Technically, it’s not a trend. We share our photos with OpenAI to build huge datasets for training AIs. We see how we look in 1980s AI-generated images; AI reads our face, skin color, skull type, etc to categorize us by race, gender, emotions, ability, sexuality, and personality. Each category is a dataset for further study.
As of now, AIs are good at object detection and facial recognition, deciding whom to interview for a job, which students are paying attention in class, which suspects to detain etc. But this is where bias kicks in based on politics, ideology, prejudices, etc. A man with a beer is “seen” as an alcoholic, and a woman in a bikini is ‘seen” as a sl*t by AI. So now we’re trying to train AI not to be judgmental and biased by creating more specific datasets.
Why is it complicated? - Images do not describe themselves. It can mean different things depending on who looks and where they are located (Monalisa). Images are open to interpretation and reinterpretation. Objects don’t have emotions, so it’s easy to categorize images of apples and oranges. But humans have names, and humans emote. A single person can emote in several ways. He may also emote neutrally without any expression. A person smiling in a photo is a “performed” expression—not relating to any interior state. So AI needs millions of human photos to train datasets; every photo is open to question and is contested.
The 1st image dataset was JAFFE in the 1980s. Since 2009, one of the most significant training sets is ImageNet, which has 14 million labeled images organized into more than 20,000 categories; however, it does not include a “person” category. why?
Politics and controversy - ImageNet is based on WordNet, a database of words. These words are grouped under ‘synsets”. For example, “chair” is nested as artifact>furnishing>furniture>seat>chair. However, ImageNet is restricted to nouns. In the ImageNet hierarchy, every concept is organized under one of 9 top-level categories: plant, geologic formation, natural object, sport, artifact, fungus, person, animal, and miscellaneous. Below these are layers of additional nested classes.
For example, “ Human body” falls under Natural Object>Body> Human Body. This is further classified as “male body”; “person”; “juven**e body”; “adult body”; and “female body”. The “adult body” includes “adult male body” and “adult female body”. What’s political and controversial here? - The AI calls only male and female natural. Whereas trans people fall under “Hermaphro***e”, further under “ pseudohermaphro***e”. Basically, the AI categorizes L***Q as “Abnormal Se***l Relations, Including Sex**l Crimes.” Such labelling is problematic, illogical, and cruel, especially when it comes to labels applied to people
AI further categorizes humans as Bad Person, Call G**l, Drug Add**t, Clo**t Queen, Conv**t, Crazy, Failure, Flop, *ucker, Hypocrite, Je**el, Klepto***iac, Loser, Mela***olic, Nonperson, Per**t, Prima D**a, Schizop***ic, Second-Rater, Sp***er, Streetwalker, S*ud, To**er, Unskilled Person, Wa**n, Wa**rer, and W**p. There are many ra**t sl**s and misogynistic terms. This practice has only become more common in recent years, inside the big AI companies, where there is no way for outsiders to see how our images are being ordered and classified.
The images we upload become part of “trends” aren’t simply raw materials to feed algorithms - they’re controversial and political. There is no “neutral,” “natural,” or “apolitical” point training data can be built upon. There is no easy technical “fix” by shifting demographics, deleting offensive terms, or seeking equal representation by skin tone.
The whole idea of collecting our photos (or images) through the “1980s trend”, categorizing them, and labeling them is itself a form of politics, filled with questions about who gets to decide what images mean and what kinds of social and political work those representations perform.