Leaders in the Public Authority: Mandate, Legitimacy and Democratic Accountability

All courses / AI Leader / Module 3 / Chapter 18

Learning objectives

After the lesson the participant can:

• Explain why direction on what AI is for in a public service is a political decision and what a committee may not delegate.
• Describe how reasons, hearing and appeal constrain the design of AI support in case handling.
• Summarise the evidence on whether human review performs the check the law assumes.
• Draft a decision record for an AI-supported public decision and a quarterly view for elected oversight.

Key concepts

Chain of delegation

The council decides matters of principle; a committee may delegate decisions to members, employees or, since 2022, an automated decision function, except where Chapter 6, Section 38 of the Swedish Local Government Act forbids it. A caseworker's decision is the Region's decision under that chain.

Reasons, hearing and appeal

Under the Administrative Procedure Act the party is informed of material and may respond, a significant decision carries a clarifying statement of reasons, and it can be appealed within three weeks. The decision must rest on circumstances the official can name and defend without the system.

Necessary and not sufficient

The evidence on human review is mixed: experiments found selective adherence to advice matching stereotypes, and a survey of 41 oversight policies argued people cannot perform the oversight assumed. The caseworker is a necessary check, so the leader answers also for the system.

Deployer duties at procurement

For the procured benefits-triage system the AI Act makes the Region the deployer: use according to instructions, monitoring, logs kept at least six months, information to persons concerned, a fundamental rights impact assessment under Article 27 and registration under Article 49.

The council's quarterly view

Elected oversight sees the system and not the case. Its specification: the register with degrees of delegation, changes since the last report, departure rates by group of applicant, appeal outcomes, Explain and Reach test results, and the age of evidence behind each claim.

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. The Local Government Act allows delegation to an automated decision function in some matters, and the book's ceiling is stricter. What reasons does the chapter give for holding decisions about a named person's support at Level 0 Advisory?
2. A supplier update can change a system without a council decision. How would your authority's elected level learn that a model or a degree of delegation had changed?
3. Hansa's legitimacy comes largely from supervision and contract and shows only when something fails. What would an insurer gain from borrowing the reasons log and the quarterly view voluntarily?

Written exercise

Self-test

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

1. What does the chapter propose that Nordvik Region's decision AIDR-1018 should do?

   1. Raise the benefits-triage system from Level 0 to Level 2 Supervised
   2. Delegate living-support decisions to an automated decision function under the delegation order
   3. Keep the triage rank out of the grounds and out of the reasons of the decision
   4. Publish each caseworker's departure rate from the system's advice
   Show the answer

   C. Keep the triage rank out of the grounds and out of the reasons of the decision
   The system orders the queue; it is not a ground for the decision. Keeping the rank out of grounds and reasons preserves a decision that rests on circumstances the official can name and defend without the system.

2. In the council's quarterly view, what does a rate near zero at which caseworkers depart from the system's advice signal?

   1. That the system has reached the accuracy needed for Level 3 Bounded
   2. That rubber-stamping may be taking place in the case decisions
   3. That appeals against decisions will fall in the next period
   4. That the reasons log can be discontinued for routine cases
   Show the answer

   B. That rubber-stamping may be taking place in the case decisions
   The table reads a departure rate near zero as a signal of rubber-stamping and a pattern across groups as a signal for the equality review, with sample method, period and size stated.

3. According to the chapter, what did Green argue after surveying 41 policies that prescribe human oversight of government algorithms?

   1. That people cannot perform the oversight the policies assume, so the policies can legitimise faulty algorithms
   2. That trained caseworkers reliably correct the errors of ranking systems
   3. That human oversight is unnecessary once an algorithm has been registered
   4. That oversight works when the reviewer sees an explanation with each output
   Show the answer

   A. That people cannot perform the oversight the policies assume, so the policies can legitimise faulty algorithms
   Green proposed institutional oversight, with justification of the algorithm and democratic review before adoption, as the central mechanism. The chapter holds this beside experiments that found no automation bias as such but selective adherence.