Calibration
Making a model's stated confidence match how often it's right
- Calibration means lining up the confidence a model states with how often it's actually right.
- The method is fixed: gather every case where it stated the same number, then count how many of those actually turned out right.
- Being right a lot and stating things accurately are two different things. Some models score well while consistently overstating their confidence.
- Today's models tend to lean toward overconfidence. The number they state runs higher than reality.
- The fix usually isn't rebuilding the model. It's shifting the scale from the outside.
Contents
1The analogy
The thermometer at home says it's 25 degrees today. Should you trust that number? The way to check isn't to look at one day. Gather up every day this thermometer said 25 over several months, and separately measure what the real temperature was on each of those days.
Say the real average on those days turns out to be 30. This thermometer has a habit of reading low. If it comes out close to 25 instead, this is a thermometer that tells you what it actually is. Whether a thermometer is trustworthy isn't something one day's number can tell you, only gathering the days it said the same number can.
Calibration is finding that habit and shifting the scale to correct it. The same thing gets done to the probabilities an AI states. Gather up every case where it said 80, and count how many of those really were the eighty out of a hundred it claimed.
2In detail
Gather up the cases that stated the same number
A single case tells you nothing. If a model said 80 and got that one wrong, that doesn't make 80 a bad number. It's a number that's supposed to be wrong two times out of ten anyway.
So many cases get pooled together. Group a model's stated confidences into similar-value bundles, and count the actual hit rate inside each bundle. In the bundle where it said 60, six out of ten should be right; in the bundle where it said 90, nine out of ten should be right, for the scale to be sound.
Comparing the stated value against the actual hit rate, bundle by bundle, shows exactly where things go off. Some models are fine at the low end and only inflate at the high end; others inflate evenly across the board. Average the size of the gap across all the bundles, and it comes down to a single number too.
The habit of leaning toward overconfidence
Today's large models tend to state numbers higher than reality across the board. Gather up everything they said 90 on, and only seven or eight out of ten turn out right. Training pushes toward nailing the correct answer as firmly as possible, so the tone at the end of a statement keeps getting harder.
Models that answer in words show this bias even more visibly, because they say wrong things in the exact same tone as right ones. When a confident-sounding tone isn't tied to the actual odds of being right, whoever's listening has no way to know where to stop and double-check.
Leaning the opposite way happens too. A model that's overly cautious and always states things low makes even its correct answers look unconvincing, so people wave the answer off. Either direction, the number can't be taken at face value.
Shifting the scale
Fixing this usually happens outside the model itself. Instead of buying a new thermometer, the dial gets shifted. Using data that was never touched during training, the stated values get compared against the actual hit rate, and one transformation gets attached that undoes the gap.
The simplest version softens or firms up everything the model outputs with a single knob, dividing the values right before they become probabilities by one adjustment number, so the model states things less firmly overall. When the gap differs from one range to another, remapping the scale range by range is used instead.
What matters is what data this gets done with. Calibrate using the same data the model trained on, and it looks better than it really is, because those are memorized answers. Calibration needs its own untouched data too.
Being right a lot and stating things accurately are different
Say two models both get nine out of ten right. One always says 99, the other says something around 90. Their skill at getting the right answer is identical, but only the second model has a sound scale. Trust the first model's numbers and skip human review, and things slip through.
The reverse happens too. A model can have a sound scale and terrible skill. A model that says fifty-fifty no matter what it's asked really will be right about half the time, so its scale is technically sound, and it's still useless. Skill and scale always need checking separately.
Calibration matters most in specific places: wherever a probability decides whether a human steps in, wherever several models' judgments get combined as probabilities, wherever a probability gets weighed against a value to balance gain and loss. In spots like these, even a little inflation throws the whole decision off.
3More precisely
Calibration is making a model's stated probabilities line up with actual frequencies. A common approach bins probabilities into ranges, measures the gap between the average stated value and the actual hit rate in each bin, and weight-averages that gap by bin size to get a single measure of miscalibration, the smaller it is, the better the scale. The simplest fix divides the last-stage value before it becomes a probability by a single number to soften everything at once; remapping bin by bin with a monotonic lookup table is another common approach.
The analogy breaks in one place. A thermometer always has a real temperature to check against, but a single probability from a model has no such thing to check against on its own. Saying there's an 80 percent chance an answer is correct can't be verified for that one case. It only means something once many cases are pooled. Calibration is always a property of a group, never of one single case.
A scale that's sound overall isn't sound everywhere either. Split the data into subgroups and one subgroup can run high while another runs low, canceling out to look fine overall. Checking each subgroup separately is needed alongside the overall check.
4Try it yourself
5Common misconceptions
It's easy to think a higher stated probability means more trustworthy, but actually without a sound scale, a stated 90 doesn't mean nine out of ten will actually be right.
It's easy to assume a model that's right a lot also has a sound scale, but actually skill at getting things right and accuracy in stating confidence move independently, so each needs measuring on its own.
It's easy to think calibrating means retraining the model, but actually shifting the scale from the outside, using untouched data, solves it in most cases without retraining at all.
7One-line summary
In shortCalibration means gathering up every case where a model said 80 and checking whether eighty out of a hundred really turned out right, then shifting the scale to close the gap.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02