20 years after Mark Zuckerberg’s infamous ‘hot-or-not’ website, developers have learned absolutely nothing.
Two decades after Mark Zuckerberg created FaceMash, the infamously sexist “hot-or-not” website that served as the precursor to Facebook, a developer has had the bright idea to do the exact same thing—this time with all the women generated by AI.
A new website, smashorpass.ai, feels like a sick parody of Zuckerberg’s shameful beginnings, but is apparently meant as an earnest experiment exploring the capabilities of AI image recommendation. Just like Zuck’s original site, “Smash or Pass” shows images of women and invites users to rate them with a positive or negative response. The only difference is that all the “women” are actually AI generated images, and exhibit many of the telltale signs of the sexist bias common to image-based machine learning systems.
For starters, nearly all of the imaginary women generated by the site have cartoonishly large breasts, and their faces have an unsettling airbrushed quality that is typical of AI generators. Their figures are also often heavily outlined and contrasted with backgrounds, another dead giveaway for AI generated images depicting people. Even more disturbing, some of the images omit faces altogether, depicting headless feminine figures with enormous breasts.
According to the site’s novice developer, Emmet Halm, the site is a “generative AI party game” that requires “no further explanation.”
“You know what to do, boys,” Halm tweeted while introducing the project, inviting men to objectify the female form in a fun and novel way. His tweet debuting the website garnered over 500 retweets and 1,500 likes. In a follow-up tweet, he claimed that the top 3 images on the site all had roughly 16,000 “smashes.”
Understandably, AI experts find the project simultaneously horrifying and hilariously tonedeaf. “It’s truly disheartening that in the 20 years since FaceMash was launched, technology is still seen as an acceptable way to objectify and gather clicks,” Sasha Luccioni, an AI researcher at HuggingFace, told Motherboard after using the Smash or Pass website.
One developer, Rona Wang, responded by making a nearly identical parody website that rates men—not based on their looks, but how likely they are to be dangerous predators of women.
The sexist and racist biases exhibited by AI systems have been thoroughly documented, but that hasn’t stopped many AI developers from deploying apps that inherit those biases in new and often harmful ways. In some cases, developers espousing “anti-woke” beliefs have treated bias against women and marginalized people as a feature of AI, and not a bug. With virtually no evidence, some conservative outrage jockeys have claimed the opposite—that AI is “woke” because popular tools like ChatGPT won’t say racial slurs.
The developer’s initial claims about the site’s capabilities seem to be exaggerated. In a series of tweets, Halm claimed the project is a “recursively self-improving” image recommendation engine that uses the data collected from your clicks to determine your preference in AI-generated women. But the currently-existing version of the site doesn’t actually self-improve—using the site long enough results in many of the images repeating, and Halm says the recursive capability will be added in a future version.
It’s also not gone over well with everyone on social media. One blue-check user responded, “Bro wtf is this. The concept of finetuning your aesthetic GenAI image tool is cool but you definitely could have done it with literally any other category to prove the concept, like food, interior design, landscapes, etc.”
Halm could not be reached for comment.
“I’m in the arena trying stuff,” Halm tweeted. “Some ideas just need to exist.”
Luccioni points out that no, they absolutely do not.
“There are huge amounts of nonhuman data that is available and this tool could have been used to generate images of cars, kittens, or plants—and yet we see machine-generated images of women with big breasts,” said Luccioni. “As a woman working in the male-dominated field of AI, this really saddens me.”
I’m just curious if we could collect demo data on the raters and have that train the generation algorithm. Could we find regional, statistically significant differences in aesthetic preferences? Would we be able to trace the cultural influence of different groups by their preferences?
Do some groups prefer specific shapes/sizes/colors? What’s the most predictive feature of perceived attractiveness across different groups?
I dunno, sounds like some interesting research into the visual aspect of attraction and also implicit biases.
I think I may have participated in research like this once. It collected a lot of demographic data on me and then I had to rate a bunch of people and in the next section make value judgements about whether each of a series of people looked like they were likely to be smart, trustworthy, bad tempered etc. It was all on a timer.
Besides “pure” Psychology, Aesthetics research is dealing with such kind of questions and has been going on for a while.
This does not only encompass visual features in humans to which people feel some form of attraction, but also stuff like music or visual art.
It’s not my field and it has been some time since I read good literature on this, which is why I am not giving you any possibly erroneous summaries. But I am sure that this has been investigated and is still a topic of active research. So you can take my comment as a pointer.
First hit, when searching for: “cultural differences in visually appealing facial features”
https://www.sciencedirect.com/science/article/pii/S0960982221003523
Have fun researching and reading.