What are AI 'hallucinations' and can we fix LLMs so that they don't happen?
LLM hallucinations stem from their probabilistic next-token nature and cannot be fully eliminated, only mitigated.
“This minimizes hallucination, but it doesn't eliminate it.”
A Microsoft Build session explains that LLM hallucinations arise from the stochastic, next-token-prediction nature of models that have no inherent sense of accuracy. It notes that RAG reduces but cannot eliminate hallucinations, since context can still be misprocessed, irrelevant, or incorrect. The framing is educational and reinforces a known limitation rather than announcing anything new.