Season · AI, Trust & Human Oversight · 12 of 16
An AI summary can look reviewed when someone has only agreed to review it.
Concept of the Week · Accepted is not checked
When handing over an AI-generated summary, preserve what has and has not been checked. A reviewer’s name indicates who has taken the task; a review finding records what they compared and what they found. Confusing the two can make an unchecked statement appear ready to use.
Handoff rule: Keep the statement visibly unchecked when review responsibility changes; record task acceptance and review completion separately.
In an intensive care unit (ICU), an outgoing physician prepares to hand over overnight changes so the incoming physician can assess the patient’s current condition. An AI summary assistant drafts those changes from clinical notes, observations, and results, but the incoming physician still needs to verify a statement about overnight blood pressure against the source observations.
The incoming physician agrees to accept the check, and the interface records their name. If the summary now receives a general “reviewed” marker, accepting the task can make it appear that the comparison is complete.
The incoming physician may need to reopen the notes and reconstruct which checks occurred, or use the unchecked statement in their assessment without recognizing the gap.
A calmer design could keep the statement, the source observations, and the open check visible after the incoming physician accepts the task. After comparing the statement with the source observations, the physician records whether the statement is supported, needs correction, or remains unresolved, without marking the rest of the summary as reviewed. If the intended reviewer cannot take the task, it remains open, with the current reviewer visible, while clinicians arrange cover through their usual escalation route.
Separating task acceptance from the recorded source comparison could help clinicians identify which statements still need checking without reconstructing a colleague’s work.
Ship this by developing the handover with ICU physicians, testing whether they can distinguish accepted tasks from completed checks during interruptions and reviewer changes, then revising and retesting against agreed usability and safety criteria before release.
For ICU physicians handing over an AI summary, the left view indicates the review is complete; the right keeps the unchecked statement and its sources visible after task acceptance.
Asset & Resource
Design and human factors teams can use the Review Handover Map with clinical colleagues to trace a single unchecked AI-generated statement throughout a handover. Revisit it when the summary, source access, or review workflow changes.
Keep the statement and source together: Identify the exact statement and the observations or notes needed to verify it.
Show what remains unchecked: Keep the open comparison visible, even after someone accepts the task.
Confirm who has accepted the check: Show the current reviewer until the next reviewer explicitly accepts it, with room for questions.
Record what was found: Capture the comparison, any corrections, and unresolved points without approving unrelated statements.
Keep exceptions open: Show what happens when the reviewer is interrupted or unavailable, and how another person takes over.
Use the Review Handover Map alongside the actual interface and the local handover procedure. Ask representative clinicians to identify what has been checked, what remains open, and what they would do next. Use misunderstandings and repeated searches to revise and re-evaluate the design.
The article Meaningful oversight of medical AI beyond human in the loop argues that assigning a clinician is not enough: they also need knowledge, time, authority, and practical ways to act. The Review Handover Map applies that broader principle to one proposed handover; the article does not establish that this interface works.
Light Wisdom & Reflection
“Communication is a process of sharing experience till it becomes a common possession.”
Where could a handover make an AI-generated statement look more thoroughly checked than it is?
Mindful Practice
When someone at home offers to water a plant while you are away, pause beside it for a moment.
What do you know now, and what have you assumed?
What would help you leave room for a question or for a change in plans?
Next time, notice how an agreement can reassure before the promised action is taken.
Best,
Andreas Walden
Share this with someone designing how AI-supported work changes hands.
Before changing a shared workflow, revisit how teams prepare people for AI updates.