Context windows and hallucinations

Why models forget, and why they sometimes sound sure and are wrong.

The context window

A model can only look at a limited amount of text at once: its context window, measured in tokens. Your messages, any documents you paste, and its own replies all share that space. Past the limit, older text has to be dropped or summarised.

Models don't remember you between separate chats on their own. Apps that seem to remember are saving notes and feeding them back in as part of the prompt.

Hallucinations

The model's job is to produce plausible text, not to look facts up. So when it doesn't know, it can still write something confident-sounding that is simply wrong. These are called hallucinations.

• Paste the source text and ask it to answer only from that text.
• Ask it to say "I don't know" when unsure.
• Check names, numbers, dates and quotes against a real source.
• Be extra careful for medical, legal and financial questions.

Quick check

Which habit reduces made-up answers the most?

1. Asking louder
2. Giving it the source text and asking it to stick to it
3. Using more emojis
4. Asking the same question ten times

Answer: Giving it the source text and asking it to stick to it. Grounding the model in text you provide gives it real material to work from instead of guessing.