Skip to content

4. Confidence Scores and Uncertainty

The Preview panel doesn't just give you an answer — it gives you a confidence score for every class, and they always add up to 100%.

Reading confidence correctly

Result What it means
One class at 98%, others near 0% The model is very sure — and probably right, if it's been tested well
Two classes close together, like 55% and 45% The model is genuinely unsure — this input is ambiguous to it
A class gets a non-trivial score for something totally unrelated The model has no concept of "none of the above" — it must pick from the classes it knows

The "nothing" problem

Teachable Machine always forces a prediction among your defined classes, even for an input that matches none of them. Show your "water bottle vs. coffee mug" model a photo of a dog, and it will still confidently assign it to one of the two — because it has no third option. If "could be neither" matters for your use case, you have to build that in yourself, by adding an explicit extra class (something like "none of the above," trained on a variety of unrelated images).

Try it

Show your Lesson 2 model a test image that's neither of its two classes (a third, unrelated object). Record what it predicted and its confidence, then explain why it couldn't say "I don't know."

Tip

A model that's wrong but 60% confident and a model that's wrong but 99% confident failed in very different ways — the second is far more dangerous to trust, because nothing about the number itself warns you it's wrong.

Next: Exporting and Using Your Model →