How machines learn
A threshold is the cutoff score a model's confidence must reach before it counts as a 'yes'; moving it trades fewer false alarms for more missed cases.
Many models do not simply say yes or no. They output a confidence, such as '82 percent likely spam.' Someone has to decide where the line is. Above the line, the email goes to junk; below it, it stays in the inbox. That line is the threshold, and it is chosen by people, not learned by the model.
Moving the threshold changes behavior without retraining. Set it high and the model only acts when very sure, so fewer innocent emails are junked but more spam slips through. Set it low and it catches more spam but also more real mail. Neither setting is 'correct'; it depends on which mistake costs more.
Thresholds are everywhere in deployed AI: fraud alerts, medical screening, content moderation. When a system feels too aggressive or too lax, often the threshold, not the model, is what was tuned. It is a business and ethics decision hiding inside a number.
A spam filter set to act only above 70 percent confidence lets a suspicious 65 percent email through; lowering the cutoff to 50 percent would catch it.