The ease at which bias creeps into AI

Removing AI bias is a critical component in the design and training of AI models, however despite the care taken, it can be incredibly easy for bias to creep in.

This is because we often don't see our own biases and, even at a macro level as a species, we may hold biases we are not aware of that fundamentally impact how our AIs perform.

A great example I came across the other week was in image generation AIs, when asked to create a selfie of a group of people. There is a tendency for AI models is to portray a smiling group, whatever their era or cultural background.

This shows a bias in that many groups in history traditionally don't smile in photos - however the western norm has largely become smiling, and so the training sets these AIs use are biased towards smiling, regardless of the culture or timeperiod of the people represented.

AI generated image of French WW1 soldiers selfie