Hallucination Types & Causes

Hallucination is when a model produces content that contradicts facts, its context, or logic. Three types: factual (invented facts — wrong dates, people, numbers), logical (broken reasoning chains — contradictions, wrong steps), and contextual (conflicts with the user input, retrieved documents, or system instructions, including stale post-cutoff knowledge).

credibility checks in high-stakes domains (healthcare/finance/law); hallucination-rate and attribution testing before shipping RAG, agent, or summarization systems; a favorite interview route into why models lie and how training objectives and data distributions cause it.

hallucination is the biggest obstacle to trustworthy LLM deployment. The root causes live in the training objective $\max_\theta \sum_t \log P_\theta(y_t \mid y_{<t}, x)$ — maximizing conditional likelihood rather than truth — plus noisy training data, stochastic sampling, and knowledge cutoff/forgetting. Classifying the three types makes mitigation and detection tractable.

Name the three hallucination types and give one concrete example of each.

Explain hallucination from the training objective: why doesn't the MLE objective max Σ log P(y_t|y_<t,x) guarantee factual correctness?

Which type does cutoff-induced hallucination belong to, and how do you mitigate it via interface design?

How do sampling parameters affect hallucination rate? Does greedy decoding fully eliminate it?

Three hallucination evaluation lenses — fact-based, context-based, logic-based: what tests does each map to?