This is it. A human being can identify the context. Taking Fountain as an example -- a photo of a urinal on a building merchant's website is an illustration of

This is it. A human being can identify the context. Taking Fountain as an example -- a photo of a urinal on a building merchant's website is an illustration of the product they're selling, not an infringement. But the filter doesn't know about context, it just correlates two images... and you get "Comparison with copyrighted work 'Fountain', 75% match". 75% > 0%, so the filter says "non".