Definitions
Self-Improvement. An AI system modifying its own code, weights, training, or successors so that each gain speeds the next. Humans using AI to build better AI is engineering; the line is crossed when the system directs the loop. The danger is a fast takeoff: capability gains arriving faster than humans can evaluate them, with the next improvement beginning before the last has been reviewed.
Human control. The ability to inspect, pause, and halt at each step. Pressing “go” is not control.
Independent scientific review. Review by qualified people outside the developer, whose findings cannot be suppressed. Where publishing a method is itself dangerous, the findings are published, not the method.
Individual accountability. Named persons answer for each decision to proceed: the CEO, the other decision makers, and the people who carry it out.
Please consider adding yourself as a signatory.