077 - Image Filters

Image data on the internet is already heavily poisoned, and surprisingly "Generative AI" might not be the leading source of it.

"Filters" have become both extremely popular across common image posting platforms, and extreme in the ways that they alter photos. Some of these filters are the obvious ones, where it looks like a dog nose is drawn on someone, or a weird white dotted mask is over their face. However, some of them alter the photos in extreme and disturbing ways that you might not be able to spot.

Filters that "whiten" the skin of people in them have become increasingly common in some parts of the world, but more extreme than altering skin color is altering facial features such that they conform much more closely to Western Caucasian stereotypes.

In previous years it was already demonstrated how the earlier and less extreme versions of filters had a significant and very clearly negative psychological impact on an entire generation. The new filters entering common usage on social platforms are much more extreme than this, and they confer an increasingly potent stigma against looking like anything other than a Barbie Doll. The direct scale of this impact is billions of people, and entire regions of the planet.

This trend is horrifying both from an ethical perspective, but also from an AI and data science perspective. How are you going to build a less biased Generative AI model when most of the images of people elsewhere in the world have been pushed through one or more extreme "Westernizing" filters?

New filters are also relatively trivial to produce, and very difficult to robustly detect and counter, shy of imposing extreme measures and cutting your available supply of data to less than 0.1% of what is available. While some "ethnicity filters" have been removed from apps in the US, they are thriving elsewhere. These filters also serve to boost sales of "skin whitening" treatments, Botox, and luxury goods for those aiming to double down.

Humans have a relatively massive amount of brain tissue dedicated to visual processing relative to other sensory capacities, and yet human visual systems are already being fooled to this degree. If the human brain's largest processing capacity is being fooled, how bad off might your other capacities already be?

This is another large-scale problem, and cleanup of the massive and rapidly growing pool of poisoned data won't be easy. In the interim, before viable technologies to counter this issue are deployed at scale, people building AI systems desperately need to verify that the visual data they're training on hasn't been Westernized through extreme filters.

I joked about how SDXL has an absurd bias toward giving every man a cleft chin, which is annoying and often unsightly, but AI systems trained on such extreme filters will no doubt prove far worse, even as the consequences to global society rise.