Decide with AI: Advice, Automation Bias and Delegated Decisions

All courses / AI Leader / Module 2 / Chapter 8

Learning objectives

After the lesson the participant can:

• Distinguish AI as adviser, drafter and decider and state the record each role requires.
• Summarise what studies measured about automation bias, explanations, aversion and appreciation, and human-AI team performance, with their populations.
• Design a seeded-error drill to calibrate trust in a drafting assistant.
• Classify candidate decisions as suitable or unsuitable for Level 3 Bounded delegation.

Key concepts

Adviser, drafter, decider

The three roles AI holds in a leader's decisions. The drafter is the most underrated, because framing is a choice. The decider exists only where a named person formally delegated a class under a recorded profile.

Calibrated trust

The weight a leader gives to AI advice tracks how often that kind of advice has been right in that kind of situation. It can be measured and it expires. It is built by reproduction by a person, scripts before a second model, seeded errors and outcome tracking.

Seeded-error drill

A sample of the assistant's documents is seeded with known errors, such as an altered figure or a missing dissent, and the share readers catch is measured. Results are process data and are never used against an individual.

Routine delegation

A decision qualifies for Level 3 Bounded when it is narrow, reversible within a window, governed by an existing rule that can be written as a check outside the model, small in reach and frequent enough that sampling produces evidence.

Delegation by drift

The concern in the debate: the leader stops reading the source, then stops forming a view, then the approval threshold rises, and at no point does anyone decide to hand over the decision, so nobody writes a record.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Teaching plan

Discussion

1. Readers at Hansa Mutual caught 7 of 12 seeded errors and the board raised no ceiling but added a mechanical figure check. What would your board have done with a catch rate of 58 per cent, and why?
2. Explanations raised acceptance of recommendations whether or not they were correct, while cognitive forcing reduced overreliance and was rated lowest by participants. What does that imply for how a board pack should present AI-prepared options?
3. Nordvik chose the lowest ceiling after an incident whose record could not name its cause; Hansa delegated two classes under a numeric stop condition. Which reasoning fits your organisation's forums, and what evidence would move you?

Written exercise

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

1. What did the meta-analysis by Vaccaro, Almaatouq and Malone of 106 experimental studies find on average?

   1. Combinations outperformed both humans alone and AI alone, most clearly in decision tasks
   2. Combinations did worse than the best of humans or AI alone, losing in decision tasks and gaining in content creation
   3. Combinations matched AI alone in every type of task that the reviewed experiments covered
   4. Combinations improved only where an explanation accompanied each recommendation shown to the person
   Show the answer

   B. Combinations did worse than the best of humans or AI alone, losing in decision tasks and gaining in content creation
   The average effect was a Hedges' g of -0.23 against the best of either alone. Losses appeared in decision tasks and where AI alone was better; gains in content creation and where humans alone were better.

2. Which two classes did Hansa Mutual's executive board allow at Level 3 Bounded in DLG-HM-EXE-01?

   1. Shortlisting candidates for board posts and scheduling board meetings
   2. Accepting risk exceptions below a threshold and routing incoming letters to the responsible unit
   3. Supplier approvals below EUR 2,500 and setting the board's agenda
   4. Travel approvals within policy below EUR 2,500 per trip and scheduling of board meetings
   Show the answer

   D. Travel approvals within policy below EUR 2,500 per trip and scheduling of board meetings
   Both are narrow, reversible, rule-governed and frequent. Supplier approval above a threshold, risk exceptions, agenda setting and shortlisting exercise judgement about interests, risk or persons and do not qualify.

3. Which condition does Nordvik Region's profile DLG-NR-EXE-01 state for any level above Level 1 Assisted?

   1. An evidence age of no more than 90 days, with provenance for every figure and a seeded-error drill
   2. A vote of the regional council and approval by the national supervisory authority for each decision class
   3. A second model from another provider that confirms every summary before release
   4. A catch rate of at least 58 per cent in the first drill on the summariser
   Show the answer

   A. An evidence age of no more than 90 days, with provenance for every figure and a seeded-error drill
   The profile asks for provenance for every figure, a mechanical comparison of figures with the source, evidence no older than 90 days and a seeded-error drill. The Region has measured nothing yet that would justify Level 2.