Probability, Uncertainty, and Confidence
A number that looks like 80% sure is a score until someone shows it is calibrated. Fluent language is not confidence.
Four different things people call "confidence"
People use one word for four different objects.
- Chance in the world. A fair coin. A sampling process. A flood after rain. A number in [0, 1] assigned to an event under a probability model, or a long-run frequency story. Say which story you mean.
- A model score. A classifier outputs 0.85. Softmax, or a similar squash, turned internal numbers into a list that adds to 1. That 0.85 is a score.
- A human "I'm sure." A person saying the words. That is speech. It is not a measurement.
- A product UI percent. A screen that prints a percent, or a chat that writes "I'm 80% sure." Display is not calibration.
Those four are not the same object. Mixing them is how a score becomes a fake certainty.
AI is a category of machine-based systems. [nist-ai-glossary] It is not defined as learning. A search engine, a rule list, and a planner can all be AI systems without a confidence slider. See What is AI? and AI Is More Than Machine Learning.
This page is literacy. It is not a redo of Mathematics for AI. That lesson is the toolkit. This one is how not to fake a percent.
A fair coin is 50/50. A score is a score.
If a coin is fair, heads and tails are equally likely under that model. Probability 0.5. That is a claim about a repeatable event. You can say it as a model of the coin, or as a long-run frequency of flips. Say which story you mean.
A weather app percent is not automatically that kind of number. A chat that says "I'm confident" is not that kind of number. A classifier score of 0.85 is not that kind of number. Those displays can be useful. They are not a fair coin.
Uncertainty is not a vibe. It is not a percentage you invent to fill the silence.
Conditional probability is "given that"
Given that it rained, the flood chance on Main is higher than the flood chance if you ignore the rain. That is conditional probability: P(flood given rain). It is not P(rain given flood). Those two numbers answer different questions.
Illustration, not a study. A bag holds 3 red cards and 1 blue card. If every card is equally likely, the chance of drawing red is 3/4. You draw one card and it is red. If you do not put it back, the bag has changed. The chance the next card is red is now 2/3. The number updated because the evidence arrived. That is not a mood.
Bayes is the same idea with names. You start with a prior (what you thought before the evidence). You have a likelihood (how expected the evidence is if a hypothesis is true). You get a posterior (what you think after the evidence). The posterior is not the likelihood. Do not smash them.
Illustration, not a study. A drawer has two coins. One is fair. One has heads on both sides. You pick a coin at random. You flip it and see heads. Heads is more likely under the two-headed coin, so that coin is now more plausible than it was before the flip. Prior: which coin, before the flip. Likelihood: chance of heads given which coin. Posterior: which coin, after seeing heads. Three numbers. A toy. Not a clinic.
A model score is a transformation. Calibration is an extra claim.
A classifier has internal numbers for each option. Softmax, or a similar squash, turns those numbers into a list that is non-negative and adds to 1. A product may print the largest as a "confidence." That printout is a transformation of internal numbers. [goodfellow2016]
Calibration is the extra claim: among cases where the system outputs a score near p, about p of them were correct, on the data and conditions studied. [guo2017] [niculescu2005]
Most of the time the product has not shown you that measurement. Then 0.85 is a score. It is not calibrated confidence. This page does not invent a calibration curve, and it does not invent a reliability-diagram number.
A score on a test set is not a calibrated probability of being right in your clinic. The test set is still a sample. For how to inspect a claim about a named system, see How to Check an AI Claim. Benchmarks and leaderboards belong on the evaluation lesson, not here.
Fluent language is not calibrated confidence
A paragraph can sound sure and be wrong. Fluency is a style of output. It is not evidence of certainty. The OECD definition of an AI system is a machine-based system that infers from inputs how to generate outputs. [oecd-ai-2024] Fluent prose is an output. The definition does not make fluency a probability.
That is why How to Check an AI Claim exists. Ask for a first-party fact. Do not treat a confident tone as a measurement.
What to do as a reader
- Ask which of the four objects you are looking at: chance in the world, a model score, a human "I'm sure," or a UI percent.
- Ask whether a calibration was published for that score, on data like yours.
- Do not invent a percent to fill the silence.
- A fair coin may be 50/50. A model's 0.85 is a score until someone shows it is calibrated.
- Related lessons: Mathematics for AI for the toolkit. Data: What Machines Actually Learn From for "the dataset is a sample." Machine Learning Fundamentals for training. How to Check an AI Claim for a named-system claim. What is AI? for the category. AI Is More Than Machine Learning for methods that are not this pipeline. Tokens, Embeddings, and Context for scores over next tokens.
References
National Institute of Standards and Technology (NIST) (2026). Artificial Intelligence — Glossary. NIST Computer Security Resource Center.
National Institute of Standards and Technology (NIST) (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1.
Organisation for Economic Co-operation and Development (OECD) (2024). Explanatory memorandum on the updated OECD definition of an AI system. OECD Artificial Intelligence Papers.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning. MIT Press.
Niculescu-Mizil, A., and Caruana, R. (2005). Predicting Good Probabilities With Supervised Learning. Proceedings of the 22nd International Conference on Machine Learning (ICML).
Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. Proceedings of the 34th International Conference on Machine Learning (ICML), PMLR 70.
Angelopoulos, A. N., and Bates, S. (2021). A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. arXiv preprint.
Citation Note: Some citations are open access (arXiv or DOI, where linked). Others are books or journal articles that may sit behind a publisher paywall. Use the linked DOI or arXiv when available. If you notice any citation errors, please let us know.