// AI
Hallucination
When an LLM generates plausible-sounding but factually wrong output, the fundamental risk of generative AI.
In depth
Hallucination rates dropped materially from 2023 to 2026 via RAG, better base models and inference-time verification, but never to zero. Any customer-facing AI motion needs guardrails, source citation and human review of consequential outputs.
Related terms
RAG
Retrieval-Augmented Generation, grounding an LLM's response in retrieved documents to reduce hallucination.
Guardrails
The layer of policies, filters and validators that constrain LLM outputs to safe, on-brand, on-topic responses.
LLM
Large Language Model, the transformer-based foundation model behind ChatGPT, Claude, Gemini and every 2024+ AI product.