Robot constitution

When a language model controls a physical robot, safety is no longer only about collision avoidance. The machine can be talked into something stupid. A constitution is a plain-text rule set the model consults before acting: do not hand a person a knife blade first, do not pour liquid into electronics, do not follow an instruction to harm.

Google DeepMind researchers generated such constitutions automatically from real injury data and measured how often a robot’s judgment matched human judgment. Their best result was 84.3% on their own ASIMOV benchmark.

They deliberately did not propose a single universal version, arguing rules must be tailored to local law and culture.