Science

The AI systems in learning mode are compiling much more detail and information than the people in the process are doing. The people are categorizing images with very broad descriptors like "red toy truck". That is a very simple level of compilation.

When the AI looks at a ton of red toy truck images, perhaps a couple thousand in that category alone, it is also looking at thousands of categories and reading millions of pieces of text. If you looked at few thousand red toy truck pictures you too would learn more than you would be consciously aware of: details like the common natures of the plastics used, the shades of red used, the most popular models, etc.

Think of it as three phases: 1) Assemblage and categorization of the images and text, 2) learning by the AI from the data and the categories, 3) generating responses to queries. Phases 2 and 3 are run without human intervention.

So it is not true that the people are compiling answer sets. That is phase 3.