Ranking student attention in a classroom
A school installs software that scores each student's attention from facial video and ranks the class weekly.
Bias & ethics
Two working modules. The first shows how an accuracy number can hide complete failure on most categories. The second asks you to judge three situations — there is no score, only the considerations that follow from your answer.
Move the slider to change how much of a fictional 1,000-image dataset belongs to the largest class. The 'model' below is the laziest possible one: it always predicts the majority class.
70.0%
overall accuracy
| Class | Recall |
|---|---|
| Neutral-type (majority class) | 100% |
| Smiling-type | 0% |
| Frowning-type | 0% |
| Surprised-type | 0% |
The headline number can look respectable while three of four categories are never predicted at all. That is why per-class recall and a confusion matrix matter: they show where the errors land, not just how many there are.
Class balance is not the same as demographic fairness. A perfectly balanced set of expression categories can still be collected mostly from one age group, one region or one lighting setup — and a model can score well on every class while failing for particular groups of people. These need to be measured separately.
Fictional teaching data. No real dataset is described here.
Choose a position for each scenario. There is no moral score: the point is to surface consent, the right to opt out, privacy, the cost of errors, and what a less invasive alternative would look like.
A school installs software that scores each student's attention from facial video and ranks the class weekly.
A screen in a station selects adverts based on the inferred mood of people walking past.
An individual chooses to use an app that suggests phrases, can be corrected, and can be switched off at any time.