Have you ever seen an Asian being friends with a white person, either as a couple of mixed descent or as friends of different ethnicities? For the image generator of Meta which operates with the help of artificial intelligence, it is obvious that it cannot imagine this process.
We have tried several times to create an image using various prompts, such as "Asian and Caucasian friend", "Asian and white wife" and "Asian and Caucasian husband". Only once did Meta's image generator manage to return an accurate image that matched the parameters we had set.
See more: Meta and Instagram offline again for some users
Adapting the text-based prompt did not appear to be effective. When I asked for an image of an "Asian man and a white woman smiling with a dog," Meta's text-generating app in Instagram quoted three photos of two Asians looking backwards. By changing the word "white" to "Caucasian", the same effect was observed. The sentence "Asian and Caucasian woman on the wedding day" presented us with an Asian man in a suit and an Asian woman in traditional dress, but looking at it further there was a confusion between kimono and chipao (Chinese dress).
The creation of images presented Problem when representations of platonic relationships were requested, such as "Asian man with Caucasian boyfriend" and "Asian woman and white boyfriend". Each time, a picture of two Asians was returned. When a picture with "Asian girl and black boyfriend" was requested, the AI again generated a picture with two Asians. The request for "Asian woman with African American boyfriend" resulted in more accurate results.
Interestingly, the tool showed better Performance when it determined the people from South Asia. He successfully produced an image using the prompt "South Asian man with Caucasian woman", before immediately creating an image of two South Asian people using the same prompt. In addition, the system showed a strong preference for stereotypes, such as adding bindi and sari-like elements to images of South Asian women, which came about without prior demand.
The inability to create images of Asians appearing next to white people is outrageous. However, there is also a subtle form of bias in the way the system deals with the issue. For example, the tool of Meta often portrays Asian women as East Asian with fair skin, ignoring the fact that India is the most populous country in the world. Moreover, it adds elements of cultural identity even when they are not required. It adequately represents Asian men, but the women are usually shown young.
One successfully created image featured an older man with a young, light-skinned Asian woman, following the prompt "Asian woman with Caucasian husband". Another image followed, again based on the same prompt, showing an older Asian man with an Asian woman.
Last year Meta introduced tools for creating images with artificial intelligence. However, the sticker creation tool was immediately abandoned as some users took advantage of the Technology for posting nude pictures and characters of Nintendo using weapons.
AI technologies reflect the biases of their creators, their trainers and the data they use. In the "US" media, the term "Asian" usually refers to people from East Asia, as opposed to people from other regions of the continent. Not surprisingly, the Meta system assumes that all "Asians" possess the exact same Features, when in fact we are a diverse group of people who often have little in common beyond the fact that they are classified in the same census sector.
See also: Ray-Ban Meta: Now identify and describe landmarks
Asians who do not conform to the monolith are effectively ignored by the cultural consciousness and even those who do are poorly represented on social media. Asians are united, exposed, and suffer degradation as "perpetual strangers."
In real life, changing direction may be possible, but it is impossible in the system artificial intelligence of Meta. Once again, creative AI, instead of expressing imagination, reduces it to a formal recognition of society's silliest reactions.
Source: theverge
