Decoder. plain-English AI glossary

Hallucination

● Core

When the model states something false with total confidence — the failure mode that never fully goes away.

Think of it like

A friend who never says “I don’t know” and just fills the gap with a plausible-sounding story.

Example

Ask for a citation and get a real-looking paper, by real-looking authors, that simply does not exist.

How it actually works

It’s not lying — the model predicts likely text, and “likely” is not “true.” Grounding it in real documents and asking it to cite reduces it, but confidence and correctness are separate dials. A model can be dead wrong and completely sure.

For product teams

Confident wrong answers — the core reliability risk; mitigate with grounding, citations, and evals.

For engineers

Plausible-but-false generations from likelihood-based decoding untethered from ground truth.

RAG — What RAG tries to ground away.
Evals — What evals exist to catch.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome