Who pays when a robot injures someone?

Four entities sit in the liability chain when a robot injures someone. The AI vendor who built the model. The OEM who built the machine. The integrator who installed it. The operator who used it. Each has a different relationship to the harm — and each has been tested in court.

The reason this matters now is that Physical AI blurs boundaries that traditional product liability assumed were clear. In a conventional machine, the manufacturer designed the behaviour and the operator followed procedures. Fault was binary: either the machine was defective or the operator was negligent.

With AI-controlled machines, the picture fractures. The AI vendor provided a model that was trained on data they selected, with objectives they defined. The OEM integrated that model into a physical system with actuators that can exert real force on real bodies. The integrator configured the safety parameters at the deployment site. The operator followed — or didn't follow — the procedures written for a system whose behaviour might change with each software update.

Each of these relationships creates a distinct liability arrow. And each arrow has been tested. Tesla faces lawsuits as both AI vendor and OEM for Autopilot-related crashes. Industrial robot manufacturers in South Korea and the US have faced product liability claims after fatal workplace incidents. Integrators have been held liable when commissioning errors led to injuries. And operators — or their employers — bear responsibility when lockout/tagout procedures are violated.

The question compliance teams, insurers, and founders should be asking is not "will we be liable?" but "which arrow points at us — and can we prove where our responsibility ends?" That's where a hardware-enforced safety boundary becomes more than a safety mechanism. It becomes a legal boundary: an auditable, independent layer that records whether the physical outcome was permitted or blocked, regardless of what the AI decided.

The liability chain in Physical AI is not hypothetical. It's being drawn right now, case by case, courtroom by courtroom. The companies that map their own chain before an incident happens will be in a fundamentally different position than those who discover it after.