Confidence Score
A confidence score is a number with which a computer program indicates how certain it considers its own result to be. It is usually given between 0 and 1 or as a percentage, and it often determines whether a human needs to review the result again.
When a computer program classifies an image or reads out text, it rarely delivers just a bare result. Almost always, it also outputs an additional number that says: this is how certain I am. This number is called a confidence score. It typically lies between 0 and 1, often also written as a percentage. A program that classifies a photo as a cat and outputs 0.97 alongside it is reporting a very high level of certainty. At 0.41, on the other hand, it is barely more than a guess.
The distinction between certainty and correctness is important. A high value does not mean the result is correct. It only means that the program, by its own standards, has hardly any doubt. These two things often coincide, but not always.
The number that the handover to humans depends on
Without this value, every result would be worth the same. A certain result and a guessed one would look identical to the user. The score makes the difference visible and thus usable. Companies therefore set a threshold, a so-called cutoff value.
An example from practice: a bank has invoices read out automatically. Everything above 0.95 is booked without further inquiry. Everything below lands on an employee’s desk. This way, the machine handles the simple cases, and humans take care of the tricky ones. That is exactly what saves costs without driving up the error rate.
Choosing the threshold is a trade-off. If it is set high, few errors slip through, but a lot of work remains with humans. If it is set low, more runs automatically, and more errors go unnoticed. In medicine or credit decisions, therefore, different thresholds are chosen than in a photo app.
How calculation results become a percentage
A model, meaning a trained program, does not simply pick one answer. It calculates a numerical value for every possible answer. For an animal picture, these might be cat, dog, and fox. These raw values are then converted so that together they add up to exactly 1. The result then looks like a probability: 0.97 for cat, 0.02 for dog, 0.01 for fox.
The highest of these values is output as the confidence score. You can picture it like the distribution of votes in a ballot. If almost everyone agrees, the value is high. If the votes are spread evenly, it is low.
A common misconception is to read this number directly as the probability of being correct. Many models are overconfident: they output 0.99 and are still regularly wrong. Experts call this poor calibration. A well-calibrated model is actually correct in around 90 percent of all cases with a score of 0.90. This is checked by comparing many predictions against the actual outcomes.
Where the number shows up in everyday life
The value becomes visible, for example, in translation and subtitling programs that display uncertain passages in gray or with a marker. Speech recognition on mobile phones also works with it. In automated facial recognition at the airport, the score decides whether the gate opens or an officer steps in.
In business news, the term often comes up in connection with liability and regulation. Authorities require that, for high-risk applications, uncertain results be reviewed by humans. The confidence score is then the technical basis for this rule.
With chatbots, the situation is different. These systems usually do not display their internal value at all, and they phrase even wrong answers confidently. A phrase like “I’m pretty sure” is not a confidence score, just text. Anyone using such a system professionally should be aware of this difference.