At 7:42 p.m. in San Francisco, a waymo taxi rolled away from the curb, paused behind a delivery van, and waited for a gap that never arrived. The rider was not in danger. The planner was simply conservative, and the app eventually offered a safe reroute. That small moment is more informative than a glossy demonstration because it exposes the real product: a constrained autonomous system operating inside a defined service area, not a human replacement with unlimited judgment.
I evaluate systems with the same basic loop every time: what happened, what should have happened, and how large was the gap? For a waymo taxi, that means separating the rider experience from the autonomy stack. Pickup accuracy, route choice, braking, yielding, remote assistance, and recovery from ambiguity all deserve separate scores. A vehicle can feel smooth while still making a poor decision at an unprotected turn.
What the ride actually involves
The customer workflow is deliberately ordinary. Open the Waymo app, enter a destination, review the estimated fare, request a vehicle, and follow the pickup instructions. When the car arrives, the app identifies it and explains how to unlock the doors and begin the trip. A rider still needs to choose a legal pickup point; the vehicle is not guaranteed to stop exactly where a human rideshare driver might improvise.
Inside, the most important difference is the absence of a driver to negotiate exceptions. There is no person to ask whether a blocked driveway is acceptable or whether you can stop briefly at a different corner. The cabin has controls for requesting assistance and ending or modifying a ride, but those controls do not turn the passenger into a supervisor. The operational design assumes the vehicle can handle normal traffic and escalate unusual cases through a remote support process.
That distinction matters for insurance and liability as well. A passenger should not treat the ride like a privately owned car. The operating company carries commercial coverage, while a rider's personal auto policy generally does not replace the operator's coverage for the autonomous trip. Exact terms depend on the service and jurisdiction, so read the current rider agreement before relying on assumptions about collision claims or lost property.
The autonomy is good at structure, not magic
The strongest part of a waymo taxi is repeatable behavior in a mapped operating domain. Lane geometry, traffic signals, curb space, and common pedestrian patterns give perception and prediction models useful context. The vehicle usually has more time than a distracted human driver to process its surroundings, and it does not become impatient after a long shift.
The weak point is the long tail. A construction worker may wave traffic through a closed lane. A police officer may override a signal. A moving truck may occupy the only drivable path. These events are not impossible for an autonomous system, but they create conflicts between map priors, visual evidence, and social cues. The planner has to decide whether the temporary instruction is trustworthy and whether proceeding creates a new hazard.
My severity scale treats a harmless hesitation as a 1, an inefficient but safe maneuver as a 2, a confusing interaction requiring remote help as a 3, a near-conflict or unsafe road positioning event as a 4, and an actual collision or serious safety failure as a 5. Most rider-visible oddities should not be placed in the same category as a dangerous event. A car taking an extra 20 seconds to yield is annoying. A car entering a cyclist's path is a different class of bug.
What to watch during a waymo taxi trip
Do not judge the system only by smooth acceleration. Watch the complete decision chain. At pickup, does the vehicle select a location that leaves room for traffic and pedestrians? At intersections, does it establish right of way before moving, or creep forward until another road user reacts? Around parked cars, does it provide a reasonable buffer? During a merge, does it communicate intent early enough for nearby drivers to understand it?
Also record recovery behavior. A blocked lane is not interesting because the car stops; it is interesting because of what happens next. Does the vehicle wait indefinitely, reverse unnecessarily, request help, or choose a legal alternative? A good autonomous system should fail in a predictable direction. Conservative behavior is valuable only when it eventually resolves the situation without creating a second problem.
For a useful personal log, record the date, city, weather, road type, traffic density, event description, and severity from 1 to 5. Do not record only memorable failures. Ten ordinary trips provide a baseline for pickup delays, unnecessary stops, and route deviations. Without that baseline, one strange interaction becomes an anecdote rather than evidence.
Pricing, availability, and the comparison problem
A waymo taxi fare is not a fixed national price. It depends on the city, trip distance, time, demand, and the service rules shown in the app. Compare the quoted total with a conventional rideshare only after matching pickup location, destination, wait time, and vehicle category. A cheaper quote that adds a long walk to a distant pickup point is not necessarily the better trip.
The more useful comparison is operational. A human rideshare driver can interpret a hand signal, move around a temporary closure, or answer a question immediately. An autonomous vehicle offers consistent attention and removes the variability of driver behavior, but it is bounded by its operating domain and escalation procedures. Neither advantage eliminates the other's failure modes.
Availability also changes faster than online reviews. A neighborhood can be inside a service area while particular streets remain excluded because of road geometry, construction, event traffic, or mapping limits. Check the live app rather than trusting a screenshot from a launch announcement. If the trip matters, leave enough time for a canceled pickup or a remote support delay.
A practical scorecard for riders
I would score a waymo taxi on five separate dimensions rather than one star rating. Safety margin asks whether the car maintained sensible space and speed. Intent communication asks whether its movements were legible to other road users. Progress measures whether it completed the trip without excessive hesitation. Recovery measures how it handled an unexpected obstruction. Rider control measures whether assistance, cancellation, and destination changes were understandable.
A sample result might read: safety margin 4 out of 5, intent communication 3, progress 4, recovery 2, rider control 4. That is a mostly successful ride with one meaningful weakness, not a declaration that the entire system is unsafe. Engineers need that resolution because aggregate satisfaction scores hide the exact behavior worth fixing.
The best test cases are repeatable: an unprotected left turn, a narrow street with parked cars, a four-way stop with poor lane markings, a curbside pickup near a bike lane, and a route interrupted by construction. Run each scenario more than once when possible. I have run supposedly identical evaluations three times and received three different answers because traffic participants changed. That variability is not noise to discard; it is part of the operating environment.
Bottom line: useful transportation, bounded autonomy
A waymo taxi is compelling when the task is a normal trip inside a supported area and the rider values a consistent, driverless experience. It is less compelling when the trip requires improvisation, a precise curbside interaction, or a quick answer from a human. The system should be judged by its boundaries, not by claims that autonomous driving is solved.
Your car talks. I check his homework. For this service, the homework is mostly solid on routine geometry and still incomplete on messy social coordination. Try it with a real destination, observe the edge cases, and log what happened. Then separate the harmless robot awkwardness from the events that deserve an engineering ticket. That is how a waymo taxi becomes test data instead of hype.