Hallucination
Still wrong.
Ground it.↘
What exactly is a hallucination, and what makes it
dangerous to a reader?
A hallucination is output that's stated with total confidence but is not true. What makes it dangerous is that it arrives in the same fluent, assured tone as a correct answer, complete with plausible names, dates, or citations, so the reader gets no warning.
A bluff reads identically to a fact, so you can't catch it by feel. That's why you need grounding, citations, and automatic verification instead of trusting the output's confidence.
eliminate it?”
→ No. It reduces the rate. Verify
important claims and keep a
check in the loop.
Why does a model trained to predict plausible next tokens
produce confident falsehoods when it lacks knowledge?
Its only training objective is to predict the next token that fits the pattern. Nothing in training checks a claim against reality.
Fluency and accuracy are separate skills, and the objective only optimizes fluency.
So when the model hits something it doesn't know, it emits the most statistically natural continuation, which often reads like a real answer with invented specifics. There's no built-in uncertainty signal, so the confident guess comes out in the exact same voice as a fact.
Which mitigations reduce hallucination, and why is grounding
with RAG the biggest lever?
Four mitigations work together. Ground with RAG, so the model fetches real documents and answers only from what it retrieved.
Allow "I don't know" as a valid answer. Lower temperature for factual tasks.
And ask for citations so every claim is traceable.
Grounding is the biggest lever because it removes the root cause: it changes where the facts come from. The model reads real documents instead of answering from memory about things it was never given.
The other steps tighten behaviour around that.
"I don't know"
temperature
citations
documents
from them
root cause
This is a small sample.
600+ interview questions live inside Skillumen.
Every concept card carries its own interview questions, graded as you answer.
Free to start.