What this prompt does
This prompt casts the model as an AI ethics researcher and asks it to build a full bias-audit framework around the system you describe. You supply [system_purpose], [model_type], [training_data], [protected_attributes], and [deployment_context], and it returns a structured audit covering data assessment, fairness metrics, testing methodology, mitigation, monitoring, documentation, and stakeholder communication.
The structure works because fairness is impossible to evaluate in the abstract — it depends entirely on who the system affects and how. By keying every section to your [protected_attributes], the prompt produces slice-based tests, a fairness-metric table (demographic parity, equal opportunity, predictive parity, calibration) with thresholds, and a model card grounded in your real use case. Crucially, it flags that these metrics can conflict and asks the model to recommend which one matters most for [system_purpose] — a decision teams too often skip.
When to use it
- You're deploying an AI system that affects people's outcomes (hiring, lending, screening) and need fairness baked into engineering.
- A regulator or
[compliance_framework]such as the EU AI Act or NYC Local Law 144 requires documented bias testing. - You need a model card and demographic performance breakdown before launch.
- You suspect your
[training_data]reflects historical inequities and want a structured way to surface it. - You're setting up ongoing monitoring and need a fairness dashboard and audit cadence, not a one-time check.
- You have to brief executives and need both a technical report and a plain-language risk matrix.
Example output
You get a multi-section audit: a data-composition analysis by [protected_attributes], a fairness-metrics table with definitions, thresholds, and score placeholders, a slice-based and adversarial testing plan including counterfactual tests, stage-by-stage mitigation options (pre-, in-, and post-processing), a monitoring plan with drift detection and an audit schedule, a Model Card template, and stakeholder materials — an executive summary, technical report, and likelihood-by-impact risk matrix.
Pro tips
- List every relevant attribute in
[protected_attributes]; the slice and intersectional tests are generated directly from this list. - Be honest in
[training_data]about where it came from and who was excluded — that's what surfaces historical bias. - Use
[additional_metric]and[adversarial_test]to add measures specific to your domain, like individual fairness or name-swap tests. - Name your real
[compliance_framework]so the documentation section maps findings to actual control requirements. - Remember the prompt computes nothing — the metric tables are templates you must fill with real model outputs.
- Push back on the recommended priority metric; ask the model to argue the trade-offs before you commit to one definition of fairness.
