Part 1: Who Gets to Label?
Kate Crawford and Trevor Paglen use Magritte’s painting of a pipe — labeled “this is not a pipe” — to make one point: an image does not carry a fixed meaning. Someone has to decide what it means. In machine learning, that someone is the person who labels the training data.
This matters because the labels are choices, not facts. Whoever builds the dataset decides the categories, draws the lines between them, and decides what counts as “normal.” The model then learns those choices and repeats them at scale, as if they were objective truth.
This has real effects on society. If a dataset was labeled by a small group of people, the model carries their assumptions and blind spots. People who were not included in the data get misclassified or ignored. And because the output looks like a neutral number, people trust it — even when it is just one group’s opinion dressed up as science. The power to label is the power to define reality for everyone the system touches.
Part 2: Cough Severity Classifier
For my project I trained a sound classifier in Teachable Machine and connected it to a p5.js sketch. It listens through the microphone and sorts each cough into four classes I made: Mild cough, Severe cough, Throat clearing, and Background noise.
It is not a real medical tool, and that is the point. I recorded all the samples myself, at my own desk, using my own body. So the model’s idea of a “severe” cough is really just my idea. When you cough, it shows a big confident number like “Severe — 87%,” which makes a personal guess look like a measurement.
I designed it to expose its own problems:
- The confidence bars always show all four classes. The model can never say “this is none of these” — it must force every sound into a box, even throat-clearing, which is not a cough.
- The Background Noise class is required. Everything I didn’t name becomes “noise” — just like real datasets ignore the people they never sampled.
- Pressing the spacebar marks “says who?” and counts how many times you doubt the verdict, because the real question is not is it right but who gave it the power to judge.
The best way to use it is to try to fool it — fake a cough, clear your throat, play someone else’s cough. It stays confident anyway. That moment of being fooled is what shows the real lesson from Crawford and Paglen: the label does not describe the truth. It just has the power to decide it.
Sketch: https://editor.p5js.org/ervin021013/sketches/-79zPMZni
Leave a comment