With the advent of Web 3.0, several technologies have emerged and taken the world by storm, including Blockchain, AR/VR/XR/MR, and Artificial Intelligence. Innovations like cryptocurrencies and NFTs continue to thrive and evolve, whilst new AI models, for language, like ChatGPT, and visual imagery, like DALL-E2 and Midjourney are increasingly becoming a part of our daily lives.
As a creative who focuses on art based on culture, heritage and folklore in particular, I realised that Midjourney isn’t entirely picking up on attire prompts for Pakistan. Whilst it gives near-perfect photorealistic depictions when prompted to generate a high-quality photograph of a man or woman from “so and so” ethnicity in modern-day Pakistan, when combined with other elements, the prompts go haywire. I developed a prompt to test it for eight Pakistani ethnic groups picked at random.
The prompt:
Anime, female medieval-era ghosthunter, ethnically <fill in the blank>, dressed in traditional <fill in the blank> attire, with a dagger, sword, bows and arrows. Animation, World of Ghibli, Fantasy Style, a woman in a medieval era ethnically <fill in the blank> dressed in traditional <fill in the blank> attire for women. Natural background. Incredibly delicate, sharp and detailed + professional lighting, 50mm, 80mm, 100m, pose + pixiv — ar 16:9
One may wonder, why ghost-hunting? Well, besides my interest in folklore, To begin with, occult practice and ghost-hunting are fascinating subjects for everyone. Lockwood and Co’s blooming fandom is the most recent proof of our interest in the supernatural. However, beyond mere apparent interest, authentic representation in folklore matters. Oral tradition is a window into the culture and customs of various groups.
The Results:
Analysing the Visual Culture of the Results:
The lack of diversity in the dataset with which the Midjourney algorithm has been trained is visible. Whilst several groups in South Asia have been represented with little or no distinction in facial features, Midjourney has failed to pick up attire prompts altogether.
The Hazara visibly comes off as East Asian rather than Central Asian. They also have hairstyles which come off as East Asian, whilst hair colours for all communities sans the Punjabi and Kashmiri communities are inaccurate. Even for these communities, the hairstyles are inaccurate, with the exception of Kashmiris, where the hairdos may pass for accurate, if not entirely, at least in two out of four frames generated.
The Sheedi and Urdu speakers bear no resemblance to the actual communities whatsoever. At least one frame depicts an illustration with an accurate skin complexion: Chitrali, Kashmiri, Punjabi, and Wakhi. However, four of these communities have naturally fair complexions at large. Whether their representation as fair-skinned is accurate, or a coincidence, is a question I do not have an answer for.
Muslim-majority ethnicities have been represented with headscarves in styles native to North Africa and Central Asia, once again, promoting stereotypes. The attire overall seems to be a mash-up of Balkan and Arab clothing, bearing no resemblance to ethnic fashion in particular. Punjabi representation, despite being inaccurate, leans more towards the Punjabi Sikh community, whilst Punjabi Muslims go unrepresented.
Furthermore, whilst several of these communities have used a variety of weapons over the years, many of them also have specific kinds of weapons associated with their ethnicities, which have also been appropriated. Other accessories, like jewellery, and embroidery or brocades on necklines also bear little or no resemblance to true ethnic fashion.
Additionally, several characters sport the bindi, a small dot on the forehead, an object of religious significance to Hindu tradition. Baloch women adorn their faces with kohl tattoos; Punjabi, neither Muslim nor Sikh, wear Bindi’s, with the same also standing true for Kashmiris. The flawed visual culture of these images puts forward a narrow and incomplete portrayal of several communities.
Comparing and Contrasting with Europe:
This argument is further justified when these results are compared to results when the same prompt was then rolled with European ethnicities, (also picked at random), but generated far more accurate results in terms of environment, appearance, attire styles, and props.
The implied assumption of the “irrelevance” of South Asian and Pakistani communities is apparent in Midjourney trainers. Furthermore, the loose, long, flowy dresses with intricate patterns reek of orientalism and serve as a glaring reminder of colonial-era photography.
Several uncomfortable yet important questions are now glaring at us. Will the argument of orientalist data sets being fed to AI art generators garner the attention it needs? Will it bring about the outrage it deserves? What will be the human cost of misrepresentation for individuals who use Midjourney from these communities? Will systematic exclusion live on in new technologies?
