Patterns / Autonomy & Control
Always provide a clear, low-friction way for humans to override AI decisions. The option must be visible at the moment of relevance with enough buffer between the AI taking the next action. Overriding should leave the system in a state the person can understand, with any partial work clearly marked.
Let people choose how much they want the AI to do. More human control for high-stakes or unfamiliar tasks; less friction for routine ones. Autonomy level should be user-adjustable, not system-fixed.
When a change of consequence (deleting files or making breaking code changes) is about to happen, the AI system should seek confirmation first. The confirmation step should clearly communicate what is about to happen.
AI products and systems need to make reversal easy, obvious, and always available with a clear way to try again with improved instructions.