COMPEL Glossary / hallucination
Hallucination
Hallucination is the phenomenon where an AI system, particularly a large language model, generates output that is plausible-sounding and confidently stated but factually incorrect, fabricated, or unsupported by the model's training data or any real-world source.
What this means in practice
Hallucinations range from minor factual errors to completely invented citations, statistics, or events. For organizations deploying generative AI, hallucination is a significant operational and reputational risk because users may trust and act on AI-generated information without verification. Mitigation strategies include grounding through RAG architectures, confidence scoring, human review workflows, and output verification against authoritative sources. In COMPEL, hallucination risk is addressed within the Technology and Governance pillars, with detection and mitigation forming part of the responsible AI infrastructure designed during Model and operationalized during Produce.
Why it matters
Hallucination is a significant operational and reputational risk because users may trust and act on AI-generated information without verification. When AI systems confidently state fabricated facts, cite nonexistent sources, or invent statistics, organizations face liability for decisions made based on false information. The challenge is compounded by hallucinations being grammatically perfect and contextually plausible, making them difficult to detect without verification.
How COMPEL uses it
Hallucination risk is addressed within both the Technology and Governance pillars. During Model, mitigation strategies including RAG grounding, confidence scoring, and human review workflows are designed into AI system architectures. The Produce stage implements detection and mitigation infrastructure. The Evaluate stage tracks hallucination rates as a key Quality KPI, and production monitoring during ongoing operations alerts on hallucination spikes.
Related Terms
Other glossary terms mentioned in this entry's definition and context.