← Research
Research

Calibration

Accuracy is how often a system is right. Calibration is whether it knows.

Last reviewed: 4 September 2026

The definition, the standard measure, what the numbers actually are for language models asked to state confidence, and why a calibrated model does not produce a calibrated reader.

Question this page answersAll 811 questions this research covers

A system is calibrated when the confidence it states matches how often it turns out to be right. Take every case where it said it was 80 per cent sure; if it was correct in about 80 per cent of them, it is well calibrated. This is a different property from accuracy, and the two come apart in both directions. A system can be accurate and badly calibrated, by being right most of the time while claiming a certainty it has not earned. It can be inaccurate and well calibrated, by being wrong often and saying so. Only one of those two is safe to rely on, and the accurate one is not it.

The answer, in one line

A system is calibrated when its stated confidence matches how often it turns out to be right. If it says 80 per cent across a thousand cases and is correct in about 800 of them, it is well calibrated.

Share as a card

How it is measured#

The standard summary is expected calibration error. Group the predictions into confidence bins, compare the average stated confidence in each bin against the accuracy actually achieved in that bin, and average the gaps, weighted by how many predictions fall in each. Zero is perfect. The measure is a summary and behaves like one: it can hide compensating errors, where a system is overconfident in one band and underconfident in another, so it is read alongside a reliability diagram rather than on its own.

A second measure answers a different question. AUC, the area under the receiver operating characteristic curve, asks whether the confidence signal separates the cases where the system is right from the cases where it is wrong. Calibration asks whether the number is honest in absolute terms; AUC asks whether it is useful for ranking. A system can discriminate well and be badly calibrated, which is a fixable problem, and the reverse, which is not.

What the numbers are#

Xiong and colleagues tested five models across eight datasets. The figures for plain verbalised confidence deserve the space, because the summary "models are overconfident" understates them. Average expected calibration error: 0.520 for GPT-3, 0.461 for Vicuna, 0.436 for LLaMA 2, 0.377 for GPT-3.5 and 0.180 for GPT-4. Even GPT-4's average AUROC of 62.7 per cent sits close to the 50 per cent chance line.

The shape of the numbers is as informative as their size. Stated confidences arrive "as multiples of 5 and with most values ranging between the 80% to 100% range", which the authors suggest means the models "might be imitating human expressions when verbalizing confidence". A system asked how sure it is produces a plausible human-sounding percentage. That is a text-generation act rather than a measurement.

One distinction decides how to read all of this. Internal confidence signals, derived from the model's own probabilities, are considerably better calibrated than the number the model writes in a sentence. The user almost always sees the written one.

The second gap, which is the one that hurts#

Even a calibrated model does not produce a calibrated reader. Steyvers, Tejeda, Kumar, Belem, Karny, Hu, Mayer and Smyth measured both ends on the same items with 301 participants. Model confidence separated correct from incorrect answers at an AUC of 0.751 for GPT-3.5 and 0.781 for GPT-4o. Participants reading those models' explanations reached 0.589 and 0.592, which the authors describe as "only slightly better than random guessing".

They name the shortfall the calibration gap, alongside a discrimination gap, and attribute the human side of it to overconfidence: people "generally believe that LLMs are more accurate than they actually are". Their most practically awkward finding is about length. Longer explanations significantly raised participant confidence while leaving discrimination unchanged, at a mean participant AUC of 0.54. More detail made readers surer without making them righter, which is the opposite of what an interface designer would predict.

Why it belongs in a governance conversation#

Calibration is usually treated as a machine-learning property and filed with the engineers. It decides something organisational. Every oversight arrangement that asks a person to check the cases where the system is unsure depends on the system knowing when it is unsure, and on that knowledge reaching the person in a form they read correctly. Both links are measurably weak, and neither is required by law: nothing in Regulation (EU) 2024/1689 obliges a system to tell the person in front of it how confident it is in the specific output.

So a confidence score is not a control until somebody has checked that it is calibrated on their own data, and checked that the people acting on it read it as intended. The estate's fuller treatment is on how an AI agent should communicate uncertainty.

On what to do with the signal, how an AI agent should communicate uncertainty, does explaining an AI decision help and how to know when AI is wrong. On the reading side, over-reliance, automation bias and why AI sounds so confident. On the underlying behaviour, what an AI hallucination is.

Key research and primary sources

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Calibration and expected calibration error are standard terms in statistics and machine learning. The calibration gap and the discrimination gap are Steyvers and colleagues' terms. Nothing on this page is a SuperSkills coinage. The expected calibration error values from the Steyvers paper are absent here on purpose: they sit inside a figure rather than in the text, so only the AUC figures, which are stated in the body, are quoted.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.

Explainer · SS-2026-172 · Graded against the published rubric

Cite this page

Hirji, R. (2026). Calibration. The SuperSkills evidence base, SS-2026-172. https://thesuperskills.com/research/what-is-calibration. Last reviewed 4 September 2026.

An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.

How citations and IDs work
Questions answered on this page

What is calibration in AI?

A system is calibrated when its stated confidence matches how often it turns out to be right. If it says 80 per cent across a thousand cases and is correct in about 800 of them, it is well calibrated. Calibration is separate from accuracy: a system can be accurate and badly calibrated, by being right often while claiming certainty it has not earned, and it can be inaccurate and well calibrated, by being wrong often and saying so.

What is expected calibration error?

The standard summary measure. Predictions are grouped into confidence bins, the average stated confidence in each bin is compared with the actual accuracy in that bin, and the differences are averaged, weighted by how many predictions fall in each bin. Zero is perfect. Xiong and colleagues report average expected calibration error for plain verbalised confidence of 0.520 for GPT-3, 0.461 for Vicuna, 0.436 for LLaMA 2, 0.377 for GPT-3.5 and 0.180 for GPT-4.

Are language models well calibrated?

Not when asked to state confidence in words. Xiong and colleagues found stated confidences clustering in multiples of five between 80 and 100 per cent, and suggest the models may be imitating human expressions rather than reporting anything measured. Internal confidence signals do considerably better than verbalised ones, which matters because the number a user sees is usually the verbalised one.

Does a calibrated model give you a calibrated reader?

No, and the gap has been measured. Steyvers and colleagues found model confidence separated correct from incorrect answers at an AUC of 0.751 for GPT-3.5, while participants reading that model's explanations reached 0.589, which the authors describe as only slightly better than random guessing. They name the shortfall the calibration gap and attribute it mainly to human overconfidence, people believing the systems are more accurate than they are.

In this hub

Definitions

The terms this field uses, defined against their primary sources.

Ask the evidence
What does the evidence actually show?What should our board be asking about this?Where does Rahim disagree with the consensus?
Bring this into your organisation

If this describes something happening in your teams, say so.

Keynotes, board sessions and advisory work, drawing on research across more than 200 organisations in 30 countries. Tell me the room, the date and the shift you need. A reply within 24 hours.

Start a conversation

Topics and audiences  ·  All research

This is the question underneath the rest of them, and the one a leadership team is least likely to have named out loud. There is the judgement version, and the full range of topics and audiences.

Box of Amazing

Rahim’s free weekly letter on AI and human capability

If this was useful, the weekly letter is where the thinking happens first. Most of what ends up on this site starts there. Weekly essays on AI, capability and the future of work. Read by 25,000 people, every week since 2017. Free, and one click to stop.

Opens Substack to confirm. No pitch in it, unsubscribe in one click, and nobody follows up because you read something.