AI and programming classes in Chatham

The learner downloads two Census 2021 tables from the Nomis API for all 853 output areas in Medway, covering 111,425 households: the type of each home, and how many cars or vans each household has. An area counts as mostly terraced when at least half its households live in a terraced house, which is true for 296 areas, 34.7% of them. A logistic regression then gives every area a score between 0 and 1 from just two clues: the share of households with no car, and the share with two or more. Every score is produced by a model that did not see that area during training.

An ROC curve shows, for every possible cut-off, how many of the true cases are caught against how many false alarms are raised. A useless model runs along the diagonal; a perfect one hugs the top-left corner. The area under the curve, AUC, sums this up in one number: 0.5 is guessing, 1.0 is perfect. Random scores give 0.515, the no-car share on its own 0.642, and the two-feature model 0.673. The clues carry real information, but not much.

The default cut-off of 0.5 is nearly useless here. Because scores rarely rise that high, it catches only 11 of the 296 terraced areas, and its accuracy of 61.8% is worse than simply answering no every time, which scores 65.3%. Lowering the cut-off catches far more, but every extra catch costs false alarms: catching 80.1% means wrongly flagging 287 other areas. Which point is right depends entirely on what a mistake costs. A spam filter and a medical screening test would sit at opposite ends of the same curve.

Reading an ROC curve shows Chatham teenagers how AI decisions are actually made, which is a sound reason to take up coding in 2026. The longer argument is in why teenagers should still learn to code in 2026.