Preserve proportionate user agency over AI action
Users can preview, edit, confirm, cancel, stop, undo, retry, take over manually and escalate in proportion to impact.
実践ガイダンス
Place explicit approval before external or irreversible side effects. Summarize the actual action, target and consequence, then retain an audit record and reversible checkpoint where possible.
推奨
- Offer manual takeover
- Make cancellation immediate and visible
- Require scoped approval for external side effects
非推奨
- Bundle materially different actions into one consent
- Silently retry a side effect
- Remove manual paths after automation begins
検証契約
Cancel, undo, manual-takeover, duplicate-prevention and high-risk approval tests
根拠リファレンス
ref.anthropic.building-effective-agents— Building Effective Agents — Anthropic Engineeringref.apple.hig-machine-learning— Apple HIG — Machine Learningref.google.pair-guidebook— Google PAIR Guidebookref.intuit.design-system— Intuit Design Systemref.microsoft.fluent-ai-harm— Fluent 2 — AI Harmref.microsoft.human-ai-guidelines— Microsoft Guidelines for Human-AI Interactionref.salesforce.lightning-guidelines— Lightning Design System 2 patterns