06
The public must always be able to switch off an AI system. The most powerful models need democratic approval before release and a shutdown mechanism tested by an independent authority, which can suspend them at any time. Executives who bring a system to market without these safeguards are personally liable for the harm it causes.