Where do AI's answers actually come from?
Think of your phone's autocomplete, scaled up enormously.
In a 2025 survey of about 2,300 American adults, the most popular answer was that it looks up the exact answer in a database. 45% of people picked that.
Only 28% picked the right answer.
Your phone keyboard.
Type "I'll be there in" and it offers 5, a, an hour.
It has no idea where you are. It has seen a lot of messages that start like yours, and it guesses what usually comes next.
AI was trained on an enormous pile of text: websites, books, code, forums. From all that, it learned which words tend to follow which.
When you ask it something, it writes the answer a small piece at a time, picking a likely next piece each time.
Large language model (LLM): the technical name for this kind of AI. It's "large" because of the amount of text it learned from and the size of the thing that stores what it learned.
AI goes round this loop once for every word in the answer.
The capital of France is …
| Next word | How likely |
|---|---|
| Paris | 97% |
| a | 1% |
| located | 1% |
| … | … |
AI picks Paris, adds it to the sentence, and does the whole thing again for the next word.
For common facts, the most likely next word is also the true one. The internet says "Paris" an awful lot.
AI gets shakier the rarer the fact gets. A niche regulation, a small company's revenue, a paper from last year: there's less text to learn from, so "likely" and "true" stop lining up.
Ask it for a source and it'll write something that looks exactly like a source.
Smith, J. & Patel, R. (2019). Remote work and the erosion of team trust. Journal of Applied Psychology, 104(6), 812–829.
The journal is real. The names are plausible. The page numbers look right.
The paper might not exist. It wrote what a citation on that topic usually looks like.
There's no warning light when it's guessing.
A made-up fact comes out in the same confident tone as a real one, because both come from the same process. It's always doing the same thing: writing the likely next word.
Some apps can search the web first and then answer from the pages they found. That's a separate step the app adds on top.
If you don't see it searching (usually a "Searching…" line or a list of sources), assume it's answering from what it learned in training.
Check the specifics. Names, numbers, dates, quotes and links are where it fills in the blanks. The general explanation is usually fine.
If a fact matters, make it search. Ask it to look it up and show its sources, then open one.