First a question: why does human brain theorize things in 3 or 4 terms most often ex: CAP, CMM , C-4, ACID, BASE etc. Most reserach papers and PPTs will have 3 letter or 4 letter acronyms rarely 5,6 (hexagons) or 8(octagons). (Interestingly 7s are missing and so are higher numbers 9+)
Plausible explanation: because geometrically closed spaces start with 3 and then 4 sides. A topic can be interpreted and is of interest to human brain only if it’s a closed space (you want to sustainably keep it going). The reality is a circle or sphere with infinite sides but the first approximation is a 3 or 4 sided closed space which immediately gives the comfort of a closed space and therefore stability i.e. factors that pull and push to create a sustained phenomenon including one ore more eventual outcomes. Visualize pyramid where outcome is the central vertical pole.
The above 2 paragraphs are an example of meta-thought, i.e. thoughts on patterns used by human brain to approximate reality and design/build systems based on its understanding a.k.a Software and now AI system design. Bias creeps-in when you have limiting constructs, thus coming up with limiting model and eventually ethics suffer. You can argue feature selection is now automated, there is deep learning, that’s a great point. However, what you feed, what you choose to feed, what you choose to test, what you choose to test with, how you test is still in your realm i.e. training data selection, test data selection, test design and even more importantly how you create data i.e. how you choose to encapsulate/represent a physical phenomenon in data is where bias/ethics conversation need to start from. ex: are you using a black/white picture, grey-scale picture or a color-yet-smoothed out picture. To stretch it a bit further is it a 2-D or a 3-D view.