TL;DR - An LLM predicts the next word from patterns it learned. It doesn't "look up" facts. That single idea explains hallucinations, inconsistency, and why context is everything.
Why it matters
If you think ChatGPT is "searching for the answer," you'll trust it like a search engine and get burned. Once you know it's predicting likely text, its quirks stop being mysterious and you learn to steer it.
How it works (no math)
The model read a huge amount of text and learned which words tend to follow which. When you prompt it, it generates one likely word at a time. That's it. It's astonishingly good at language because language is full of patterns - and unreliable on facts because a plausible-sounding word isn't always a true one.
Worked example
Prompt: "Give me a quote about persistence from Albert Einstein."
Risk: the model produces a clean, inspiring quote... that Einstein never said. It sounds right, so it generated it. That's a hallucination - confident output not grounded in fact.
Safer prompt: "If you're not sure the quote is real, say so and give me a verified one with the source."
Steal this - the trust dial
Predicting words is reliable for: phrasing, structure, rewriting, brainstorming.
Predicting words is risky for: names, numbers, quotes, citations, recent events.
Rule: the more factual/verifiable the claim, the harder you check.Common mistakes
- Treating it as a database. It has no live lookup unless the tool adds web access.
- Expecting identical answers. Ask twice, get two phrasings - it's sampling, not retrieving.
- Trusting specifics. Exact stats and quotes are where hallucinations hide.
Good to know
Some tools now do search the web (ChatGPT with browsing, Perplexity, Gemini) and cite sources - much safer for facts. Plain chat without browsing is working from memory of its training data, which has a cutoff date. Check whether the tool you're using can actually look things up.