Hallucinations aren't bugs; they're a consequence of how LLMs work. The model is trained to predict plausible next tokens, not to verify truth. When uncertain, it generates something that fits the shape of an answer -- citations to nonexistent papers, API methods that don't exist, legal cases that were never filed.
Mitigations, roughly in order of effectiveness: (1) retrieval / grounding (give the model the facts), (2) citation-forced output (quote sources verbatim), (3) structured output + validation (narrow generation space), (4) self-critique / reflection, (5) LLM-as-judge for fact-check, (6) human review for high-stakes. Stacked, these can reduce hallucination to single-digit rates.
Example Prompt
# Reducing hallucination with citation-forced grounding
SYSTEM PROMPT:
"For each factual claim, quote the exact supporting sentence from the
provided documents and cite it as [n]. If no source supports the
claim, mark it UNVERIFIED and do not include it in the final answer."
USER:
{question + provided docs}When to use it
- Any domain where factual accuracy matters
- User-facing Q&A where confident-but-wrong damages trust more than "I don't know"
- Evaluating which model / prompt combination hallucinates least for your corpus
When NOT to use it
- Creative writing / brainstorming where "confabulation" is the feature not the bug
- Assuming any mitigation eliminates hallucination -- it reduces, never eliminates