King's College London · FACULTY OF ARTS & HUMANITIES

7AAVCD42 Chap.7 Automated Decisions in Social Services

- one subject, every graph, every model, every mark
5 Chapters3-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 7 of 10 · 7AAVCD42

Automated Decisions in Social Services

Benefits and risks share mechanisms

AI in education, healthcare, social care and public services is often justified through accessibility, personalisation and cost. The same data and optimisation can create privacy, security, trust and inequality problems. This chapter defines the decision, distinguishes support from substitution and follows data, recommendation, human discretion, consequence and appeal.

Its evidence section shows why aggregate performance is insufficient when error burdens and access differ across groups.

Define the decision and baseline

Begin with the public purpose and the decision being changed. A recommendation for extra practice differs from an automatic denial of treatment or eligibility. Identify the data input, target, output, reviewer and consequence.

Then define the previous barrier or service baseline. More processed cases do not necessarily mean better access or appropriate support.

Benefits should be testable. Personalisation needs an outcome, evidence about individual need and a way to recognise cases outside the model categories. Lower cost should include work shifted to applicants, frontline staff and appeals.

Consistency should be tested against the possibility that uniform rules reproduce an unequal history.

Test whether human review is meaningful

A person in the loop is not automatically a safeguard. The reviewer needs time, evidence and authority to depart from the system. Examine what information appears on the interface, how often outputs are overridden and whether organisational pressure rewards agreement.

A rubber-stamp review preserves the appearance of discretion while automation determines practice.

Support becomes substitution when the recommendation effectively fixes the outcome or when no practical route exists to challenge it.

Describe that transition explicitly rather than classifying the whole system with one label.

Evaluate unequal consequences

Overall performance can hide which groups receive false positives or false negatives, how serious those errors are and whether some people were absent from the data. Compare relevant error types, service access and data quality.

A group with fewer records may appear difficult to predict because the institution historically served it poorly.

Metrics do not decide which difference is acceptable. Rights, institutional purpose and the cost of each error shape the judgement. Show who benefits when a threshold changes and who becomes exposed.

If evidence is missing, propose monitoring rather than asserting equal treatment.

Make appeal usable

Notice, explanation, human review and correction form an accountability route. Formal availability is insufficient if a person lacks time, documents, language support or assistance.

Ask whether review occurs before the harmful consequence and whether repeated errors trigger institutional change.

Close by naming a governance condition and a non-model alternative. Improved staffing, simpler forms or wider service access may address the stated problem without prediction. Comparing alternatives prevents automation from appearing to be the only possible reform.

In this chapter

What this chapter covers

  • 01

    Decision Support

  • 02

    Personalisation

  • 03

    Public Services

  • 04

    Unequal Error

  • 05

    Appeal Rights

Worked example · free

Audit a debt-risk flag

Q [8 marks]. A public service flags households for investigation. Evaluate the process. The marks shown are an AskSia study allocation, not an official King's marking scheme.
Benefit and risk balance.Benefit and risk balance. The beam places access, personalisation and cost beside privacy, trust, security and inequality rather than treating either side alone.access and supportpersonalisation, costprivacy and trustsecurity, inequality
Benefit and risk balance. The beam places access, personalisation and cost beside privacy, trust, security and inequality rather than treating either side alone.
  • 2Define the target and decision.
  • 3Inspect data and discretion.
  • 3Test error and appeal.
Check whether the predicted target matches the legal decision, examine historical data and missing circumstances, identify reviewer discretion and compare error burdens. Require notice and meaningful review.
Sia tip — Do not report an average benefit without locating who receives each error.
Glossary

Key terms

Decision Support
Decision Support provides information or recommendations while leaving accountable judgement with a person or institution.
Automated Decision
Automated Decision is a consequential outcome made or materially determined by a computational process.
Appeal Route
Appeal Route allows an affected person to obtain explanation, review and correction of a decision.
FAQ

Automated Decisions in Social Services FAQ

How is support different from substitution?

Support provides evidence or recommendations that a responsible person can inspect and override. Substitution occurs when the system effectively makes the consequential decision or removes meaningful judgement.

Why can personalisation create inequality?

Personalisation depends on chosen outcomes, available data and categories. People poorly represented in the evidence may receive unsuitable recommendations while appearing difficult to predict accurately.

What should a fairness analysis include?

Compare relevant error types and access across groups, connect them to data and thresholds, identify the consequences and state which institutional response or monitoring rule should follow.

Study strategy

Assessment move

Map one public decision from data to appeal, then rewrite the benefit claim after adding group-specific errors, human discretion and an alternative non-model intervention. Define the service baseline and decide which outcome captures the public purpose rather than processing volume. Mark where the reviewer receives evidence, how they can override the recommendation and whether organisational pressure makes that authority real.

Compare false positives and false negatives for affected groups and explain the consequence of each. Walk through the appeal as a person with limited time or documents to test practical access. Finish by setting a monitoring trigger and naming an institutional repair that does not depend on better prediction alone. Compare the automated proposal with improved staffing, simplified forms or wider eligibility.

This makes the recommendation answer the public problem rather than merely selecting the most advanced tool. Rewrite the final recommendation after imagining that appeal capacity is already overloaded.

A+Everything unlocked
Unlocks this Bible + all 18 of your subjects at King's College London, plus 1,000+ more Bibles across the AskSia library.
Sia - your 7AAVCD42 tutor, unlimited, worked the way the exam marks it
The full 3-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works