0.7.15 Modeling AI and Human Judgment With an Activity Diagram

Model AI Assistance with Human Review

A Wildcats coach asks an AI system to draft a training announcement. The system can propose wording, but the coach remains responsible for checking the time, location, audience, and suitability before using it.

An activity diagram can make those responsibilities visible. Put the AI's generation work in one partition and the human's evaluation and final decision in another. This is a workflow model, not evidence that an AI answer is accurate.

Keep Evaluation and Final Choice with the Human
Coach specifies the task, AI generates a draft, coach verifies it against the schedule and requirements, then approves, retries, or rejects it.
Review can lead to approval, revision, or rejection
  • New
  • Changed
  • Removed
Follow the Responsibility Changes

The prompt crosses from coach to system; the draft crosses back. Checking an announcement against the actual team schedule is human verification using a relevant source. Asking the same system whether its own time is correct is not equivalent evidence.

A suitable draft reaches approval. A defective draft first reaches a clear description of the problem, then a choice about another attempt. Rejection is a valid endpoint: the loop should not imply unlimited retries until the system is accepted.

Inspect One Failure