10How should I use it?

Wrong, and sure about it

So far you've seen what AI makes. Now let's look at what to watch out for.

Ask AI something. Watch how it answers.

Press Really? Let's check. Some answers are right, some are wrong.

But AI sounds just as sure when it's wrong. It says I'm sure! either way. The tone doesn't change at all.

Look at the bars. They're the same bars from chapter 6. AI picks the most likely-sounding answer. How likely something is, is called .

Likely and right are not the same thing. Always check the things that matter.

A little deeper

For grown-ups, and for children who want more.

An AI inventing something plausible is called hallucination. It is not a malfunction — it is a property of how language models work.

A language model does not look facts up. As you saw in chapter 6, it continues with whatever is most likely to come next. Likely and correct overlap often enough that it is usually right; where they diverge, it is confidently wrong. Its tone carries no information about its accuracy.

The short I-do-not-know bar has the same cause. Human writing contains answers, not admissions of ignorance, so that phrase carries low probability. Recent models saying they are not certain did not happen on its own — people trained that in deliberately.

The six questions here are the one thing on this site written in advance. Which way it goes wrong has to be chosen for the lesson to work, and this is the only exception among the fourteen chapters.