Why Is AI So Bad at Generating Images of Kamala Harris?

When Elon Musk shared an image showing Kamala Harris dressed as a” communist despot” on X last year, it was quite obviously a fake, seeing as Harris is both a socialist nor, to the best of our knowledge, a Russian cosplayer. And, as some observers noted, the woman in the photo, probably generated by X’s Grok application, had only a passing connection to the vice president.

” AI also is unable to properly describe Kamala Harris”, one X user wrote. ” Looks like they’re posting some strange Latina person”.

Another person reacted, noting that the “dictator” pictured in the picture was a Desperate Housewives star, and that” Grok put ancient Eva Longoria in a slick dress and called it a day.”

” AI only CANNOT replicate Kamala Harris”, a second posted. ” It’s uncanny how the algorithm is at an AMERICAN ( of South Indian and Jamaican heritage )” says the author.

Harris is frequently depicted in artificial intelligence in similar ways. Almost 28 million people have read a tweet with an AI-generated video of Harris and Donald Trump having a loving relationship that culminates with her holding their child, which appears to be Trump, on X. Harris transforms into what appear to be various people throughout the clip, while the noticeably better Trump imagery remains remarkably consistent.

When we tried using Grok to make a picture of Harris and Trump putting their differences aside to learn a version of WIRED, the results consistently captured Harris properly and incorrectly. The evil president appeared with varying functions, wigs, and skin tones. On a dozen times, she looked more like former First Lady Michelle Obama.

Grok differs from some well-known AI photo creators in that it enables users to create fake images of political images. Midjourney began preventing people from displaying pictures of Trump and Joe Biden earlier this year. ( The ban extends to Harris. ) The application could be used to create a range of politically charged pictures, according to a report released by the Center for Countering Digital Hate.

Also, OpenAI’s ChatGPT and Google’s Gemini refused to make pictures of Harris or Trump in WIRED’s tests. However, a number of available resource image generators did, like Grok, produce pictures of politicians. WIRED found one for unit, Stable Diffusion, even produced not-great images of Harris.

What are known as propagation models to create images from language prompts in contemporary AI image generators. These models are fed a large number of labeled images, commonly pulled from the web or taken from different sources. Joaquin Cuenca Abela, CEO of Freepik, a business that hosts several AI resources, including several picture generators, tells WIRED that the difficulty for generators have conjuring up Harris, compared to Trump, is that they have been fed fewer well-labeled pictures.

Despite being a prominent figure, Harris has n’t been as widely photographed as Trump. WIRED’s research of picture dealer Getty Images bears this over, it returned 63, 295 pictures of Harris compared to 561, 778 of Trump. Given her comparatively recent entrance into the political competition, Harris is” a fresh celebrity”, as far as AI graphic designers are concerned, according to Cuenca Abela. ” It usually takes a few months to find up”, he says.

That Harris is a Black person, of Caribbean and Indian origin, also may be a factor. Hugging Face’s head of international policy, Irene Solaiman, claims that “poorer visual recognition for darker skin tones and woman features” may have an impact on how Harris ‘ photos are sorted for automatic labeling. The issue of visual identification failing to recognize women and people with darker skin was initial outlined in the 2018 Gender Shades study, which was co-authored by MIT researchers Joy Boulamwini and Timnit Gebru, who is now the founder and executive director of the Distributed Artificial Intelligence Research Institute.

Another factor could be that Harris ‘ depictions in AI are not very nice. Hany Farid, an expert on algorithmic detection and cofounder of GetReal Labs, a startup that offers software to catch fraudulent media, says that the images are being created to drive a narrative rather than be realistic.

In other words, those who share Harris ‘ AI-generated images may be more concerned with creating meme-worthy situations than with improving the image of her. Both the film in which Harris holds her Trumpy girl and the” communist tyrant” photo shared by Musk and disseminate the Democratic nominee go too far, to distribute propaganda.

Some individuals may even be using bad Harris pictures on purpose, according to Ari Lightman, professor of electronic media and marketing at Carnegie Mellon University’s Heinz College, to further emphasize the idea that she is a scam. ” This is an AI-generated connections era”, Lightman says. ” If it’s done hastily, it’s designed to deliver a information”.