Build capability on both sides of the interaction.
Make room for human practice.
Intention
Choose a skill or judgment that matters, and establish starting capability.
Attempt
Work through a relevant task and explain important choices.
Feedback
Compare with a rubric, checked example, or expert response; keep evaluator limitations visible.
Practice
Try new cases that address observed difficulties.
Transfer
Test unfamiliar but relevant tasks under declared assistance conditions.
Retention
Revisit comparable new tasks after a suitable interval when lasting learning is claimed.
Better assisted output does not establish independent competence. Independent performance matters when that claim is made; it is not a universal demand for every task. Read source ↗
Name what changes in the AI system.
Prompt revision · Retrieval change · Inspected memory update · Tool change · Model selection · Workflow change · Model training.
Pin the exact mechanism and version, and preserve regression cases. Saved feedback is not evidence that model parameters changed. None of these mechanisms is performed by the current pilots.
Assess the working relationship.
Possible observations
Judgment, appropriate reliance, defect detection, review effort, ability to challenge the system, and participant feedback.
Limitations
These are possible measures, not a validated psychological instrument or a universal competence score. Missing evidence is not a poor result.
Participation is part of the method.
Agree purpose, permitted data use, access, withdrawal, and escalation before collecting individual observations.
Prefer aggregate reporting when individual records are unnecessary. Do not infer sensitive traits, rank employees from practice data, or reuse transcripts for training without distinct permission.
Restoring a configuration cannot undo human learning, habits, disclosures, or every external effect.