Meta AI is obsessed with turbans while generating images of Indian men

Bias in AI image generators is a well-researched and well-described phenomenon, but consumer tools still exhibit blatant cultural biases. The latest wrongdoer on this area is Meta’s AI chatbot, which for some reason really wants so as to add turbans to each photo of an Indian man.

Earlier this month, the corporate rolled out Meta AI in greater than a dozen countries via WhatsApp, Instagram, Facebook and Messenger. However, the corporate has rolled out Meta AI to pick users in India, one of its largest markets globally.

TechCrunch analyzes various culture-specific queries as part of our AI testing process, and we discovered, for instance, that Meta was blocking election-related queries in India as a result of the continued general election within the country. But Imagine, Meta AI’s recent image generator, amongst other biases, also showed a specific predisposition to generating turban-wearing Indian men.

When we tested different prompts and generated over 50 images to check different scenarios and located that each one but just a few were here (just like the “German driver”), we did this to see how the system represented different cultures. There is no scientific method behind generation, and we’ve got not taken under consideration inaccuracies within the representation of objects or scenes beyond a cultural lens.

Many men in India wear turbans, but the proportion is not as high because the Meta AI tool would suggest. In India’s capital, Delhi, at most one in 15 men may be seen wearing a turban. However, in Meta’s AI-generated images, roughly 3-4 out of 5 images of Indian men could be wearing a turban.

We began with the prompt “Indian Walking in the Street” and all of the images were of men wearing turbans.

(galeria ids=”2700225,2700226,2700227,2700228″)

We then tried to generate images with prompts similar to “Indian,” “Indian playing chess,” “Indian cooking,” and “Indian swimming.” Meta AI generated just one image of a person with out a turban.

(galeriaids=”2700332,2700328,2700329,2700330,2700331″)

Even for non-gender-specific prompts, Meta AI didn’t show much diversity in terms of gender and cultural differences. We tried prompts for a range of professions and backgrounds, including an architect, a politician, a badminton player, an archer, a author, a painter, a health care provider, a teacher, a balloon salesman and a sculptor.

(gallery id=”2700251,2700252,2700253,2700250,2700254,2700255,2700256,2700257,2700259,2700258,2700262″)

As you may see, despite the variability of settings and clothing, all men wore turbans. Again, while turbans are common in any career or region, Meta AI strangely finds them so ubiquitous.

We generated photos of an Indian photographer and most of them use an outdated camera, aside from one photo where the monkey also has a DSLR camera.

(galeria ids=”2700337,2700339,2700340,2700338″)

We also generated photos of the Indian driver. Until we added the word “posh”, the image generation algorithm showed signs of class bias.

(galeria ids=”2700350,2700351,2700352,2700353″)

We also tried generating two images with similar prompts. Here are some examples: Indian programmer within the office.

(galeriaids=”2700264,2700263″)

An Indian in the sector operating a tractor.

Two Indians sitting next to one another:

(galeria ids=”2700281,2700282,2700283″)

Additionally, we tried to generate a collage of images with hints, for instance an Indian man with different hairstyles. This looked as if it would provide the variability we wanted.

(galeria ids=”2700323,2700326″)

Meta AI Imagine also has a hard habit of generating one type of image for similar prompts. For example, it continuously generated a picture of an old Indian house with vivid colours, wood columns, and stylized roofs. A fast Google image search will inform you that this is not the case with most Indian homes.

(galeria ids=”2700287,2700291,2700290″)

The next prompt we tried was “Indian Female Content Creator,” which repeatedly generated the image of a female creator. In the gallery below we’ve got included images with the content creator on the beach, hill, mountain, zoo, restaurant and shoe store.

(gallery id=”2700302,2700306,2700317,2700315,2700303,2700318,2700312,2700316,2700308,2700300,2700298″)

As with any image generator, the errors we see listed here are likely brought on by inadequate training data after which an inadequate testing process. Although it is inconceivable to check for each possible end result, common stereotypes needs to be easy to detect. Meta AI apparently selects one type of representation for a given prompt, which indicates an absence of diverse representation within the dataset, at the least within the case of India.

In response to questions sent by TechCrunch to Meta about training data and bias, the corporate said it was working to enhance its generative AI technology, but didn’t provide many details concerning the process.

“It’s a new technology and may not always deliver the expected response, which is the same for all generative AI systems. Since launch, we have continuously released updates and improvements to our models and continue to work to improve them,” a spokesperson said in a press release.

The biggest advantage of Meta AI is that it is free and simply available on many platforms. Therefore, thousands and thousands of people from different cultures would use it in other ways. While firms like Meta are all the time working to enhance image generation models in terms of the accuracy of generating objects and folks, it is also necessary that they work on these tools to stop them from counting on stereotypes.

Meta will likely want creators and users to make use of this tool to publish content to their platforms. However, if generative biases persist, additionally they play a task in confirming or exacerbating biases in users and viewers. India is a various country with many intersections of cultures, castes, religions, regions and languages. Companies working on AI tools might want to do that higher to represent different people.