable places and weather. Comparing their crash rates with those of all human drivers, including on rural roads and in snow, is not like for like.
Edge cases. Unusual events, such as a pedestrian partly hidden behind a vehicle, unclear road works or unexpected behaviour by other drivers, are where automation can fail. Because there are so many possible odd situations, testing cannot cover them all.
Human factors. With partial automation, drivers may over-trust the system and look away, even though they are meant to supervise. This is a recognised risk, and some systems use cameras to monitor driver attention.
Data and transparency. Companies report their own figures, and the way incidents are defined and counted varies. Independent studies, regulators and incident investigators provide additional views, but publicly available data are still limited.
It is accurate to say that there have been serious and fatal crashes involving both partially and highly automated vehicles, and investigations have led to software changes and, at times, suspended services. It is also accurate that some operators report fewer injury-causing collisions than human benchmarks in their service areas. Both statements can be true, and neither settles the question for driving in general.
Cybersecurity and privacy
A connected vehicle can be a target for hackers, so developers must protect software and receive timely updates. The wider topic of cybersecurity threats applies here too. Vehicles also collect data about routes and surroundings, which raises privacy questions that depend on local rules and company policies.
Rules and responsibility
Regulation differs across countries and, in some places, between regions. Questions include which tests a system must pass before use, who is liable after a crash, how incidents are reported, and what happens when a vehicle blocks emergency services. These rules continue to develop, and they are a major reason automation arrives gradually.
What to take from this
For someone buying or riding in a car with automation, the practical points are modest. Learn which level of system it is. Understand when you are expected to be in control. Do not treat driver assistance as a replacement for attention. For fully driverless services, remember they operate within limits that the operator sets.
The technology is real and improving, but the claim that it is already everywhere is not accurate. Automation is only one part of a broader shift in how people and goods move.