Bias & fairness
AI systems learn from data, and biased data → biased decisions — which, in high-stakes uses (hiring, lending, healthcare), is both harmful and often illegal. Engineer against it:
- Measure it — evaluate outcomes across protected groups (a fairness eval, like the Evaluation course but sliced by demographic). You can't manage bias you don't measure.
- Sources of bias — unrepresentative training data, historical bias encoded in labels, and proxy features (a zip code standing in for race). Audit data and features.
- Fairness metrics — demographic parity, equal opportunity, etc.; they can conflict, so choosing one is a decision, not just math.
- Bias audits are increasingly mandatory (NYC Local Law 144 for hiring tools; EU AI Act data-governance duties).
Transparency & explainability
- Disclose AI use (EU AI Act limited-risk transparency): tell users they're interacting with AI / that content is AI-generated.
- Explainability — for consequential decisions, be able to give the basis (the factors, the retrieved evidence, the audit trail). "The model said so" isn't acceptable for a loan denial.
- Model/system documentation — model cards, data provenance, known limitations.
Hallucination & reliability as a governance issue
Confident wrong answers (misinformation, OWASP LLM09) become a governance problem when people rely on them for decisions. In regulated/high-stakes domains, control it: grounding + abstention (RAG/Prompt courses), human review, and clear disclaimers about limits.
Building the practice
Responsible AI is ongoing, not a launch gate:
- An AI inventory (what AI you run, its risk tier, its owner).
- Pre-deployment review (risk, bias, privacy, security) for new/changed AI.
- Continuous monitoring (bias drift, quality, incidents) — governance in production.
- Clear accountability and an incident process.