Hallucination

Executive Definition: Hallucination — A phenomenon where an AI model generates false, nonsensical, or unverified information while presenting it as fact.

Hallucinations occur because Large Language Models generate text based on statistical probabilities rather than a genuine understanding of facts. They are especially common when a model lacks sufficient training data on a topic or is prompted to generate highly specific details.

Frequently Asked Questions

Can hallucinations be completely eliminated?

Not entirely, given the probabilistic nature of current generative models. However, they can be mitigated using techniques like RAG, lower temperature settings, and rigorous human-in-the-loop validation.

How do I detect hallucinations?

By cross-referencing AI outputs against trusted source documents. Automated evaluation pipelines, fact-checking APIs, and model confidence scores can also help flag potential hallucinations.

Are some models more prone to hallucination?

Models tend to hallucinate more when queried on niche topics or recent events outside their training data. Integrating models with external knowledge sources (grounding) reduces hallucination frequency.