What should machines be allowed to do alone?

What should machines be allowed to do alone?

Physical autonomy should be granted inside a defined operating envelope, not as a universal level of intelligence. The same machine may work safely alone in a fenced, mapped facility and require close supervision near people, unstable terrain or unfamiliar objects. Authority must follow conditions.

The simulation�reality gap remains decisive. Stanford�s 2026 AI Index reports 89.4% robotic-manipulation success in a software benchmark but only 12% across real household tasks. Autonomous vehicles reached scale�about 450,000 weekly Waymo trips across five US cities in 2025�yet deployments still operated in bounded geographies and generally favourable conditions with remote human support available.

Define the operational design domain through environment, location, weather, speed, payload, human proximity, connectivity and foreseeable misuse. Combine it with the energy a machine can release, the reversibility of harm and whether failure is detectable before contact. Tasks outside that envelope should trigger a safe stop or controlled handover, not improvised confidence.

Layer learned perception and planning inside deterministic safety controls: geofencing, speed and force limits, separation monitoring, redundant sensing, health checks and an independently tested emergency stop. Give operators an accurate view of state and enough time to intervene; a nominal human-in-the-loop is not protection when attention or latency makes takeover unrealistic.

Validate rare events through simulation, controlled trials and field evidence, then monitor near misses, boundary excursions, disengagements and recovery time. Expand the envelope in small, reversible increments. A machine has earned autonomy when the complete system can recognise the edge of its competence and fail without exporting unacceptable risk to people around it.