A burglar in San Francisco used a Waymo robotaxi as a getaway vehicle and walked free, partly because the footage captured by the car's cameras sits in a legal and procedural gray zone. The incident is being reported as a quirky crime story, but it is actually a precise case study in what happens when an autonomous system is designed to observe everything and is accountable to nothing in particular.
Seeing Everything, Proving Nothing
Waymo's robotaxis run continuously on sensor arrays that would make a surveillance state blush. Cameras, LiDAR, radar. And yet the footage did not straightforwardly produce an arrest. The reason is structural, not technical: there is no pre-established chain of custody, no agreed evidentiary framework, no societal consensus on what the machine is a witness to. A new arXiv paper by Luong Tuan and Sanyal titled "Toward Pre-Deployment Assurance for Enterprise AI Agents" identifies exactly this gap. It argues that pre-deployment verification of AI agents remains a critical failure point, that we deploy first and discover accountability voids second. The yoga clothes are gone. The robot watched the whole thing. Nobody was certified to ask it what it saw.
Trust Certification as the Missing Layer
The same logic haunts the broader AI deployment wave. A Delphi study of 272 international AI risk experts published this week on arXiv ranked accountability gaps as among the highest-priority near-term AI risks, not rogue superintelligences, but mundane deployments that generate data no one has legal standing to use. The Waymo incident is almost too clean as an illustration. The system performed exactly as designed. The criminal benefited. The accountability infrastructure simply was not there. As AI systems proliferate into homes via robots like Hello Robot's fourth-generation Stretch, and into grid infrastructure via repurposed EV batteries, this is not a niche problem. It is the substrate condition of the next decade.