Why hallucination is structural rather than a rare bug, prompt injection and jailbreaking as an adversarial security concern, bias in training data, guardrails and their limits, human-in-the-loop design, and the EU AI Act's risk-tiered regulatory approach.
Large language models generate text by predicting the next plausible token, not by looking up verified facts — which is why hallucination is a structural property of the generation mechanism, not a rare bug a future update will quietly fix. This module works through why grounding (Module 9's RAG) reduces but never eliminates that risk, why prompt injection and jailbreaking are a real adversarial security concern unique to LLM-based systems, how bias in training data surfaces in model outputs, and why guardrails, input validation, and output filtering are each a genuinely useful but individually incomplete layer of mitigation.
By the end of this module you should be able to explain when an AI system should recommend an action for human review rather than act autonomously — especially relevant to Module 15's engineering applications and Module 7's tool-calling agents — and describe the EU AI Act's risk-tiered approach as the regulatory landscape's most concrete example, a topic Module 19 (Professional Practice) returns to as an organizational process rather than a one-time technical checklist.