Hallucination
A hallucination is content an AI model states with full confidence but that is not true: invented facts, fabricated citations, wrong numbers, delivered in exactly the same tone as its correct answers.
No jargon, no math
In plain words, what is it?
A language model is a pattern machine trained to produce the most plausible continuation of a text, not to verify it. When it writes, it predicts word by word what should come next, and plausibility is its only compass. If your question sits in a dense area of its training data, the pattern is solid and the answer is right. If the question is niche, recent or specific (your part numbers, last Tuesday's fault log), the model still produces a fluent answer, but that fluent gap-filler can be fiction wearing a confident tone. Nothing inside the generation loop checks the answer against reality; that check must come from outside: retrieval that grounds the answer in your documents, guardrails that compare claims to sources, or a human who verifies. Readers mistake fluency for accuracy, which is why hallucination is the most expensive failure mode in business AI.
- 01
Your question. You ask something the model only half-knows.
- 02
Generation. It writes the most plausible continuation, word by word.
- 03
The fork. Is the claim grounded in real sources? If yes, you get a reliable answer.
- 04
The invention. If no, the model fills the gap with something that sounds right: the hallucination.
- 05
The net. Guardrails and retrieval check the claims; caught ones get flagged, missed ones reach you as fact.
Terms that live nearby.
Retrieval-Augmented Generation
RAG is an architecture where the AI retrieves relevant documents from your knowledge base before writing an answer, so it quotes your data instead of guessing from memory.
ReadGuardrails
Guardrails are the checks and limits wrapped around an AI model so its behaviour stays inside your rules: they screen what goes in, constrain what the model can do, and verify what comes out before it reaches a person or a system.
ReadFine-tuning
Fine-tuning is additional training that teaches a general model your specific patterns: you show it hundreds or thousands of examples of your tasks, your tone and your formats, and it adjusts its internal weights until that behaviour becomes its default.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.