In classification, accuracy is the proportion of evaluated predictions that match the reference labels. Divide correct predictions by all evaluated predictions. This is a defined test metric, not a universal score for how truthful an AI assistant is.
How to calculate classification accuracy
If a model classifies 80 of 100 labeled emails correctly, its accuracy on that test set is 80 ÷ 100 = 80%. For binary classification, correct predictions include both true positives and true negatives.
The result depends on the test examples, reference labels, and evaluation conditions. Comparing scores is meaningful only when those conditions are comparable.
Why 99% accuracy can be useless
Imagine 100 emails: 99 legitimate messages and one spam message. A filter that always predicts “not spam” gets 99 correct answers and 99% accuracy, but detects none of the spam. This is a constructed example illustrating class imbalance, not a measured product result.
| Metric | Question it answers |
|---|---|
| Accuracy | What fraction of all classifications were correct? |
| Precision | Of the items flagged positive, how many really were positive? |
| Recall | Of the actual positive items, how many were found? |
How do you check a generated answer?
A free-form answer can contain several claims, calculations, and omissions. A single correct/incorrect label may hide those differences. Define an evaluation that fits the task: check factual claims against sources, recalculate totals, and assess whether required details were included.
Speech recognition, forecasting, translation, and search ranking also use task-specific measures. Do not treat their reported scores as the same metric as classification accuracy.
The itinerary exercise checks time arithmetic separately from opening hours. Use the verification routine to identify claims that need an external source.
Before trusting an accuracy claim
- Ask what task, dataset, sample size, and reference labels were used.
- Check which errors matter most, not just the overall percentage.
- Look for performance on rare and relevant cases.
- Do not generalize a test-set score to every future task or user.