Taking a subset to stand in for the whole, and the ways that goes wrong. If the way you collected your examples relates to what you are trying to measure, no amount of data will fix it. A survey of people who answer surveys tells you about people who answer surveys.
Technical view
Random, stratified and cluster sampling differ in how they trade variance against practicality. Selection bias, survivorship bias and non-response bias are all cases where the sampling mechanism is not independent of the outcome. In machine learning this determines who a model works for: training data collected from one population produces disparate performance on others, and this is a data collection problem that modelling cannot repair.
Related
Part of AI Terminology Index, section: Maths and Statistics.