AI-Assisted Decision Safety Checklist
A plain-language checklist for keeping AI-assisted decisions traceable, reviewable, challengeable, correctable, reversible, and accountable.
Use this checklist before relying on AI to support a decision that affects people, resources, rights, safety, access, discipline, eligibility, money, health, education, housing, public services, infrastructure, or public trust.
The Checklist
Before using or accepting an AI-supported recommendation, ask:
- What decision is actually being made?
- Who could be affected if this is wrong?
- What evidence supports the AI's output?
- What evidence contradicts it?
- What assumptions are being made?
- Can a human reviewer challenge the result?
- Can the decision be paused before harm occurs?
- Can the decision be corrected or reversed?
- Is there a record of how the conclusion was reached?
- Who is accountable if the decision causes harm?
Risk Levels
Low
Example uses: Drafting, brainstorming, summarizing, formatting.
Minimum safeguard: Human review.
Moderate
Example uses: Internal recommendations, prioritization, triage, workflow support.
Minimum safeguard: Evidence check plus named reviewer.
High
Example uses: Hiring, benefits, discipline, healthcare support, finance, housing, education, public services.
Minimum safeguard: Audit trail, human challenge process, and correction path.
Critical
Example uses: Weapons, emergency systems, mass surveillance, infrastructure, large-scale automated decisions.
Minimum safeguard: External review, strict deployment controls, and clear accountability.
Decision Resilience Test
A decision is more resilient when it can be:
- Traced
- Reviewed
- Challenged
- Corrected
- Reversed
- Accounted for
If a decision cannot be traced, challenged, corrected, or reversed, it is not decision-resilient.
Practical Use
This checklist is designed for:
- Public officials
- Schools and universities
- Healthcare administrators
- Nonprofit leaders
- Business teams
- Researchers
- Journalists
- Civic technology projects
- Anyone reviewing AI-assisted decisions that may affect real people
Boundary
This checklist is a public governance aid. It is not legal advice, medical advice, safety certification, compliance certification, or a substitute for professional review.