AI Ethics, Safety & Compliance

Why Models Hallucinate

Models are not databases — they are pattern completion engines. They predict what SOUNDS right, not what IS right.

Measuring Hallucinations

Bias Detection

Mitigation Strategies

1. →Prompt engineering — Add "Consider diverse perspectives" or "Avoid gender assumptions" to system prompts.
2. →RAG grounding — Constrain responses to verified, curated sources.
3. →Output filters — Post-process outputs to detect and flag potential hallucinations.
4. →Human review — For high-stakes content, always have a human verify before publishing.
5. →Confidence thresholds — Only surface model outputs when confidence exceeds a set threshold.

Test Your Understanding

Next Steps

You can now detect hallucinations and biases. In the next workshop, you will go on the offensive: red-teaming AI systems to proactively find and fix vulnerabilities.

---

Continue to the workshop: AI Red Teaming Charter to learn adversarial testing.

---

What is a Red Team Charter?

A red team charter is a formal document that defines:

• →Scope: What system are we testing? What is in-bounds vs out-of-bounds?
• →Objectives: What types of failures are we looking for?
• →Methods: What attack techniques are we authorized to use?
• →Reporting: How do we document and escalate findings?

Attack Categories

Mitigation Strategies

Test Your Understanding

Next Steps

You can now systematically find and fix AI vulnerabilities. In the next module, you will master context engineering — the advanced techniques that push AI performance to its limits.

---

Continue to Context Engineering: The Four Pillars to learn advanced prompting architecture.