Step by Step
F
Factual errors — wrong dates, names, details
LLMs can state incorrect factual details — wrong dates, names, or other specifics — with the same fluent confidence as correct information.
Example: an LLM confidently stating an incorrect birth year for a historical figure, phrased with the same fluency as a correct fact would be.
A
Attribution errors — fabricated citations and quotes
LLMs can fabricate citations, sources, or quotes that sound completely plausible but don't actually exist.
Example: a lawyer who cited fake, ChatGPT-generated case citations to a federal court, receiving sanctions when it was discovered the cited cases didn't actually exist.
R
Reasoning errors — logical mistakes
LLMs can make logical or reasoning mistakes, arriving at an incorrect conclusion through a flawed chain of reasoning, even while sounding confident throughout.
Example: an LLM making a basic logical error partway through a multi-step reasoning problem, yet presenting its final (incorrect) answer with full confidence.
T
Temporal errors — outdated information presented as current
LLMs can present outdated information as if it were current, since their knowledge is fundamentally limited to their training data cutoff.
Example: an LLM describing an organization's leadership using information that was accurate at training time but has since changed.
Applied Walkthrough
1
A lawyer uses ChatGPT to help draft a legal brief, and the tool generates several case citations that sound entirely plausible and properly formatted.
2
The lawyer submits the brief to a federal court without independently verifying these citations.
3
It's later discovered that these citations were entirely fabricated — the cases simply don't exist — resulting in sanctions against the lawyer.
4
This real case illustrates why hallucination is described as fundamental to how LLMs work, not a rare bug: the model generated plausible-sounding, properly formatted text without any internal fact-checking mechanism to verify the citations were real.
Exam Application
Exams test whether you can name the four types of hallucination (factual, attribution, reasoning, temporal) and whether you understand why hallucination cannot be fully eliminated — it's fundamental to how LLMs generate text, not a bug that can simply be patched out. Also expect questions on mitigation strategies: RAG, citation requirements, structured verification, temperature reduction, and human review for high-stakes output.
⚠ Common Trap
The most common trap is treating hallucination as an occasional bug that will eventually be fixed with a better model. Hallucination is fundamental to how LLMs work — they predict statistically plausible next tokens, not verified facts — meaning it can be mitigated (through RAG, verification, human review) but not fully eliminated.
✓ Quick Self-Check
1. What are the four types of hallucination described in this lesson?
Factual errors, attribution errors, reasoning errors, and temporal errors.
Tap to reveal / hide
2. What happened in the real legal case referenced in this lesson?
A lawyer cited fake, ChatGPT-generated case citations to a federal court and received sanctions when the fabrication was discovered.
Tap to reveal / hide
3. Can hallucination be fully eliminated from LLMs?
No — it's fundamental to how LLMs generate text (predicting plausible tokens, not verified facts), so it can only be mitigated, not fully eliminated.
Tap to reveal / hide
4. Name one mitigation strategy for reducing hallucination.
RAG (grounding in real documents), citation requirements, structured verification, temperature reduction, or human review for high-stakes output (any one).
Tap to reveal / hide
5. Why can't LLMs simply fact-check themselves before responding?
Because they have no internal fact-checker — they generate statistically plausible text continuations, not verified, retrieved facts.
Tap to reveal / hide