Eddie Ungless

Eddie Ungless

Looking for work A collection of my thoughts on the topics of ethics, fairness and social bias in NLP and beyond.

The Ethics of Stock Image Generation

In October 2022 it was announced that Shutterstock would soon be offering licensed images generated using D-ALLE 2, a powerful image generation model1. This suggests a commerical future for D-ALLE 2 that until recently had mostly been the plaything of tech enthusiasts and digital artists. Researchers have already expressed a concern about the biases and stereotypes the generated images might exacerbate2, but I am also particularly worried about another ethical issue: the removal of marginalised individuals from the process of generating images of these communities.

To explain further, currently at least some stock images of Black, queer, disabled or otherwise marginalised individuals are produced by taking photos of models belonging to these communities. And likely at least some of the stock images are produced by creators belonging to these communities (though the world of photography, like tech, is dominated by straight white men3). Compensation for modeling in these images might be relatively poor - but when new images of marginalised individuals can be produced without the need to actually employ a marginalised individual, any compensation seems preferable to none.

Whilst lack of representation is in itself an issue2, I argue that this kind of false, inauthentic representation is also harmful, such that even a perfectly “unbiased” model (if such a thing could exist) would be complicit in causing harm by erasing marginalised models and creatives from the stock image production pipeline. “Shudu Gram” (a virtual Black model created by a white man) was controversial back in 20184, but these kinds of fake models may become commonplace. Since I drafted this blog post in 2022, Levi has stated they plan to “increase diversity” in their product images using AI5. In reality, they will be increasing the appearance of diversity whilst actually marginalised models continue to be passed over for work opportunities.

  1. https://www.forbes.com/sites/johanmoreno/2022/10/26/shutterstock-will-soon-offer-ai-generated-images-showing-what-the-future-of-dall-e-might-look-like/ ↩

  2. http://arxiv.org/abs/2211.03759 ↩ ↩2

  3. https://www.zippia.com/professional-photographer-jobs/demographics/ ↩

  4. https://www.newyorker.com/culture/culture-desk/shudu-gram-is-a-white-mans-digital-projection-of-real-life-black-womanhood ↩

  5. https://internetretailing.net/levis-ai-diversity/ ↩

Updated: