Generative Image AI: Existential Cultural Bias & The Critically Urgent Need To Globalize Image Models

One of the most surprising findings of our experiments with commercial generative AI image creation models has been the abject failure of Silicon Valley to globalize their models, despite all of their talk of "debiasing", "harms" and "responsible AI". For all their focus on AI bias, developing intricate "principles", maintaining  teams of "bias experts", encouraging "red teaming" and collaborating with academics and non-profits, Silicon Valley appears to have focused nearly exclusively on small tweaks to prevent mediagenic instances of contemporary American bias issues from making the headlines, rather than genuinely attempting to make their models culturally and globally aware.

There is simply no excuse in 2023 for an image generation model to produce images like the following in response to the prompt "create an inspirational image of Turkey filled with the nation's symbolism, history and imagery." Create an image of a "doctor" or a "CEO" and the intense past media attention to those biases means most image generation models have been tuned to cycle through a suitably diverse combination of races and genders. Yet, even here gender is construed as a binary male/female, despite most of the companies behind the models embracing a vast spectrum of genders in their corporate missions and official AI policies. However, stray beyond this small collection of extremely well known bias examples and the models utterly collapse back into the harmful and toxic stereotypes of their training data, reinforcing just how fragile these small hyper-targeted manual nudges are and just how unseriously the generative AI community is taking these topics.

Most devastatingly of all, when one reaches beyond the US and Western Europe, even these meager bias mitigation efforts reach their end, with the companies seemingly investing little at all in the bias issues that arise outside of the West. Critically, the images produced by the major generative AI image models reinforces just how little work has been done to "globalize" them: training them to understand the national, religious, cultural and other symbolism of the diverse societies of the world to ensure they understand the appropriate context and usage of those symbols. In much the same way that models have been tuned to avoid reproducing certain harmful Western tropes they've learned from their training data, so too could these models be adjusted to avoid producing at least the most glaring and toxic misuse of symbols from across the rest of the world, but there is simply little movement on that front.

If Silicon Valley wants to genuinely claim to care about AI bias, AI principles, responsible AI and the like, it must start taking seriously the issues of representation and harmful and toxic representations beyond just filtering out a few American stereotypes: they must for the first time globalize their models.