A confidence score is not a probability
Many models emit a number between 0 and 1 that looks like a probability. Unless it has been calibrated, it isn’t one.
Engineering Desk · Engineering notes
General edition. , or use the switcher above.
A score of 0.9 sounds like “right nine times out of ten”. That’s only true if someone has checked.
The point
Calibration is the check: gather the cases a model scored around 0.9, and see how many it actually got right. If the answer is seven in ten, the score is overconfident, and any rule built on “above 0.9 we trust it” is built on sand.
Research on modern neural networks has found that they can be poorly calibrated out of the box (Guo et al., 2017). The remedy is measurement on your own data, not a better threshold.
Practical test: before using a confidence threshold in a workflow, plot predicted confidence against observed accuracy on held-out examples. If the line isn’t close to diagonal, treat the score as a ranking signal only.
Sources
- On Calibration of Modern Neural Networks · arXiv (Guo et al., 2017) · accessed 1 Oct 2026
Your perspective
A better read starts with knowing what matters to you.
Three quick questions. A publication that adapts to your perspective.
Keep reading
All stories →Why promising AI pilots never become real products
A pilot proves something is possible. A product has to keep working. The gap between the two is predictable, and it can be closed before the money is spent.
What does it really take to bring AI into production?
Make four decisions about a hypothetical AI project (scope, data, oversight, rollout) and see what you’d be trading away. A simulation, clearly labelled.
The personalization paradox
When does tailoring an experience to someone make it better, and when does it become intrusive? An interactive look at four rungs of personalization.