How Self-Driving Taxis Work in San Francisco: Level 4, Not Level 5

Driving automation is graded on a five-level scale, and the vehicles carrying passengers in San Francisco sit at level 4, one step below the top. The distinction is not academic. A level 5 vehicle performs every driving task in every condition with no human involvement of any kind. A level 4 vehicle performs every driving task within a defined operating envelope, and outside that envelope it does not operate at all.

Two consequences follow. The first is geofencing: a virtual boundary, maintained by GPS, inside which the vehicle is permitted to run. The second is the remote operator, a human who can intervene from elsewhere. In practice that intervention is usually mundane, such as halting a vehicle whose passengers have not fastened seatbelts or who exceed the permitted occupancy.

A robotaxi builds its picture of the road from three sensor families that fail in different conditions, which is the point of having all three. LiDAR – light detection and ranging – measures distance by timing reflected light. Radar handles range and closing speed and is largely indifferent to weather. Cameras, some of them full 360-degree arrays, supply the detail that neither of the others can: colour, text, the state of a traffic signal.

The output is a rendered model of the surroundings, and the useful part of it is predictive. Detecting a cyclist is the easy half; estimating where that cyclist will be in three seconds and acting on the estimate is the half that determines whether the vehicle behaves acceptably in traffic.

The experience diverges from a conventional taxi in ways that have little to do with the driving. There is no negotiation about the route, no conversation to decline, and no moment at the end where a card comes out, because the fare settles in the application. Occupancy is enforced, not judged: a fifth passenger is not a matter for discussion.

The driving itself reads as cautious in a way that is legible from inside. The vehicle leaves larger gaps than a human would, hesitates at ambiguous intersections, and yields in situations where a person would assert. Regular passengers describe the same two impressions in roughly the same order: unsettling for the first few minutes, then unremarkable, then slightly slow.

The failure modes are also different in kind. A human driver misjudges; an autonomous vehicle encounters a situation outside its model and stops. Stopping is the safe response and also the disruptive one, which is why the incidents that draw attention tend to be vehicles halted in inconvenient places, not collisions.

“Driverless” describes the front seat, not the operation. Level 4 deployment assumes a support layer: staff who can inspect a stalled vehicle’s surroundings, advise it, halt it, or dispatch someone to it. The ratio of humans to vehicles is not one to one, which is the entire economic argument, but it is not zero either.

This matters for how the service should be read. A vehicle that hands an unsolvable situation to a remote human is not failing; handing off is the design. What the design does mean is that availability depends on a support organisation as much as on the software, and that an incident is resolved at the speed a person can reach it, not at the speed a driver already in the seat could.

It also complicates the question of who is responsible when something goes wrong, in a way that a metered taxi does not. A cab has a driver, a licence, a vehicle number and a dispatch agency, and the receipt names enough of them to start from. An autonomous trip has an operator and a software version.

Two operators were cleared by California regulators to carry paying passengers without a driver in San Francisco: Waymo, which grew out of Google and sits under Alphabet, and Cruise, majority-owned by General Motors. Both had spent years testing with safety drivers before the permits arrived, and the California Public Utilities Commission cleared paid driverless service for both in August 2023 over objections from the city’s own transport and fire departments.

The two diverged within months. In October 2023 a Cruise vehicle struck a pedestrian who had already been thrown into its path by another car, then pulled over with the pedestrian beneath it and dragged her. California suspended Cruise’s driverless permit days later, and General Motors subsequently ended its funding of the robotaxi programme. Waymo continued, and is the operator a passenger in San Francisco meets today.

That sequence is the clearest thing this page can offer about risk: the record that matters is not the collision rate in a brochure, it is what a regulator did when one vehicle behaved badly.

The deepest difference between a robotaxi and a licensed cab is not the absence of a driver. It is who decides the fare. A Seattle taxi charges what a jointly issued city and county rule says it charges, and every licensed vehicle charges the same, which is why a passenger there never has to compare. A robotaxi fare is set by the operator and moves with demand, in the manner of any app-dispatched service, where a Seattle meter reads the same in every licensed vehicle.

Both arrangements have a defensible logic, and the rule-making behind the older one is public. A fixed public tariff protects a passenger who cannot shop around, which is exactly the position of someone stepping off a plane. A dynamic price allocates a limited fleet toward whoever needs it most at that moment. What they do not share is comparability: two metered cabs quote the same number, and two applications do not.

It is worth noting how recent the settled version of the older model is. In Seattle the separate for-hire vehicle category, with its rate books and hourly charges, only ended on 31 March 2026. The regulated-tariff model is not a relic that autonomous vehicles are displacing; it is still being consolidated.

A level 4 vehicle is defined as much by where it will not go as by where it will. The operating design domain is the set of conditions the system is certified for, and it is narrower than a map suggests: it covers geography, but also weather, time of day, road types and situations the software has been validated against.

Three categories account for most of the exclusions. Weather that degrades the sensors, heavy rain and fog in particular, because each sensor family fails differently and the redundancy that makes the system safe is exactly what the weather removes. Unstructured environments, such as active construction zones where lane markings and signals have been replaced by a person waving. And the long tail of situations that occur rarely enough that no fleet accumulates the miles to validate against them.

That last category is the one that keeps the gap between level 4 and level 5 open. Extending the envelope is not a matter of better software in general; it is a matter of validating a specific new condition, one at a time, which is why the boundary has crept outward instead of jumping.

Commercial autonomous passenger service opened in San Francisco under a permit regime, and the permits are granular. Whether a vehicle may carry paying passengers, whether it may operate without a safety driver, which hours it may run and which streets it may use have all been governed separately and adjusted repeatedly.

That granularity is why blanket statements about the city age badly. The set of vehicles allowed to charge fares, in a given area, at a given hour, has changed more than once, and a description written at any single point describes that point and not the regime.

The comparison with a metered taxi is less about technology than about who sets the price. A Seattle cab charges what an ordinance says it charges, identically across every licensed vehicle. A robotaxi fare is set by the operator and moves with demand, which is a different arrangement wearing the same word.

Are San Francisco robotaxis fully self-driving?

They are level 4, not level 5. That means they drive themselves within a defined operating envelope, with geofencing and a remote operator who can intervene. Outside that envelope they do not operate at all.

Is there really nobody supervising?

Nobody in the front seat. There is a support layer of remote staff who can inspect a stalled vehicle’s surroundings, advise it, halt it or dispatch someone to it. The ratio is not one to one, but it is not zero.

What stops them operating in bad weather?

Heavy rain and fog degrade the sensors, and the redundancy between LiDAR, radar and cameras is exactly what the weather removes. Construction zones are the other common exclusion, because a person waving replaces the markings the system reads.

How does the fare compare with a metered taxi?

Differently in kind. A licensed Seattle taxi charges what an ordinance says, identically across every vehicle. A robotaxi fare is set by the operator and moves with demand, so two applications need not quote the same number for the same trip.