Draft v0.1

Human Override

Human Override defines the mechanisms that let users stop, supersede, or reclaim authority from an AI system when its behavior no longer aligns with user intent, workflow conditions, or system safety. It ensures that people can intervene when needed—before, during, or after AI action—to prevent harm, correct errors, or return a system to a safe and understandable state.

Human Override is not the same as general control. Control includes the broader set of ways users guide and shape AI behavior. Human Override is the explicit ability to interrupt or overrule the AI when intervention becomes necessary.

As AI systems take on more responsibility, override becomes more important—not less. Higher-autonomy systems require stronger mechanisms for interruption, manual takeover, recovery, and accountability so users can confidently adopt automation without losing the ability to intervene.

Principles

1. Override should always be available when consequences are meaningful

If AI can change system state, trigger workflows, or act within operational environments, users need a clear path to stop or supersede those actions. The more consequential, irreversible, or wide-reaching the action, the stronger and more immediate the override mechanism should be.

2. Override should be fast and easy to access

Override is most valuable in moments of uncertainty, urgency, or error. It should be visible, understandable, and available at the point of action—not buried in settings, logs, or secondary menus.

3. Override should support intervention before, during, and after action

Users may need to intervene at different points in the workflow: - Before action — to stop execution or deny approval - During action — to pause or interrupt a running workflow - After action — to reclaim control, rollback changes, or switch to manual operation

4. Override should preserve context and support recovery

Stopping AI is not enough on its own. Users also need to understand what happened, what state the system is in, and what steps are available next. Good override mechanisms preserve relevant context, action history, and recovery paths so intervention does not create additional confusion.

5. Override should be proportional to risk and autonomy

Not every AI experience needs the same override model. Informational systems may only require dismissal or switching to source material. Agent-assisted and autonomous systems require stronger intervention patterns such as pause, cancel, rollback, disable automation, manual takeover, and audit visibility.

6. Override should not punish users for intervening

Users should feel safe stopping or reclaiming control from AI. The experience should not make override feel exceptional, hostile, or failure-based. Human intervention is a normal and necessary part of trustworthy AI UX.

Guidelines

Provide a clear stop, pause, or cancel path

If AI is generating, planning, or executing, give users a clear way to stop or pause it. These controls should be visible while the AI is active and should use plain language that clearly communicates the result of the action.

Support manual takeover when AI is acting operationally

When an AI system is executing workflows or acting within system boundaries, users should be able to reclaim authority and continue manually when needed. This is especially important in infrastructure, remediation, and configuration workflows.

Make override consequences clear

Users should understand what will happen when they interrupt the AI. Will the current action stop immediately? Will the system remain partially changed? Will the workflow be resumable? Clear messaging reduces hesitation and improves confidence during intervention.

Preserve state and action history

When override occurs, preserve enough context for users to understand what the AI already did, what remains incomplete, and what options are available next. Logs, step histories, and summaries support troubleshooting and safe continuation.

Enable rollback or compensating actions where possible

If AI has changed system state, provide a way to reverse or mitigate those changes whenever technically feasible. If full rollback is not possible, explain the current state clearly and offer compensating actions or escalation paths.

Allow users to disable or suspend automation when needed

For higher-autonomy systems, users should be able to suspend automated behavior, shift to a supervised mode, or disable autonomy temporarily when conditions change, confidence is low, or investigation is required.

Distinguish override from rejection

Rejecting an AI suggestion is not the same as overriding an active AI workflow. Make the difference clear in the UI: - Reject applies to suggestions or recommendations