All courses / AI Boss / Module 6 / Chapter 26
Learning objectives
After the lesson the participant can:
- Explain why a requirement of a human in the loop does not by itself make a decision human.
- State the six conditions under which oversight is real and how each can be recorded.
- Distinguish oversight of the system, which can be sampled, from the decision in the case, which cannot.
- Set up an oversight-quality metric with floors, seeded cases and aggregation agreed with the representatives.
Key concepts
- Selective adherence
- The finding that advice aligned with group stereotypes was adopted more readily, whether it came from an algorithm or a person. The record of departures must therefore show for whom overseers depart, not only how often.
- Six conditions of real oversight
- Time per decision; authority to deviate without heavier justification; information independent of the system; a measured deviation rate; consequences that do not punish deviation; and a stop path outside the system.
- Decision in the case
- The decision about one person, such as whether this applicant is invited. It cannot be sampled, because a sample showing ninety-five sound decisions gives the other five people no decision.
- Oversight record
- For one delegation profile and one period, who oversaw, what they saw, what they did and with what authority. It holds the evidence of the observed degree and names no overseer in its figures.
- Designing volume
- Volume is set by the owner: decisions multiplied by minimum time give the hours to plan. Where reading is not possible at the given volume, the class is reclassified, the volume reduced or fewer decisions produced.
Interactive exercises
Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.
Interactive exercise 1 · Put in order
When oversight fails: measures of rising strength
Chapter 26 gives the owner five measures of rising strength when oversight is found to be failing. Put them in order from the mildest to the strongest.
Interactive exercise 2 · Scenario
Who approved too fast?
Apply the chapter's three rules for measuring the review, not the reviewers, to the situation below.
Teaching plan
| 15 min | The teacher presents the survey of oversight policies, the reported case of approvals at 1.2 seconds each, and the research on automation bias, selective adherence and algorithm aversion. |
| 15 min | Presentation of the six conditions with their anchors in Article 14(4) and Article 26(2) of the AI Act and Article 10(2) of the Platform Work Directive, as organisational implications. |
| 20 min | Reading of OVR-NR-ROS-2027-1: participants classify the observed degree from the figures and propose measures before seeing those of AIDR-0630. |
| 30 min | Exercise: groups set up an oversight-quality metric for one decision class and compute the hours of oversight the volume requires. |
| 10 min | Debrief on the Polish recruiting record: what drifted, why volume was the cause, and what version 1.5 of the profile changed. |
Discussion
- Ingrid Halvorsen asked why the design allowed two-minute approvals, not who made them. What changes in an organisation when findings are made against the profile and not the person?
- A screen on which accepting is one click and departing needs a sign-off teaches acceptance. Where do your own approval screens make departure more costly than acceptance?
- The answer to Q9 places accountability with whoever set the volume. Who sets the volume of decisions per decider in your organisation, and do they know it?
Written exercise
Exercise · 30 minutes
An oversight-quality metric for one class
Choose one decision class in which a person approves or decides on a system's proposal: roster approval, shortlisting on a ranking, or a pay proposal on a score. Name the human role: oversight of the system or the decision in the case. Write the six conditions for this class. Choose the measures and set floors and ceilings from the chapter's defaults, including a floor in minutes suited to the class. Design the seeded cases and the gate that keeps them from reaching a person. Fix the aggregation, who sees which figures and the excluded uses. Then multiply the expected number of decisions by the minimum time and state the hours of oversight to plan, and what happens when demand exceeds them.
Deliverable: A metric specification with human role, six conditions, measures with floors and ceilings, seed design, aggregation and excluded uses, plus the volume calculation and the rule when it is exceeded.
Self-test
Choose an answer to see whether it is right, or open the answer below each question.
What did the figures in OVR-NR-ROS-2027-1 lead Nordvik Region to conclude about roster approval?
Show the answer
B. That the design was at fault: a screen let eleven flags be closed with one click and a wave went live before four of seven managers were trained
The record names no overseer and tests the design. AIDR-0630 blocks approval until each rest-rule flag has its own resolution, plans two hours per roster and requires training before a unit goes live.Why can the decision in an M3 Consequence case not be replaced by sampled oversight of the system?
Show the answer
D. Because a sample showing most decisions sound gives the remaining people no decision at all
Each consequence decision concerns one person and must be real in every case. Oversight of the system is added to it, never substituted for it; sampling is how Level 3 Bounded is made safe for M1 Direction.What caused Level 1 Assisted to drift in the Polish recruiting team, according to OVR-HM-REC-2027-Q1-PL?
Show the answer
A. Volume: nineteen reviews per recruiter-hour at peak against a cap of twelve
Recruiters rejected applications below the cut in a median of 38 seconds against a floor of 120, and only three of six seeded applications were caught. The remedy closes postings at readable volume and shows the application before the rank.