How to build trust in physical AI before regulation kills it
The market can't afford to wait for regulation. Builders need to adopt a common safety standard that defines acceptable failure before courts and regulators define it for them.
AAA's most recent surveys put fear of self-driving cars at 68% of US adults, with active trust at 9%. These figures have not improved meaningfully across two years of expanded operation.
In October 2016, Zipline turned on the world's first national drone delivery service over Rwanda. Time-to-blood at participating clinics fell from four hours to fifteen minutes. But most people heard about it years later, because physical AI needs time for validation. By the time a deployment reaches the headlines, it has usually been running in dev environments for months or years to generate the trust it needs.
That gap between deployment and headlines is also a gap between deployment and external validation: operators are running ahead of both regulators and the court of public opinion.
Each new deployment creates a new trust question
From logistics networks and public roads to factory floors and private homes, every new deployment creates a failure mode outside the existing legal framework. What happens if a drone misjudges wind shear? If an AV's perception model fails on an edge case? If a humanoid drops a 90 kg part? The chain of responsibility runs through the operator, the manufacturer, the AV-stack or robotics provider, and the foundation-model vendor. We always look for someone to blame for a "preventable" accident.
California has begun spelling this out. Attorneys Clark H. Fielding and Ryan W. Cooper at Fielding Law describe the current regime: autonomous ride-share operators must carry $5M of primary coverage, and in a Waymo crash the liable party can be the fleet operator, the manufacturer, the software developer, or any combination of the three (Fielding Law). The single accountable human, or cab company, is replaced by all the builders of its successor, and the door for litigation payouts higher than conventional road accidents is wide open.
Setting the bar before the first incident
In October 2023, a Cruise robotaxi struck a pedestrian in San Francisco after she had already been hit by another vehicle. The AV then attempted to pull over, dragging her for several metres. The incident itself was serious, but the larger fallout came from how it was handled: regulators said Cruise had failed to properly disclose key details, California suspended its permits, GM halted Cruise operations nationwide, and the company's rollout collapsed almost overnight.
What happened to Cruise was not unique. Autonomous fleets have crashes. What was unique was the absence of a shared, pre-established expectation of what counts as an acceptable failure, an unacceptable one, and what each demands in response. It also didn't help that Cruise directly lied to regulators. Without proper expectation-setting, the single incident became the whole AV story.
Aviation, also once a budding new technology, now works the opposite way. A commercial aircraft crash is a recoverable event for the industry. Fleet groundings happen, NTSB investigations ensue, mandated remediation is followed through, and public confidence remains intact because operators, regulators, and the flying public share a baseline for what to expect. We know flying is inherently dangerous, and we know that mistakes can happen. The bet of flying is worth the risk.
The FAA's Safety Management System framework, formalised across US commercial aviation by 2018, is built explicitly around proactive identification of acceptable risk levels rather than reaction to individual events. When the rate is exceeded, the system responds. When it is not, we investigate, but operations continue. This system has led to a generally accepted notion that we control the risk of crashes as best we can, but that there's only so much that can be done.
Robotics does not yet have that. The mismatch between demanded and demonstrated performance is, on its own, not a safety problem. It is a fear problem.
AAA's most recent surveys put fear of self-driving cars at 68% of US adults, with active trust at 9%. These figures have not improved meaningfully across two years of expanded operation. Fear has a track record of killing developing technologies long before their actual safety records warrant it.
The way out, for AVs, for humanoids, for delivery drones, is to set the bar in advance. Operators, underwriters, and standards bodies need to agree, while deployments are still small and incidents rare, on what an expected failure rate looks like, how it is reported, and what response a deviation triggers.
Nidhi Kalra and Susan Paddock at RAND showed in 2016 that AVs cannot drive their way to statistical safety proof: the mileage required to demonstrate safety against human-driver baselines runs into the hundreds of millions to hundreds of billions of miles (RAND). Pre-set expectations are a practical requirement.
Done well, correct expectations can make a Cruise-like incident a recoverable event for the industry instead of an extinction event for the company that had it.
Where the framework is actually being written
The healthier alternative to waiting for legislation is not the absence of standards. It is standards set early, by the parties pricing the risk (i.e., insurers), and enforced through underwriting, contracts, and claim settlement.
In some types of physical AI that is already happening.
Matthew Carrier and Dishank Jain at Deloitte size the migration: a 20% transfer of premium dollars from commercial auto into autonomous-fleet coverage represents roughly $7B/year in shifted premium, plus another ~$3B in long-haul workers' comp (Deloitte Insights).
The historical loss data underwriters built their models on does not yet exist for these systems. The first carrier with operating data sets the price.
The technology side is converging on the same view. At Fortune Brainstorm AI in December 2025, Rene Haas of Arm told the audience that "in the next five years, you're going to see large sections of factory work replaced by robots, and part of the reason for that is that these physical AI robots can be reprogrammed into different tasks" (Fortune; TechSpot). Reprogrammability, one platform, many tasks, is what makes a humanoid an underwriting unit rather than a single-use tool. It is also what creates an actuarial problem statute cannot resolve in time. This is why market-based regulation is both a prerequisite for and a result of the nature of physical AI deployment.
This is the healthier outcome. Standards set by the parties pricing the risk move at the speed of the deployment. They update with each claim and each new model. They reflect the operating data of the systems they cover.
The bet most operators are making
The operators who help set the standards before a failure are the ones who will shape the next decade's terms.
Most physical AI operators today are betting they have more time than they do. The conversation tends to run the same way: we'll get coverage when we have to; the regulator will eventually write rules and we'll comply when they do. None of that is unreasonable. It's just not strategy.
Three things will catch up:
- Claim events. Every deployment in the last decade has worked, until one didn't. Cruise was the most public failure, not the only one. When the next incident happens to a smaller operator without a dedicated risk function or established carrier relationships, the operator is the deep pocket, and "who pays" gets answered the slow and expensive way: in court.
- Exclusions. Many existing homeowners and general-liability policies already contain "motorised vehicle" or "aircraft" exclusions that carriers are starting to apply to mobile robots. Operators of physical AI today are frequently carrying less coverage than they think, and discovering it after a claim, not before.
- Regulation. The EU AI Liability Directive's withdrawal is not the end of the legislative cycle, only its current pause. When the statute does arrive, it will arrive prescriptively, and the operators it lands on will be the ones whose incidents made the news first. The more self-regulation there is, the less harsh state regulation has to be, so the earlier the industry adopts a workable framework, the better it is for everyone.
The industry has to engage with a framework either way. The choice is whether to engage with it on the public or the private level. If you ask most engineers building robotics today, I think we both know which they'll prefer.