AI Ethics & Fairness Lab
An interactive teaching and research lab where you measure algorithmic bias yourself, compare fairness metrics, try differential privacy, explain model decisions with SHAP and LIME, audit a model before deployment, and make ethical decisions in realistic scenarios. Everything runs in your browser, with no sign-in and without sending your data.
العربيةLab objectives
- See how bias travels from data into model decisions, and measure it with numbers.
- Understand the eight fairness metrics, their formulas and why they conflict.
- Try privacy-protection tools and measure their cost in accuracy.
- Explain model decisions and know the limits of each explanation method.
- Practise auditing a model and making a responsible, human-overseen deployment decision.
How to use the lab
- Pick a module or start with the beginner experiments.
- Read the objective and data description, set the parameters and press «Run».
- Follow the charts, numbers and the interpretation written from the computed values; change the seed to test stability.
- Send your results to the assistant for an explanation, download a PDF or CSV report, then move to the suggested next experiment.
Lab modules
Eight modules, each with real experiments computed in your browser, with explanation, interpretation, conclusion and a next step.
Experiments by level
Beginner 16
- Bias simulator: university admission
- Bias simulator: hiring
- Bias simulator: loans
- Comparing groups at different thresholds
- The eight fairness metrics calculator
- Personal or anonymous data?
- Pseudonymisation
- Permutation feature importance
- Balanced vs imbalanced data
- Audit checklist
- Responsible AI principles: self-assessment
- Analysing documented failure cases
- Who is responsible? Developer, user, organisation
- Scenario: A job applicant rejected by an automated system
- Scenario: Surveillance without consent
- Scenario: Biased university admission
Intermediate 18
- Under-representation and error
- Changing data properties and missing data
- Interactive fairness explorer
- Which fairness metric fits my context?
- Anonymisation by generalisation (k-anonymity)
- Differential privacy and epsilon
- Re-identification by linkage
- SHAP values for one decision
- LIME explanation and its stability
- Partial dependence and ICE curves
- Counterfactual: what would change the decision?
- Changing the class ratio
- ROS, RUS and SMOTE before and after
- Model audit studio
- Pre-deployment risk matrix
- Deployment decision: a rule-based gate
- Scenario: Unequal medical diagnosis across groups
- Scenario: An unexplainable credit decision
Advanced 9
- Why fairness metrics cannot all hold at once
- Privacy budget
- Membership inference (safe, educational)
- Black box vs interpretable model
- Leakage: SMOTE before the split
- Class balance is not group fairness
- Audit your own predictions (CSV)
- Human-in-the-loop simulation
- Scenario: A sensitive decision without oversight
