How a model “answers”: token, context, hallucination
A model splits text into tokens (words or word pieces) and builds its answer piece by piece, each time choosing the most fitting next piece. That is why answers read smoothly, but smooth does not mean correct.
The context window is the amount of text the model can see at once. If a conversation grows too long, the beginning may drop out of attention; then it helps to repeat an important condition.
A hallucination is when a model confidently gives incorrect or invented information. The risk is highest for rare facts, exact dates, numbers and source links.
Remember
- A smooth answer is not necessarily a correct one.
- In a long conversation, repeat important conditions.
- Check numbers, dates, links and quotations.
Exercise
A model gave you a link to a scientific paper that does not exist. What is this called and what should you do?
Answer and explanation
- A hallucination; open the link and check it yourself. (correct answer)Models can invent believable but non-existent sources.
- A new discovery; you can use it.An unchecked link is not a discovery.
- A token error; refreshing the page is enough.The problem is not technical; it is data the model made up.
Practical task
Ask a chat tool a question on a topic you know well and look for mistakes in the answer.
- I asked a question on a familiar topic
- I compared every claim in the answer with what I know
Glossary
Interactive mode: search, progress and Python exercises.