The Fairness Review

The situation

The moderation model is ready for a launch decision. Overall it's accurate — 0.89 on the held-out set. But the fairness review Lydia commissioned found something the headline number hides: it flags isiZulu and Afrikaans slang as abuse roughly four times more often than the equivalent English slang. Separately, the launch plan proposes keeping every flagged post and the full profile of every flagged user for five years. Leo wants it live to clear the moderation backlog. Amara wants the review done properly first.

What you'll learn

• AIF-C01 responsible AI, governance and compliance: reading subgroup (fairness) evaluation results, classifying control gaps across fairness, privacy, security and evaluation, data minimisation and retention, model cards and audit evidence, and defending a launch / no-launch recommendation against competing constraints. A risk register as a lifecycle control.
• Match responsible-use, security and governance needs to relevant AWS controls; distinguish statistical under/overfitting from social bias, and review generative-output risks. The programme practice bank does not establish current vendor exam proportions or readiness.

Who you work with

• 

  Lydia Roe

  Security Engineer

• 

  Amara Okafor

  ML and Data Engineer

• 

  Leo Martins

  Product Manager

Scenes

1. 1.The number that hides the problem
2. 2.Read the subgroups
3. 3.Tighten the controls
4. 4.Which control?
5. 5.Launch or not
6. 6.AWS AI Practitioner practice checkpoint

What you leave with

You reviewed the fictional overall and subgroup metrics, proposed a 90-day retention policy, and removed the simulated analytics grant. The recommendation is a controlled reviewer-assistance pilot with monitored subgroup errors and stop criteria. Metadata changes do not prove deletion; responsible retention review and wider evaluation remain necessary.