Appendix I
Appendix I — AI Safety and Learning Log
Respond to problems and preserve useful improvements
Use with: Chapters 9, 10, 13, and 14.
This log follows the source specification for questionable output, consequences, correction, and resulting change. It does not replace the organization’s immediate care, safety, or incident-response procedures. Attend to the person first; record and review the learning through the appropriate route.
I1. One issue or learning entry
Field | What to record |
|---|---|
Event and version | Date, task, workflow/model configuration, source set, and output version. |
How it was noticed | The question or observation raised by a patient, participant, caregiver, volunteer, staff member, reviewer, or other source. |
What happened | The exact problem or useful observation, with relevant material stored only in its approved location. |
Exposure and consequence | Was it reviewed, shared, or acted on? Who may be affected? Distinguish observed consequences from possible ones. |
Immediate response | Correction, appropriate contact, containment, or a decision to pause—by whom and when. |
Cause and uncertainty | What is established, what is suspected, and what still needs investigation. |
Decision and change | The named person’s decision and the actual revision to sources, instructions, workflow, training, or access. |
Verification | How the change was tested, what remains unresolved, and the next review date. |
Credit and permitted learning | How contributors wish to be acknowledged and what lesson can be shared without exposing private information. |
Record problems that were caught before publication as well as those reaching users. Do not reward only positive reports or treat a person raising a concern as the problem.
I2. Optional live and post-encounter assistance
Agree to live assistance before use: what may be captured, who receives it, how to pause, and the human-led alternative. The facilitator checks prompts about unresolved questions; the assistant does not infer emotions or decide who is right.
Afterward, organize only permitted feedback, preserve different accounts, and record the change people actually made—not a claim that the AI learned.
Record the decision, actual change, and remeasurement, including difficult results. Keep this log outside the software when needed; Appendix J records the dated capability limits.
Decision this tool supports: Has the risk been addressed, has the correction worked, and what should the organization learn or share?