Why General-Public Deployment Is a Different Safety Problem
A warehouse robot may safely operate around people without being ready to operate around the general public.
That is not contradictory. It is a distinction that should guide deployment policy.
Industrial environments provide controls that public environments do not.
People in a warehouse, factory, hospital unit, distribution center, or controlled commercial facility may be:
Known to the organization
Trained on procedures
Informed that automated systems are present
Subject to safety requirements
Supported by supervisors
Located in geographically bounded areas
Able to access emergency intervention
The environment itself may also be characterized. Traffic patterns, lighting, floor conditions, workflows, access points, hazards, and operating procedures can be mapped, monitored, and managed.
That makes industrial and commercial deployments valuable. They can serve as an extension of testing by exposing machines to more users and more operating conditions than a laboratory alone can provide.
But general-public deployment is different.
The General Public Is Unselected
In public settings, the humans around a machine may be:
Unselected
Untrained
Uninstructed
Unaware of the machine’s presence
Unfamiliar with its capabilities
Unaware of its limits
Unpredictable in their movement and behavior
Unable to access immediate human intervention
They may be children, older adults, tourists, delivery workers, people using mobility devices, people in distress, people navigating crowded spaces, or people who simply do not notice the robot.
They may interact with the system in ways that developers never anticipated.
That creates the central deployment question:
Can the system safely interact with humans who have not been selected, trained, modeled, or instructed to interact with it?
A “yes” should require evidence—not optimism, marketing demonstrations, or the assumption that the public will adapt.
A Deployment Ladder
Rather than treating deployment as a binary choice between laboratory testing and full public release, embodied AI should move through a structured ladder.
Level | Environment | Human population | Primary purpose |
|---|---|---|---|
Level 0: Laboratory | Controlled research setting | Selected participants | Identify early failure modes |
Level 1: Industrial | Bounded workplace or operational site | Known, trained personnel | Demonstrate safety in a defined operating domain |
Level 2: Commercial | Managed business environment | Broader users with some controls | Test generalization beyond trained operators |
Level 3: General Public | Open, dynamic public space | Unselected and untrained people | Demonstrate safe human-environment integration |
Level 3 is not simply Level 2 with more people.
It is a different class of safety problem.
At Level 3, the system must account for a much wider range of human behavior, environmental complexity, cultural norms, distractions, accessibility needs, and unexpected events. The machine cannot assume that people will know what it is, how it moves, what it can perceive, or how they are expected to respond.
The Public Should Not Become the Training Set
There is a subtle but important risk when immature autonomous systems enter public spaces.
Society may begin adapting itself to the limitations of the machine.
People may be told:
Do not approach it this way.
Do not move unexpectedly.
Do not interfere.
Do not stand in that area.
Follow the machine’s instructions.
Do not behave in ways the system cannot interpret.
In a tightly controlled industrial setting, some of those rules may be reasonable. Workers already follow procedures around forklifts, heavy machinery, chemical systems, and other hazards.
But public space is different.
Something has gone wrong if ordinary people must learn to behave like the robot’s training distribution in order for the robot to function safely.
The relationship should be reversed.
The robot should accommodate normal human variability rather than require humans to become predictable enough for the robot.
Humans do have responsibilities. No one should intentionally sabotage a system, ignore clear safety barriers, or place themselves in obvious danger.
But the central safety burden should not be shifted onto the public because a machine cannot reliably interpret ordinary human behavior.
The public was here first.
The Standard for Public Integration
The question is not whether systems should ever operate in public.
They can, and many already do in limited forms: delivery robots, autonomous mobility pilots, service robotics, public-facing kiosks, and semi-autonomous equipment.
The real question is whether the deployment matches the evidence.
A system that has demonstrated safety only in controlled environments should remain within a controlled operating envelope.
A system that has demonstrated reliable performance under supervision may be ready for supervised public pilots.
A system that has not demonstrated safe behavior across diverse people, environments, failures, and high-consequence situations should not be granted uncontrolled authority in public spaces.
That principle becomes even more important as AI systems become more capable and more autonomous.