Well-calibrated
A model is well-calibrated if its confidence in an answer matches how often that answer is actually correct.
Imagine a weather forecast that says there's a 70% chance of rain. If it rains 70% of the times the forecast says that, it's well-calibrated. If it rains only 30% of the time, it's not.
For AI models, this means if a model says it's 90% sure about an answer, it should be correct about 90% of the time when it gives that answer. If it's often wrong when it claims high confidence, it's poorly calibrated. This helps users know when to trust the AI's output.
A model is well-calibrated if its predicted probabilities align with the true likelihood of outcomes. For classification tasks, this means that for all instances where a model predicts a class with probability p, the true proportion of those instances belonging to that class is also p. Poor calibration can lead to overconfidence or underconfidence, affecting decision-making based on model outputs.
Frontier AI labs are focusing on making models more reliable. Anthropic's recent work shows Claude models can now estimate their own correctness, which is a step towards better calibration. This is important for building trust and ensuring AI systems are used responsibly.
- Language models (mostly) know what they knowAnthropic · Anthropic Research ·
- Anthropic Claude
- Anthropic's Claude models are being developed to estimate the probability that their own answers are correct, indicating a focus on calibration.
- Accuracy
- Accuracy measures the overall percentage of correct predictions, while
- Confidence
- Confidence is the model's internal score for a prediction, whereas c
The concept of calibration in machine learning originates from statistical decision theory and has been a long-standing area of research in probabilistic modeling.
Ask an AI model a question where you know the answer and ask it to rate its confidence. Then, ask a question on a topic you are unsure about and see if its confidence level seems reasonable. Repeat this with different types of questions to see if its confidence generally matches its accuracy.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- PromptRole-play a student struggling with math for AI tutor simulationSimulate a student's math struggles and uncertainty for AI tutor evaluation. · 5 min
- TemplateReusable workflow templates for healthcare teamsTeams can create and share templates for common workflows like documentation or prior authorizations. · 15 min
- TemplateTemplate for building an AI tool tailored to your workA structured approach to defining and building a custom AI tool for specific business needs. · 15 min
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
- HeuristicDouble coding agent task completion with docs linksProviding coding agents with a documentation link significantly increases their task completion rate. · 2 min
Questions people ask
- Why is calibration important?
- Calibration is important because it tells users how much to trust an AI's answer. A well-calibrated model's confidence score accurately reflects its likelihood of being correct.
- How do you measure calibration?
- Calibration is often measured using reliability diagrams or metrics like Expected Calibration Error (ECE), which compare predicted probabilities to observed frequencies.
First seen on Subvolts Aug 26, 2026. Updated Sep 15, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.