ImageNet is basically made by doing Google image searches on words from the WordNet database (a structured collection of words and their definitions), and then

ImageNet is basically made by doing Google image searches on words from the WordNet database (a structured collection of words and their definitions), and then doing quality control by crowdsourcing. So from Googling "assenter" and asking a bunch of people "here is the definition of an assenter, does this image contain one?", they have 126 confirmed pictures of assenters that can be used to see what a typical assenter looks like. Then the data is used to train a neural network to recognise these classes. Some of these steps may introduce biases.