Humans similarly can be reduced down to relationships of nearby photo receptors. That’s exactly how V1-V4 works in our human brain… building detectors that go higher and higher level from nearby bits of information from two 2D sensors. In the same way, most neural networks CNNs do exactly this as well. CV models similarly get trained to produce their own higher level abstractions of the world that they, like us, fit data too. As much like computers, we too can be subject to all kinds of gross stupid errors, like optical illusions, mistaken identities, etc.
The best CV models are often fine tuned from more generalized models that can recognize thousands and thousands of different kinds of objects, even if in the end they only need to find 1 kind. This is because they build generalized representations of the world.
None of this is to suggest there any bias that fundamental to them. We’re starting to see fruits of multi-task learning where models are integrating across domains, which ends up improving accuracy even for just one modality by incorporating more world knowledge. So it’s not like we’ve reached some kind of theoretical limit as to what they can do, and it turns out that theoretical limit is sexist and racist. Nothing could be further from the truth.