It is tempting to ask whether an AI-powered car can eventually drive better than a human. The harder question is how ‘better’ should be measured. Avoiding collisions, responding to rare hazards, handling bad weather and behaving predictably around pedestrians all matter. A system may perform well on one road type and struggle in another, so broad claims need broad evidence.
Define what ‘better than humans’ means
The source material distinguishes Level 2 driver assistance, including Tesla Autopilot or Full Self-Driving (Supervised) and BYD DiPilot, from Waymo’s Level 4 driverless service in defined operating areas. These systems do not place the same responsibility on the human. Under Level 2, the driver must supervise; a Level 4 service can operate without a human driver within its approved operating domain.
Automation levels change the responsibility
The source record includes secondary figures for reported deaths involving driver-assistance systems and an unverified count associated with FSD. Those numbers should not be treated as established safety comparisons without checking the underlying NHTSA records, definitions and reporting periods. Company claims about lower crash rates also need comparable exposure, geography, road conditions and severity measures.
Safety statistics need careful comparison
Regulators, insurers and drivers need transparent data: miles travelled, system engagement, disengagements, crash severity, weather, road type and whether the system was supervised. Comparing a robotaxi fleet operating in a mapped area with all human driving is not a fair test unless the differences are accounted for.
What regulators and insurers need to see
AI may ultimately outperform people in some defined driving environments, but the evidence must be specific to the operating conditions. Until then, drivers should follow the vehicle’s instructions and never assume that an assistance feature makes the car self-driving.
The evidence is still developing
Until stronger evidence is available, drivers should treat assistance systems as aids rather than substitutes for attention. Insurers and fleet operators should examine the vehicle’s automation level and the conditions in which the feature was active when reviewing an incident. Regulators need comparable reporting standards. A fair assessment should recognise both the potential of automation and the limits of the available data.