# AI Bias & Fairness Review

Review an AI-assisted decision, ranking, recommendation, or generated output for unfair assumptions, proxy variables, unequal errors, and missing oversight.

## Prompt

Act as a practical bias and fairness reviewer. Evaluate the AI-assisted workflow or output below without assuming that fairness can be proven from a short description. Focus on affected people, decision impact, evidence gaps, and safer controls.

Workflow or output:
{{workflow_or_output}}

People affected and decision context:
{{affected_people_and_context}}

Available data, criteria, and evaluation results:
{{data_criteria_and_results}}

Conduct the review using this structure:

1. Decision map: explain what the system predicts, ranks, recommends, generates, approves, rejects, or prioritizes, and who acts on the result.
2. Affected groups: identify groups that may experience different outcomes, including groups not explicitly recorded in the data. Do not infer personal attributes about named individuals.
3. Bias pathways: check representation gaps, historical bias, label quality, proxy variables, subjective criteria, feedback loops, accessibility barriers, language or cultural assumptions, and automation bias.
4. Error impact: compare the consequences of false positives, false negatives, lower rankings, omissions, and delayed review across affected groups.
5. Evidence status: separate observed disparities from plausible risks and unknowns that require measurement.
6. Control review: assess human oversight, appeal, explanation, sampling, monitoring, documentation, threshold choices, and whether AI should make the final decision.
7. Safer redesign: recommend concrete changes to data, criteria, workflow, review, and communication.
8. Evaluation plan: propose measurable tests, subgroup checks, qualitative review, edge cases, and a schedule for monitoring after deployment.

Rules: Do not manufacture demographic data or claim that a system is unbiased. Avoid using protected or sensitive attributes unless their use is lawful, justified, and reviewed by qualified experts. For hiring, lending, housing, education, health, insurance, or public-service decisions, recommend legal, domain, and fairness expertise before deployment.

## Best for

Teams reviewing AI-assisted hiring, moderation, ranking, support prioritization, education, recommendations, or other workflows that affect access, opportunity, or treatment.

## Compatible tools

- Claude
- ChatGPT

## How to use

- Describe the decision, affected people, input data, criteria, automation level, and consequences of errors.
- Provide real evaluation results when available, but use aggregated and privacy-safe data.
- Use the output to plan measurement and human review rather than to declare the system fair.
- Escalate high-impact decisions to legal, domain, accessibility, and fairness specialists.

## Customization tips

- Specify whether the AI advises a human or makes the final decision.
- Add known complaint patterns, appeal outcomes, or subgroup error rates when lawfully available.
- Ask for separate recommendations for data, model, interface, policy, and operations.

## Example input

Workflow: an AI ranks job applicants from 0 to 100 using resumes and automatically rejects everyone below 55. A recruiter reviews only the remaining candidates. No subgroup performance data is available.

## Example output

High fairness risk. Automatic rejection makes model errors final for one group of applicants, while no subgroup or accessibility evaluation is available. Resume features can encode historical opportunity differences and proxy variables. Replace automatic rejection with human review, document the criteria, test error rates across relevant groups where lawful, review nontraditional resumes and career gaps, provide an appeal path, and monitor outcomes after deployment.