.svg)
What are LLM Hallucinations?
LLM hallucinations are outputs from large language models that are fluent and plausible but factually wrong, unverifiable, or not grounded in the provided context-fabricated citations, invented statistics, or incorrect procedural steps.
Hallucinations arise from next-token prediction under uncertainty, gaps or biases in training data, and weak grounding to trusted sources. In high-stakes domains (law, medicine, finance), they can cause real-world harm even when the text sounds confident.
Related risks
- Misinformation presented as fact
- Amplification of training-data bias
- Ungrounded answers in RAG when retrieval fails
Practical mitigations
- Require citations and refuse when evidence is missing
- Evaluate groundedness against retrieved sources
- Monitor production for confident-but-wrong patterns
Related Giskard articles
- Practical guide to LLM hallucinations
- AI hallucinations in a French court
- When AI financial advice goes wrong
Detect hallucinations with Giskard evals
Use Giskard hallucination and misinformation checks plus groundedness metrics for RAG to catch fluent fabrications before users rely on them. Learn more.
Further reading
Authoritative reference: OWASP LLM09: Misinformation.
