AI sorted 96% of emails correctly. Which mistakes could hide a real message?

Count legitimate messages flagged as spam and spam that slips through. Overall accuracy does not tell you the cost of either mistake.

Before using an AI label to move or delete mail, define which class is “positive” and inspect each kind of error. Google’s classification guide explains that accuracy, precision and recall answer different questions, especially when class sizes differ.

Our invented 100-message example

Suppose ten messages are spam. The classifier catches eight, misses two, and wrongly flags two of the 90 legitimate messages. It gets 96 of100 classifications right, but two of its ten spam flags are legitimate mail. We calculated accuracy96%, precision80% and recall80% from these fictional counts.

Ask for the errors behind the score

Suggested input: “Using these labeled cases, show correctly flagged spam, legitimate mail wrongly flagged, missed spam and correctly unflagged mail. Calculate the metrics and explain what each denominator means.” This is a proposed input, not a tested mail service.

Choose an action that matches the risk

Our recommendation is to start with a reviewable queue when a wrong flag could hide an important message. Check the flagged examples as well as a sample of unflagged ones. A high score alone is not a reason to automatically delete messages.

Check the real workflow before trusting it

Define how to recover misplaced messages and how the labels will be evaluated on permitted data. Our toy counts do not predict your inbox. Changing the threshold or the types of messages can change the error pattern.

Published October14, 2026. Only fictional confusion-matrix arithmetic was tested. No mailbox, AI classifier, message movement, deletion or threshold setting was accessed or changed.

Sources: Google · Classification metrics.