Responsible AI, Safety, and Governance

This section covers the controls that make AI behavior testable, reviewable, and governable: factuality, privacy, security, fairness, oversight, compliance, audit evidence, and release governance. Responsible AI work is not a separate review at the end of a project; it maps risks to controls throughout the system lifecycle.

Knowledge map

The section moves from behavior risks (factuality, security, privacy) through fairness and explanation to the governance controls that gate releases.

flowchart TD
  Safety[Factuality and Hallucinations] --> Sec[Security and Prompt Injection]
  Sec --> Priv[Privacy and PII]
  Priv --> Fair[Fairness and Explainability]
  Fair --> Audit[Auditability and Risk Classification]
  Audit --> Gov[Human Oversight, Compliance, Change Governance]

Reading path

Read the behavior-safety controls first, then privacy and security, fairness and explanation, and finally governance.

1. Factual Correctness: checking claims against evidence.
2. Hallucinations: unsupported generated content and its causes.
3. Error Taxonomies: structured failure labels that make evaluation actionable.
4. Adversarial Evaluation: probing for unsafe behavior.
5. Prompt Injection: untrusted input steering the model.
6. Security: protecting the system and its tool use.
7. Privacy: protecting user and training data.
8. PII Leakage: detecting and preventing personal-data exposure.
9. Policy Enforcement: applying rules at runtime.
10. Fairness: comparable performance across groups.
11. Explainability: making decisions inspectable.
12. Auditability: keeping evidence of what happened and why.
13. Risk Classification: sizing the risk of a system or change.
14. Human Oversight: keeping people in control of consequential actions.
15. Compliance: meeting legal and regulatory obligations.
16. Governance of Model and Knowledge Base Changes: traceable approval for updates.

Connections

• Generative AI provides the guardrails and behaviors these controls constrain.
• ML Engineering and MLOps supplies the lifecycle gates, and Experimentation and Evaluation the evidence.

Nav

Learning path — Generative AI systems

← RAG Evaluation