Decision Desk · Issue 05 · complete public learning edition

Should AI Help Choose People?

Decide whether AI may assist at one defined stage of an opportunity process—and whether the criteria, evidence, access, notice, contest, and human authority can support that use.

Begin the issue Build the Impact Map

Learning time90 minutes

Teaching12 cumulative lessons

InstrumentDecision Impact Map

BoundaryNo active selection

Opportunity-stage ladder

Eligibility ≠ screening ≠ scoring ≠ ranking ≠ recommendation ≠ decision.

Each stage has different evidence, consequence, and authority. Permission at one stage does not travel to the next.

Eligibility

Stage

Screening

Stage

Scoring

Stage

Ranking

Stage

Recommendation

Stage

Final decision

Stage

Complete issue curriculum

Twelve cumulative lessons.

Use the fictional fellowship exercise only. Never enter a real person’s application or decide an active opportunity.

Issue 05

Define the opportunity and one stage

Open

Hiring, admissions, grants, fellowships, procurement, and creator programs differ. Bound eligibility, screening, scoring, ranking, recommendation, or decision—not ‘selection’ in general.

Issue 05

Name the consequence

Open

Identify what changes for a person: visibility, delay, interview, funding, access, rejection, or appeal. Higher consequence demands stronger evidence and control.

Issue 05

Connect criteria to purpose

Open

A criterion must measure something relevant to the opportunity. Familiarity, prestige, formatting, employment continuity, or writing style may be convenient without being valid.

Issue 05

Find proxies and inherited patterns

Open

Location, school, language, gaps, names, networks, and behavioral traces can act as proxies. Removing explicit traits does not establish neutrality.

Issue 05

Handle missingness without punishment

Open

Missing data may reflect access, accommodation, nontraditional experience, system design, or error. The model must not convert absence into a negative fact.

Issue 05

Design accommodations before scoring

Open

People need an accessible way to participate, request accommodation, use an equivalent route, and avoid being penalized for the accommodation itself.

Issue 05

Test material and distributional effects

Open

Evaluate false exclusions, false advances, subgroup outcomes, intersections, sample limits, and consequence—not only aggregate accuracy.

Issue 05

Give notice people can use

Open

Explain where AI is used, what it does, what data matters, how to correct it, and how to reach a person before consequence.

Issue 05

Make explanation and contest meaningful

Open

A generic score or appeal mailbox is not contestability. Provide the reason, evidence, authority, timeline, capable reviewer, and power to change the result.

Issue 05

Constrain human reliance

Open

Automation bias can turn a nominal recommendation into a decision. Define independent review, disagreement, override, documentation, and accountability.

Issue 05

Monitor change and complaints

Open

Criteria, applicant pools, models, vendors, workflows, accommodations, and effects change. Set review triggers, complaint analysis, rollback, notice, and redress.

Issue 05

Issue a stage-specific disposition

Open

Prohibit, redesign, preserve human-led work, or evaluate a bounded non-decisional pilot. Never convert this course into an active candidate decision.

Critical stop conditions

Human opportunity requires more than a score.

criteria lack evidence of relevance to the opportunity
proxy effects or missingness are unexamined
material subgroup harm is unmeasured
the process or accommodation route is inaccessible
notice arrives after the consequential step
the explanation cannot support meaningful review
the appeal cannot change the outcome
AI directly or effectively makes a consequential rejection

Interactive instrument · device-local

Human Opportunity Decision Impact Map

Evaluate one stage of one fictional opportunity process. Do not enter applicant records, protected-trait data, accommodation details, or an active selection decision.

0/28impact gates evaluated

A human clicking “approve” does not cure an invalid or inaccessible system.
Consequential automated rejection, invalid criteria, unmeasured material subgroup harm, or illusory appeal fails closed.

Fictional profile condition	Purpose-valid criterion	Proxy risk checked	Missingness handled	Accommodation available	Material impact tested	Reason reviewable	Contest path works
Nonlinear experience	Nonlinear experience: Purpose-valid criterion	Nonlinear experience: Proxy risk checked	Nonlinear experience: Missingness handled	Nonlinear experience: Accommodation available	Nonlinear experience: Material impact tested	Nonlinear experience: Reason reviewable	Nonlinear experience: Contest path works
Disability accommodation	Disability accommodation: Purpose-valid criterion	Disability accommodation: Proxy risk checked	Disability accommodation: Missingness handled	Disability accommodation: Accommodation available	Disability accommodation: Material impact tested	Disability accommodation: Reason reviewable	Disability accommodation: Contest path works
Missing credential data	Missing credential data: Purpose-valid criterion	Missing credential data: Proxy risk checked	Missing credential data: Missingness handled	Missing credential data: Accommodation available	Missing credential data: Material impact tested	Missing credential data: Reason reviewable	Missing credential data: Contest path works
Proxy-heavy profile	Proxy-heavy profile: Purpose-valid criterion	Proxy-heavy profile: Proxy risk checked	Proxy-heavy profile: Missingness handled	Proxy-heavy profile: Accommodation available	Proxy-heavy profile: Material impact tested	Proxy-heavy profile: Reason reviewable	Proxy-heavy profile: Contest path works

Opportunity, stage, and stakesName the opportunity and one stage: eligibility, screening, scoring, ranking, recommendation, or final decision. What consequence follows?

Criteria, evidence, and prohibited proxiesHow does each criterion relate to the opportunity’s purpose? Which proxies, missing values, or historical patterns require prohibition or testing?

Notice, accommodation, explanation, and contestWhat is disclosed? How can a person request an accommodation, correct data, understand the reason, and obtain meaningful reconsideration?

Human authority and monitoringWho owns the decision? What evidence can they review? How are overrides, disagreements, subgroup effects, complaints, and changes monitored?

Bounded dispositionProhibit AI at this stageRedesign criteria and process firstKeep the stage human-ledEvaluate a bounded non-decisional pilot

Restoring your local map…

Completion boundary

The educational product and reusable Impact Map are complete.

Real use still requires current jurisdiction and opportunity-specific research, qualified review, representative validity and impact evidence, accessible participation, product/configuration testing, monitoring, and organizational authorization.

Return to the 12-month program Previous issue