How machines learn
Accuracy is the share of a model's predictions that are correct, calculated as the number it got right divided by the total number it tried.
Accuracy is the most intuitive score in machine learning. If a model labels 100 photos and gets 93 right, its accuracy is 93 percent. It is easy to compute and easy to explain, which is why it is often the first number people quote about a system.
But accuracy can mislead when categories are unbalanced. Suppose one in a thousand transactions is fraud. A model that says 'not fraud' every single time is 99.9 percent accurate and completely useless. In cases like this, teams also track how many real frauds were caught and how many false alarms were raised.
Accuracy should always be measured on data the model did not train on. A high training accuracy with a low test accuracy is the classic sign of overfitting. Also ask what 'correct' means: for a chatbot, there may be many acceptable answers.
A model that sorts 1,000 bird photos and mislabels 40 of them has 96 percent accuracy, but you would still want to know which species it keeps confusing.