Glossary
Hallucination.
When a language model generates content that sounds plausible but isn't true. The model isn't lying. It has no concept of truth. It's pattern-matching what an answer to your question would probably look like, and "probable" doesn't equal "accurate."
The dangerous failure mode isn't a wrong answer. It's a wrong answer phrased with complete confidence. You can't tell from the model's tone whether it knows or whether it's guessing.
What this looks like in practice
Why models hallucinate at all.
Foundation models are trained to predict the most likely next token. They're not trained to know what's true. When asked about something they don't actually know, they generate text that looks like a plausible answer for that kind of question. The text might be right, might be wrong, might be partially right. Same confident tone either way.
Where hallucination shows up the most.
Specific facts (dates, numbers, names) the model wasn't trained on. Topics where the training data was thin or contradictory. Industry-specific knowledge. Company-specific knowledge. Recent events. Anything where "what's probable" diverges from "what's true." The model has no way to know it doesn't know.
What actually reduces it.
Retrieval helps. Give the model the actual source material so it's working from facts, not training-data probability. Grounding helps more. Force the model to cite its source, then verify the citation. Refusal helps most. Let the model say "I don't know" instead of generating something plausible. Most products implement the first, sometimes the second, almost never the third.
The opposite of hallucination is refusal.
The safest AI isn't the one that always answers. It's the one that knows when not to. Lumen tells you when the source for your question doesn't exist, instead of inventing one that sounds right.
Bring your knowledge.Simply ask.
CASA Certified · SOC 2 Infrastructure · GDPR-Aligned · Enterprise Ready