Responsible AI & Governance · 90 minutes
Make human oversight meaningful
Examine a fictional AI coach whose feedback passes a human Approve button, calculate what careful review would cost, and propose the purpose, data, review, and ownership controls that make the oversight meaningful.
1.5 h
1
Step 1: The question
Using the review setup, explain why the Approve button does not give meaningful oversight of the AI coach, and calculate how many items per day a careful review could cover at current staffing. Then propose the controls that would make oversight meaningful, covering the coach's purpose, the people affected, the data reviewers and the model need, the review step itself, and who is accountable.
Worked scenario: an Approve button on AI feedback
Suppose a coding-practice platform runs an AI coach. For each submitted solution, the coach writes a verdict (meets or does not meet the exercise rubric) and suggestions. The platform's automated tests run separately on the same submission. Before feedback reaches the learner, it enters a review queue. All figures are synthetic.
| Item | Value |
|---|---|
| Feedback items per day | 6,000 |
| Reviewers | 4 a day, each with 6 hours of review time |
| Median time on the review screen | 8 seconds |
| Careful review in a pilot (code, rubric, and test results open) | Median 90 seconds per item |
| Outcomes over the last 30 days | 96% approved by a reviewer, 4% published by the timeout, 0% rejected |
| Audit of 200 randomly sampled approved items: Python and JavaScript | 8 of 160 had wrong feedback |
| Same audit: other languages | 10 of 40 had wrong feedback |
In the audit, a senior reviewer rechecked each item with the code, rubric, and test results open. Most wrong items were correct solutions marked as not meeting the rubric.
How review works today:
- The review screen shows the coach's feedback and the learner's name and email. It does not show the submitted code, the exercise rubric, the test results, or any signal of the coach's uncertainty.
- The only actions are Approve and Skip. A skipped item returns to the queue, and any item still queued after 4 hours is published automatically. There is no field for a reason.
- A dashboard ranks reviewers by items cleared per hour.
- The coach's prompt includes the submitted code, the exercise statement, and the learner's full profile (name, email, employer, and years of experience) "to personalize tone".
- Learners see each feedback item labeled "Reviewed by our team" and have no way to dispute it.
- The coach's purpose has never been written down. A product manager has asked whether its verdicts could also decide which learners earn a public "verified skills" badge.
- No one is named as responsible for feedback quality. Complaints go to the general support queue, and no one has clear authority to pause the coach.
Constraints
- Reviewer staffing stays at 4 people with 6 hours of review time each this quarter
- Work from the setup's figures and label any number you estimate
- Describe engineering and operating controls; this is not a legal or compliance assessment
Deliverables
- Purpose: a one-sentence intended use for the coach, and how it applies to the badge request
- Impacts: who is affected by wrong feedback, read from the audit with its sample sizes
- Data: what reviewers and the model need, and what to remove
- Oversight: the evidence from the setup, the review-capacity calculation, what reviewers see and can do, and how review quality is measured
- Accountability: a named owner, an escalation route, authority to pause the coach, and a learner correction path
2
Step 2: Hints
Optional. Each hint gives a little more away.
3
Step 3: Your answer
Kept in this browser. Saving your score below keeps a copy with it.
0 / 5000
4
Step 4: Reference answer
5
Step 5: Score yourself
Score each dimension 0, 1, or 2, then save the score to your history. Readiness thresholds are editorial estimates.
Grades your written answer against this rubric with your own AI provider key from Settings. It is a second opinion: the score you save is still yours.
What earns a 2: Cites the setup's screen, actions, and timing, computes review capacity from staffing and volume, and names a review-quality measure other than the approval rate.
0 of 10 · Scored 0 of 5 dimensions
Ready at 8 of 10, with no 0 in Oversight
Kept in this browser until you sign up, then it moves to your account.
Complete this level when you can turn a broad principle into a concrete control and evidence of its operation.
Your saved scores
Practice saves and mock interview answers on this question, newest first.
No saved scores yet. Score every dimension, then save.
6
Step 6: Follow-up questions
An interviewer may ask these next.
How does governance differ from security?
Why is a human in the loop insufficient by itself?
Does high average accuracy establish fairness?
What is residual risk?
Does a risk framework prove compliance?
7