Hallucination
A confident but incorrect or fabricated output from a model, stated as if it were fact despite having no basis in the input or reality.
In more detail
A hallucination is the model producing something that sounds right and is wrong. Because a language model generates plausible continuations rather than looking up facts, it can invent citations, numbers, names, or events with full confidence. The fluency is exactly what makes the error dangerous, since the output does not signal its own unreliability.
The most effective mitigations reduce the model's need to rely on memory: retrieve real source material and have the model answer from it, require citations so claims can be checked, and constrain the task. Human review remains the backstop for high-stakes outputs where a confident error would be costly.
Where this shows up at Ceven
Ceven fights hallucination by grounding its research in retrieved sources and returning briefs with citations, so a claim can be traced to where it came from rather than taken on trust. For consequential actions, human-approval gates put a person between a model's output and an irreversible write, and the audit trail preserves what was proposed.