Answer: Accountability, Transparency and Legitimacy

All courses / AI Leader / Module 2 / Chapter 12

Learning objectives

After the lesson the participant can:

Explain answerability as a relation between an actor and a forum and identify the forums of a given decision.
Distinguish hierarchical links from forum links in a chain of answerability.
Apply Name, Explain and Reach to a sample of decisions and identify which condition makes a test fail.
Draft an accountability map for one consequential decision with forums, evidence and dated test results.

Key concepts

Actor and forum: Bovens defines accountability as a relationship in which the actor must explain and justify conduct, the forum can question and pass judgement, and the actor may face consequences. Name ties the obligation to a person, Explain gives the forum something to question, Reach makes the consequence fair.
Chains of answerability: Hierarchical links carry instruction and review; forum links carry questioning and consequence before a council, a supervisory board or a court. A decision can be sound on the hierarchical chain and fail on the forum chain when the person questioned cannot say what happened.
Misplaced responsibility: Most systems a leader meets are deployed within a mandate by an organisation that chose them, so the practical gap is not that no one is responsible but that responsibility lands on the person nearest the system while others designed the process and granted the level.
Right to an explanation: Article 86(1) of the AI Act gives a person subject to a decision based on the output of an Annex III high-risk system, with the exception of point 2, the right to clear and meaningful explanations of the system's role and the main elements of the decision, under stated conditions.
Accountability map for a decision: A small artefact in version control recording, for one consequential decision, who answers to which forum, by which evidence, and where the three tests were last run, naming the owners of the review design, the grant of the level and the stop right.

Interactive exercises

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

Interactive exercise 1 · Match

Forums and their questions

Match each forum to which a leader answers with the question that forum asks, according to the chapter's table.

Interactive exercise 2 · Scenario

Who is written into the map

Apply the chapter's account of the accountability gap to the design of an accountability map.

Teaching plan

15 min	Lecture on Bovens's definition and the six forums; participants list the forums that could question one of their own decisions and the means each has.
15 min	The teacher traces the chains through DORA Article 5, NIS 2 Article 20, the Swedish Administrative Procedure Act as illustration, and Article 26(7) of the AI Act, as organisational implications and not legal advice.
15 min	Worked case: the refusal of a living-support application at Nordvik Region. The class applies the three tests and finds the condition under which Reach fails although a named person signs.
35 min	Exercise: run the three tests on recent decisions and write the accountability map.
10 min	Quiz and the closing instruction: put the map on the agenda of the forum that would question you, before it asks.

Discussion

Nordvik and Hansa are said to fail the same way: by giving the person at the end of the chain a task and no means to carry it out. Where in your organisation does a named signatory lack the time or authority that Reach requires?
The further delegation goes, the more answerability moves from the individual decision to the design of the mandate. Who in your organisation could explain the design of a Level 3 mandate to a supervisor today?
Legitimacy in a public authority rests on how a decision was reached, since the citizen did not choose the authority. Which of hearing, mandate, reasons and appeal is most at risk when an AI system orders the queue?

Written exercise

Exercise · 35 minutes

Three tests and an accountability map

Pick one consequential decision that an AI system helps to prepare in your organisation, or the claims decision at Hansa Mutual across handler, team leader and chief claims officer. Describe how you would run Name, Explain and Reach on ten recent cases: what you would look for in the decision letter, in the reasons log and in the performance measures that apply to the decider. Then write the accountability map in the form of ACC-NR-2030-01: the decision and its class, the degree of delegation, the deciders and whom they answer to, the owner of the review design, who granted the level, who holds the stop right, each forum with the evidence it would be shown, the test results with dates and a review date.

Deliverable: A test plan for ten cases and a one-page accountability map naming decider, design owner, forums, evidence and review date.

Your work

Self-test

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

In the worked case of a benefit refusal at Nordvik Region, under which condition does Reach fail even though a named caseworker signs?
Show the answer
B. The unit's performance dashboard rewards clearing cases within minutes of the ranking
Reach needs authority, time and information to depart from the output. A measure that rewards clearing cases within minutes of the ranking removes the time, so the named signer cannot in practice set the ranking aside.
What does the chapter say is the form the accountability gap takes in most deployed systems?
Show the answer
C. Responsibility that is misplaced on the person nearest the system
Matthias's responsibility gap is an argument about learning automata. In practice systems are deployed within a mandate, and the person nearest the system absorbs consequences while others designed the process, the pattern Elish calls the moral crumple zone.
At Level 3 Bounded, what must the leader be able to do to pass Reach, according to the chapter's table?
Show the answer
A. Suspend a class of actions through a stop path outside the system
At Level 3 approval has moved from the action to the mandate and policy, so Reach means being able to suspend a class of actions through a stop path that does not run through the system being stopped.