Back to full roadmap
topicfoundation
Hallucination Anatomy
Why models 'make things up' — root causes and which prompt patterns trigger it.
2 hours3 resources1 prereqs
Hallucination = the model produces something it doesn't know with confidence. Root causes:
- Knowledge gap — events after training cutoff, niche domains
- Ambiguous prompt — model feels forced to 'guess'
- Long output pressure — when requested detail exceeds real knowledge
- Bad few-shot — wrong examples are imitated
- Contradictory instructions — model forced to pick most-likely interpretation
Mitigation:
- Add "If you don't know, say I don't know" to system
- Ground with RAG (external sources)
- Require citation-style answers + verification step
- Low temperature for critical tasks
- LLM-as-judge for post-hoc verification
Why this matters
Hallucination is the #1 enemy of production AI; you can't defend against what you don't understand.
What you'll gain
You can intuitively estimate hallucination risk and apply mitigation layers in the right order.