Deep learning, image generation, and the rise of bias automation machines

DALL-E, a new image generation system by OpenAI, does impressive visualizations of biased datasets. I like how the first example that OpenAI used to present DALL-E to the world is a meme-like koala dunking a baseball leading into an array of old white men — representing at one blow the past and future of representation and generation.

It’s easy to be impressed by cherry-picked examples of DALL•E 2 output, but if the training data is web-scraped image+text data (of course it is) the ethical questions and consequences should command much more of our attention, as argued here by Abeba Birhane and Vinay Uday Prabhu.

Suave imagery makes it easy to miss what #dalle2 really excels at: automating bias. Consider what DALL•E 2 produces for the prompt “a data scientist creating artificial general intelligence”:

When the male bias was pointed out to AI lead developer Boris Power, he countered that “it generates a woman if you ask for a woman”. Ah yes, what more could we ask for? The irony is so thicc on this one that we should be happy to have ample #dalle2 generated techbros to roll eyes at. It inspired me to make a meme. Feel free to use this meme to express your utter delight at the dexterousness of DALL-E, cream of the crop of image generation!

The systematic erasure of human labour

It is not surprising that glamour magazines like Cosmopolitan, self-appointed suppliers of suave imagery, are the first to fall for the gimmicks of image generation. As its editor Karen Cheng found out after thousands of tries, it generates a woman if you ask for “a female astronaut with an athletic feminine body walking with swagger” (Figure 3).

I also love this triptych because of the evidence of human curation in the editor’s tweet (“after thousands of options, none felt quite right…”) — and the glib erasure of exactly that curation in the subtitle of the magazine cover: “and it only took 20 seconds to make”.

The erasure of human labour holds for just about every stage of the processing-to-production pipeline of today’s image generation models: from data collection to output curation. Believing in the magic of AI can only happen because of this systematic erasure.

Figure 3

Based on a thread originally tweeted by @dingemansemark@scholar.social (@DingemanseMark) on April 7, 2022.

2 thoughts on “Deep learning, image generation, and the rise of bias automation machines

  1. Is it bias or does it reflect a biased reality? Computer science is 70% male, therefore an unbiased model would show a computer scientist being male 70% of the time. Is the model biased or is the system biased?

    Also, the point of most human advances is to replace labor(look at machines for instance), I’m not sure what makes Dall-E-2 special, except maybe the industry it’s in.

  2. It literally says in the post “does impressive visualizations of biased datasets”. And yes I think we should care about the harms of perpetuating existing biases.

Leave a Reply

Your email address will not be published. Required fields are marked *