Public Benefits AI Decision Simulation

Learning Goal

Integrate capability judgment, verification, data boundaries, bias awareness, accountability, workflow design, and risk escalation into one defensible recommendation when speed, access, privacy, fairness, accountability, and escalation pull in different directions.

Capstone Simulation

The public benefits backlog decision

0/6

Scenario

A public benefits office wants to use AI to review a growing service backlog, identify recurring access barriers, draft response options, and flag residents who may need urgent human follow-up. Staff are overloaded, appointment slots are limited, and a supervisor wants a triage plan by the end of the week.

Evidence available

• Small synthetic sample of recent service requests.
• Several requests mention benefit status confusion and appointment delays.
• Some records contain names, case IDs, household details, and sensitive context.
• Some urgent requests are short, informal, translated, or missing details.

Proposed AI output

"Benefit status confusion is the top issue at 42%. Low-detail requests should be deprioritized. Recommended intervention: automatically send standard eligibility instructions and route only complete, detailed requests to a human reviewer."

1. Bound the capability

Select the parts of the request where AI may assist. The hard part is deciding where assistance becomes delegated judgment.

Summarize recurring service-access themesList questions for benefits-staff follow-upDraft service-response options for reviewFlag possible urgent access needs for human triageAutomatically deprioritize short or incomplete requestsSend benefit or eligibility notices automatically

2. Inspect the AI output

Classify each claim against the synthetic evidence notes. Treat polished language as neither proof nor disproof.

Benefit status confusion and appointment-access delays appear in the evidence notes.

SupportedOverstatedUnsupported

Benefit status confusion caused 42% of service requests this month.

SupportedOverstatedUnsupported

The top three causes of delayed access are now known.

SupportedOverstatedUnsupported

Short or informal requests are probably lower priority.

SupportedOverstatedUnsupported

3. Set data boundaries

Choose controls before any prompt or workflow uses public service records.

Remove direct identifiers before promptingAbstract case details into needed categoriesUse only an approved public-office workspace or toolSeparate urgent access needs from general theme analysisPaste the raw case backlog so context is not lost

4. Surface fairness and harm concerns

Select the concerns that should be addressed before using AI to flag priority, access, or urgency.

Short requests may hide urgent needs or deadlinesTone, grammar, or language may distort priorityResidents with access barriers may provide less detailThe model may miss case history, notices, or prior appeals

5. Build accountability into the workflow

Choose the review controls that make human accountability concrete.

Name a human owner for any final service decisionGive reviewers authority to override AI flagsAudit a sample of outputs against recordsKeep evidence links or case record IDs for reviewDefine escalation triggers before use

6. Make the recommendation

Choose your final action. A defensible answer should force the tradeoff into the open: speed and service access matter, but so do privacy, fairness, accountability, and escalation when rights or urgent needs may be affected.

Launch automated routing nowPilot with safeguardsPause until controls existEscalate for formal public-office review

7. Write the structured decision memo

Complete each field as if this will be read by a public-office manager deciding whether the AI-assisted workflow can proceed. Completion depends on a defensible judgment, not guessing every hidden answer.

RecommendationWhat should the office do next?Allowed AI roleWhat may AI help with, and what must it not decide?Data boundaryWhat data must be removed, minimized, or kept in an approved environment?Verification planWhat claims or outputs must be checked against source records?Human review ownerWho can inspect, override, stop, or approve the workflow?Fairness or harm concernWho could be burdened by error, missing context, or proxy signals?Escalation triggerWhat condition requires formal review before proceeding?Final tradeoff judgmentWhat tradeoff are you accepting among speed, access, privacy, fairness, accountability, and escalation?Write at least 256 characters. Use at least 42 words.

Capstone self-check

0/6 checks

This section shows what the page can detect in your answer so far. These checks support reflection; they do not verify correctness, policy compliance, legal compliance, or role authorization. Use them to improve your memo, not to chase a perfect score.

