L

Learn AI

Track progress · learn offline

Open

Lesson 2 of 6

Teaching it to say 'I don't know'

5 min read

Here's a strange fact about a chatbot: left to itself, it will almost always produce an answer. Not because it knows one — because producing text is the only thing it does. Silence isn't in its nature.

It has no built-in 'I'm not sure'

A raw model has no gauge that lights up when it's out of its depth. Every question gets the same treatment: reach for the next plausible word. So a question it can't possibly answer — your private life, next week's news, a made-up name — gets a fluent reply anyway, delivered with the same calm as a fact it knows cold.

The problem isn't that it's dishonest. It's that, by default, guessing and knowing feel identical from the inside — there's no separate 'do I actually know this?' step.

It can be taught where its edge is

Modern assistants are trained on examples where the right move is to decline — to say 'I can't verify that' instead of inventing. This learned refusal is a skill, not a limitation: a good model builds a rough sense of its own boundary and, near the edge, chooses honesty over a confident guess.

A well-placed 'I don't know' is one of the most useful things an AI can say. It's the difference between a tool that guesses and one you can trust.

Under the hood, this is a balancing act. Train refusal too hard and the model turns timid, ducking questions it could answer; too little and it bluffs. Tuning that line is an active area of work.

What to take away

You ask an AI what a specific private company earned last quarter — a figure it has no access to. The most trustworthy response is:

Continue in the app

Take the whole When AI Gets It Wrong course — tracked

Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.