Hallucination

A hallucination is an AI output that reads as confident and coherent but contains factual errors, fabricated information, or invented references. The danger is not randomness: it is plausibility. Hallucinated text often passes casual scrutiny because it follows the statistical patterns of truthful prose.

Hallucination is a symptom of how LLMs work: they predict likely next tokens, not true ones. When the training data is sparse or contradictory on a topic, the model interpolates, and the result can be smoothly wrong. This is why the fluency trap is so dangerous: fluency does not imply accuracy.

In Coscientist, hallucination risk is managed through the epistemic protocol layer: traceability requires claims to connect to evidence spans, rebuttal-first search stress-tests before acceptance, and the Multi-AI Consensus Protocol uses model disagreement as a signal for closer inspection.