Why do AI models confidently make things up - and can we fix hallucinations?

The answer

AI hallucinations occur when language models generate plausible-sounding but factually incorrect or fabricated information, including fake citations, invented statistics, and fictional events described with authoritative confidence. This happens because LLMs are fundamentally trained to produce statistically likely text, not to verify truth - they model language patterns, not factual databases. A 2023 study found that GPT-4 still hallucinated in approximately 3-5% of responses on factual queries. Mitigation strategies include retrieval-augmented generation (RAG), chain-of-thought prompting, and grounding models in external knowledge bases, but fully eliminating hallucinations remains an unsolved problem.