Your chatbot just cited a paper that doesn't exist. Here's why.
Hook
Fake Citations
Your chatbot just cited a paper that does not exist. And it sounded completely sure.
- Sounds sure
- Does not exist
Cause
Next-Token Prediction
Here is why. A language model is trained to predict the next token. Not to check facts.
- Trained to predict tokens
- Not trained to verify
Plausible, Not True
It picks words by probability. Plausible is the goal. True is only a side effect.
- Objective: likely text
- Truth is a side effect
Example
Pattern Without Facts
Ask for a source on a niche topic. The model knows what citations look like: author, year, journal.
- Knows the format
- Not the real source
Fluent But Fabricated
So it fills the gaps with a perfect looking pattern. Fluent, confident, and fabricated.
- Gaps get filled
- Confidence stays high
Fix
Fix: Ground It
One fix is grounding. Retrieve real documents first, then answer from them.
- Retrieve real documents
- Answer from them
Cite Or Admit
Now the model can quote a real source, or admit it found nothing.
- Quote the source
- Or say no source
Outro
Next: Hallucination Deep Dive
Next lesson, we go deep on why models hallucinate. Follow to catch it.
- Follow for the lesson