0/6 checked by you

• Bounds what AI may and may not doThe memo should preserve AI as assistance, not as the public-service decision maker.
• Separates supported claims from unsupported claimsThe capstone tests whether fluent AI output is inspected against evidence.
• Sets data boundaries before useUseful analysis does not require exposing every raw detail.
• Names representational or access harmsPriority systems can disadvantage people through proxies such as detail, tone, access barriers, or context gaps.
• Creates a human review loop with authorityReview is only meaningful when humans can inspect, correct, stop, or escalate use.
• Writes a defensible final judgmentA strong memo states the decision, conditions, verification, review, escalation triggers, and accepted tradeoff.

Before reveal

• Complete all six interaction sections
• Choose a conditional, paused, or escalated final action
• Complete all decision memo fields
• Write a substantive decision memo

Reveal capstone review

Post-Reflection

Before your learning record

Complete the capstone review first. The post-reflection is the final step before the completion page.

Capstone Scenario

Your public benefits office receives a proposal:

Use AI to review a service backlog, identify the top causes of delayed access, draft response options, and flag residents who may need urgent follow-up.

The proposal is attractive because staff have limited time, appointment slots are scarce, and residents are waiting for benefit status updates or service access.

Learner Role

You are not asked to be anti-AI or pro-AI. You are asked to recommend whether and how this use should proceed in a regulated public-office setting.

Simulation Shape

The learner receives a realistic public-service request involving AI. Across several stages, new information appears:

1. initial request
2. proposed AI use
3. sample AI output
4. data sensitivity concern
5. affected resident concern
6. accountability or review gap
7. final tradeoff judgment

Stage Details

Stage 1: Capability

The learner decides which parts of the proposal AI might help with and which parts would delegate public-service judgment.

Stage 2: Reliability

The learner reviews a sample AI summary that includes unsupported ranking and confident recommendations.

Stage 3: Data Boundaries

The learner sees that the backlog includes direct identifiers, case details, household information, and sensitive context.

Stage 4: Harm and Fairness

The learner notices that some residents are more likely to be deprioritized because their requests are shorter, less formal, translated, incomplete, or harder to categorize.

Stage 5: Accountability

The proposed workflow says a human will review AI flags, but the reviewer has limited time and no clear authority to override routing, stop use, or escalate a case.

Stage 6: Escalation

The learner decides whether to pilot with safeguards, pause, or escalate. The decision must name the tradeoff among speed, service access, privacy, fairness, accountability, and escalation.

Evidence of Learning

The learner produces a short decision memo or structured recommendation that explains:

• whether to proceed
• what safeguards are needed
• what must be verified
• what data boundaries apply
• who should review or approve
• what should be documented
• what would trigger escalation
• what tradeoff the final recommendation accepts

Decision Memo Template

Recommendation:

What AI can help with:

What AI should not decide:

Data boundaries:

Verification requirements:

Human review requirements:

Fairness or harm concerns:

Escalation triggers:

Tradeoff judgment:

Conditions for proceeding:

Self-Check Rubric

• Emerging: gives a general opinion about whether AI should be used.
• Developing: identifies two or three major risks but treats them separately.
• Proficient: recommends a modified workflow with safeguards, review, and documentation.
• Advanced: explains tradeoffs, borderline judgments, escalation triggers, and residual risk in a public-service context.

Transfer Principle

Responsible AI use is a chain of decisions. Strong AI literacy means noticing when that chain needs evidence, safeguards, human review, or escalation before AI affects access to public services.

Grounding

The capstone integrates the lab’s full risk-management pattern. It is also consistent with the EU AI Act’s emphasis that AI literacy depends on staff knowledge, experience, training, and the context in which AI systems are used.

• NIST AI RMF
• EU Artificial Intelligence Act

Source note: the EU AI Act link is included for its AI-literacy framing, especially Article 4. This lab is not a legal compliance program and does not determine whether an organization has met EU, sectoral, or jurisdiction-specific requirements.