Data Collection, Privacy, and Bias

A data set looks complete when you're staring at rows and rows of entries, but completeness is about coverage, not row count. Every data set comes from some collection process — a sensor, a survey, an app's sign-up form — and that process always points at some population and not another. If the process systematically reaches certain people more than others, the data inherits that imbalance no matter how many entries pile up. A face-detection app trained on photos scraped mostly from one source will see mostly the people who used that source.

This becomes a program problem once you remember how these systems learn: they find patterns in whatever examples they're given. If 90 percent of training images show one demographic, the program gets heavy practice recognizing patterns common to that group and comparatively little practice on the rest. The output isn't a neutral tool — it's a tool that performs well for the 90 percent and noticeably worse for everyone else, because its 'knowledge' is really just a reflection of its training data's proportions. A machine did the sorting, but the humans who built the collection process are why the skew exists.

This connects to the unit's bigger theme of collections holding real data about real people. Arrays and ArrayLists are neutral structures — they'll happily store whatever you put in them — but the question of what you put in them, and who agreed to be in there, sits outside the code. A data set built from people who never consented to be tracked raises a separate problem: even an accurate, unbiased sample can be an ethical one only if the people inside it had a say.