Autonomous driving is where AI claims meet physical consequences, and it is also where terminology is abused most freely. The SAE levels exist to make the discussion precise. Used properly, they cut through nearly every marketing claim in the category.
The six levels
The framework comes from SAE J3016, most recently revised in 2021. Two pieces of vocabulary do the work: the dynamic driving task (everything involved in operating the vehicle) and the operational design domain (the conditions under which a system is designed to function — roads, weather, geography, speed).
- Level 0 — No automation. The human does everything. Momentary interventions such as automatic emergency braking still count as Level 0, because they do not drive sustainedly.
- Level 1 — Driver assistance. Sustained automation of either steering or speed, not both. Conventional cruise control, or lane centring alone.
- Level 2 — Partial automation. Sustained control of both steering and speed, with the human continuously responsible for monitoring and for detecting and responding to everything the system misses. Every consumer driver-assistance system on sale today is Level 2, whatever it is called.
- Level 3 — Conditional automation. The system performs the entire driving task within its domain, and the human may genuinely disengage — but must be available to take over when asked. Widely considered the most problematic level, because handing control back to an inattentive human within seconds is an unsolved human-factors problem.
- Level 4 — High automation. The system performs the entire task including fallback within a defined domain. No human attention required; if something goes wrong the vehicle handles it. Commercial robotaxis are Level 4.
- Level 5 — Full automation. Level 4 with no domain restriction — anywhere a human could drive. Nothing is close.
The meaningful boundary is not a gentle slope from 1 to 5. It is the line between Level 2 and Level 3, because that is where legal and practical responsibility transfers from the human to the system. Everything below it is a driver aid; everything above it is a driver.
What is actually deployed
Genuinely driverless commercial service — no human in the vehicle, fares charged — exists and is scaling. As of late 2026 the clear leader operates in fourteen US metropolitan areas with a fleet in the low thousands and a paid ride volume in the hundreds of thousands per week. Several Chinese operators run comparable or larger volumes across more cities, and driverless commercial service has expanded into the Middle East.
A second tier operates driverless service in narrower footprints, including a purpose-built vehicle with no steering wheel carrying passengers under a federal exemption. A third tier still places a safety operator in every vehicle, which makes it a testing programme rather than autonomous service regardless of how it is described.
The case requiring care is the manufacturer whose consumer system is named to suggest full autonomy. That product remains explicitly supervised Level 2, with hands-on requirements and driver responsibility. Its robotaxi programme is a separate stack, running driverless in some metros and with safety drivers in others where state law requires it. Conflating the consumer feature with the robotaxi service — which the naming invites — is the single most common error in coverage of this sector.
Sensors: the central technical argument
Two philosophies, and the disagreement is real rather than manufactured.
Sensor fusion — cameras plus lidar plus radar — is used by every operator running driverless service at scale. Lidar measures distance directly rather than inferring it, and works in darkness and through airborne particulates. Radar penetrates weather and measures velocity directly. The argument is redundancy: different failure modes, so no single sensor’s blind spot is the system’s blind spot. The cost is money, plus dependence on high-definition maps and tightly geofenced domains.
Vision-only uses cameras alone, on the reasoning that humans drive with two eyes and a neural network, so cameras and a sufficiently good network should suffice. The advantages are cost and scalability — no mapping dependency, and a sensor suite costing a small fraction of a fused one.
The cost argument was once decisive and has weakened considerably: lidar unit prices have fallen by more than an order of magnitude as Chinese suppliers entered. Meanwhile the deployed evidence favours fusion — every operator running at scale without a human aboard uses it.
Modular versus end-to-end
A separate question, frequently conflated with the sensor debate. A modular stack splits the problem into perception, prediction and planning, each separately trained and connected by engineered interfaces — inspectable, debuggable, and able to carry explicit safety rules. An end-to-end system trains one network from sensor input to control output, which can learn nuances nobody specified; one such transition reportedly replaced hundreds of thousands of lines of hand-written logic.
End-to-end learning has a known weakness: a policy learned by imitation can compound errors once it reaches situations unlike its training data, drifting further from recoverable states with each step. The surveyed consensus favours hybrids — learned components inside a structure that retains inspectable interfaces and hard safety constraints. Note that these axes are independent: end-to-end learning is spreading into lidar-equipped stacks too.
Is it safer? The benchmark decides the answer
The leading operator publishes extensive safety analysis across hundreds of millions of driverless miles, reporting large reductions against a human benchmark for the same roads — on the order of eighty per cent fewer injury-causing crashes and over ninety per cent fewer serious-injury crashes. Parts of the methodology have been peer-reviewed.
An independent academic analysis reached a partially contradictory conclusion: still substantially fewer injury crashes overall, but a serious-injury rate higher than that of professional for-hire human drivers. The operator disputes the methodology.
Both can be defensible, because the disagreement is about the comparison, not the data. Measured against all human drivers — including the impaired, distracted and inexperienced — autonomous vehicles look strongly favourable. Measured against professional drivers, the margin narrows or reverses. Which benchmark is right depends on what the service displaces. That choice is the story, and any single safety statistic quoted without it is incomplete.
One metric to discount: disengagement rates. Self-reported, defined inconsistently between companies, and recorded during supervised testing rather than driverless operation. Useful for tracking a single programme over time, close to meaningless for comparing companies.
Regulation is the limiting factor
The US has no federal safety standard for driverless vehicles. Purpose-built vehicles without conventional controls operate under case-by-case exemptions, granted slowly and capped in volume. A voluntary federal oversight programme was withdrawn in 2026 for lack of participation, while crash-reporting requirements were narrowed but retained. Competing bills would either establish a federal framework or pre-empt state regulation.
In the absence of federal rules, authority sits with states — which is why the same company runs fully driverless in one state and requires a safety driver in another. For a technology whose economics depend on geographic scale, fifty regulatory regimes is the constraint that matters most, and it is not a technical one.
Related reading: AI agents explained on autonomy and failure modes in software, and AI benchmarks explained on why the choice of comparison baseline decides what a safety statistic means.
Last reviewed: September 2026. Deployment footprints and fleet sizes change monthly; verify current figures before relying on them.
