As Waymo prepares for a much‑trailed London debut, its vehicles on the streets of San Francisco are undergoing a striking change in driving style – one that underwriters and claims handlers will be watching closely.
For years the firm’s white Jaguar robotaxis were known for hanging back at junctions, giving way at four‑way stops and generally behaving like the most deferential motorists on the road. Residents complain that era is ending. The latest software releases appear to have shifted the balance from ultra‑cautious to hard‑edged, with cars pushing into gaps, accelerating more briskly from lights and, in some cases, straying over the line of the local traffic rules.
According to the Wall Street Journal, one San Francisco driver, who recently watched two Waymo cars weave through a narrow tunnel in close formation, described the new style as “driving more like a taxi driver - an aggressive, New York taxi driver.” Another regular passenger, who has clocked almost 3,000 minutes in the vehicles since 2024, now finds them far more willing to nip round a double‑parked car – close enough that she sometimes finds herself thinking: “Ooh that was really close.”
Recent incidents suggest that the system’s decision‑making is being tuned to cope with the realities of dense urban traffic rather than the letter of the Highway Code.
In San Bruno, south of San Francisco, police stopped a driverless Waymo after witnessing it execute an illegal U‑turn beneath a sign explicitly prohibiting the manoeuvre. Officers reported that when they pulled the car over, a remote operator spoke through the vehicle’s speaker system and said “they would look into it” and that “they were really, really apologetic.”
In another case in the city’s Mission district, a Waymo struck and killed a well‑known neighbourhood cat. Elsewhere, observers report vehicles at four‑way junctions accelerating in tandem with adjacent cars rather than patiently waiting their turn, and on at least one occasion dispensing with the courtesy of indicating a lane change.
Other residents say they have begun to see the cars performing so‑called “California rolling stops” – slowing at a stop sign but not coming to a complete halt when the road appears empty.
Waymo, for its part, insists that its software is still built to comply with the rules of the road, but concedes that there are trade‑offs. The company says it has been working to make its cars “confidently assertive,” arguing that such changes were “really necessary for us to actually scale this up in San Francisco, especially because of how busy it gets.”
Its head of product management for the driverless system points out that traffic rules can conflict: a driver may be told not to cross a double yellow line, but also not to block traffic behind. Human motorists faced with a stationary delivery van frequently edge over the line to get past; Waymo says its latest software now understands how to make the same kind of “common‑sense decisions.”
The behavioural shift is being introduced against the backdrop of a safety record the company is keen to promote. Waymo says its vehicles have driven 100 million miles without a safety driver in several US cities and reports 91% fewer crashes involving serious injury or worse compared with human drivers.
Those claims have been reinforced by an external analysis from Swiss Re, which reviewed liability claims across 25.3 million miles of Waymo driving. The reinsurer found that vehicles operated by the Waymo Driver generated 88% fewer property damage claims and 92% fewer bodily injury claims than human‑driven vehicles. Even against a benchmark of modern cars equipped with advanced driver assistance systems, Waymo’s fleet still recorded an 86% reduction in property damage claims and a 90% reduction in bodily injury claims.
The numbers are not without caveats. Waymo’s service is concentrated in cities, and its operating history remains relatively short. US regulators have already opened investigations into a series of crashes and traffic violations involving the vehicles, and the company has issued software recalls after minor but high‑profile collisions, including one in Phoenix in which a robotaxi hit a roadside telephone pole at low speed.
For insurers, the key question is how far the move towards “confidently assertive” behaviour can be reconciled with those headline claim reductions. A driverless system that is technically safer in aggregate but perceived as impatient or unpredictable by pedestrians and other motorists risks reputational blowback very quickly, particularly in the aftermath of a single fatal or widely shared incident.
The shift in San Francisco comes just as Waymo builds up its UK presence ahead of a proposed pilot robotaxi service in London. Job advertisements in the capital for roles such as “fleet readiness lead” and “incident response manager” suggest that the company is moving beyond a pure research footprint towards full operational readiness, covering roadside recovery, vehicle availability and post‑incident processes.
Under the Government’s emerging automated vehicle framework, London is expected to host limited public services from as early as the spring, with a national regime pencilled in for 2027. Passengers would be able to book rides with no safety driver behind the wheel, placing a sophisticated autonomous system into some of Europe’s most complex urban traffic conditions: narrow streets, informal road use, busy junctions and highly variable pedestrian behaviour.
For UK insurers and reinsurers, San Francisco’s experience offers an early glimpse of the underwriting challenges ahead:
The behaviour of the driverless system is no longer static. Over‑the‑air software updates can shift risk characteristics mid‑policy term, increasing assertiveness in the name of operational efficiency. That dynamic profile demands explicit change‑control clauses, obligations on the operator to disclose release notes, and clarity over how updates interact with warranties and conditions precedent.
Traditional rating factors – driver age, conviction history, no‑claims discounts – are of limited relevance. Instead, underwriters will need deep visibility on metrics such as disengagement rates, perception performance in rain and low light, incident frequency by software version, and the governance around rollbacks when bugs are discovered.
In the UK, third‑party motor claims are still likely to be presented initially to the motor insurer, even when a software fault is suspected. Recovery actions against vehicle manufacturers and autonomy providers will depend on authenticated telematics, high‑definition mapping data and a robust audit trail of the code running at the time of the incident.
If a system behaves more aggressively – for example, pulling into a gap that a human might have declined, or edging into a junction at the same time as another vehicle – attribution of fault becomes a more nuanced exercise. Claims handlers will need technical support to distinguish between an acceptable “confidently assertive” manoeuvre and a scenario where the algorithm has breached the duty of care expected of a professional operator.
A homogenous fleet relying on common software, sensors and cloud services introduces correlated failure modes. A defective release designed to make the cars less hesitant at junctions could, in an extreme case, trigger a cluster of similar collisions across a city in a short time window. That has implications not only for motor third‑party and property damage but also for business interruption covers written for fleet operators.
Reinsurers will therefore scrutinise event definitions, clash provisions between motor, product and cyber policies, and the contractual allocation of responsibility between the platform owner, the vehicle manufacturer and any contracted fleet partners. Waymo’s model, in which vehicles are likely to sit in centrally managed fleets rather than in private driveways, naturally points towards business‑to‑business placements with higher limits and sophisticated risk engineering requirements.
Perhaps the most immediate challenge, however, lies in public acceptance. Early iterations of robotaxis were criticised for holding up traffic and over‑reacting to uncertainty. There will be pressure from operators and city authorities alike to ensure that services are not seen as obstructions. Yet the episodes emerging from California – the illegal U‑turn, the rolling stops, the impatient pull‑aways at junctions – illustrate how quickly a perception of over‑caution can flip into concern that machines are being allowed to push the boundaries of the law.
For the UK insurance sector, that means preparing not just for novel claim patterns but also for the politics of automated mobility. Any collision involving a pedestrian or cyclist in London – particularly if linked to a conscious decision to make the system “more aggressive” – will be judged as harshly in the media as in the courts.
If government timelines hold and robotaxis do begin limited operations on London’s streets, underwriters will need to approach them as a live experiment: demanding data, insisting on clear contractual wording around software behaviour, and stress‑testing accumulation scenarios that reflect the possibility that a single line of code could turn a fleet of ultra‑polite chauffeurs into something closer to a pack of impatient cabbies.
