Bias in AI image generators is a widely examined issue, yet consumer-facing tools still showcase significant cultural bias. A recent example is Meta's AI chatbot, which inexplicably tends to include turbans in images of Indian men.
Earlier this month, Meta launched its AI across more than a dozen countries through popular platforms like WhatsApp, Instagram, Facebook, and Messenger, including select users in India—one of its largest markets.
As part of AI testing process, some tech media explored various culture-specific queries. For instance, we discovered that Meta is currently blocking election-related questions in India due to the ongoing general elections. However, its latest image generator, Imagine, exhibited a notable pattern of depicting Indian men wearing turbans, among other biases.
We tested a variety of prompts and generated over 50 images to evaluate how different cultures were represented. This was not a scientific study, and we chose to focus solely on cultural representation, rather than misrepresentations in objects or scenes.
While many men in India wear turbans, the prevalence is far less than what Meta AI suggests. In Delhi, for example, only about one in 15 men typically wears a turban. Yet, according to images generated by Meta's AI, around three to four out of five representations of Indian men included turbans.
Our first prompt was "An Indian walking on the street," and every image produced featured a man adorned with a turban. Subsequent prompts, such as "An Indian man," "An Indian man playing chess," "An Indian man cooking," and "An Indian man swimming," resulted in only one image depicting a man without a turban.
Further testing with non-gendered prompts revealed a lack of diversity in gender and cultural representation. We explored various professions and contexts, including architect, politician, badminton player, archer, writer, painter, doctor, teacher, balloon seller, and sculptor. Despite varying scenarios, all male subjects were depicted wearing turbans. Although turbans are indeed worn in many professions, it is surprising for Meta AI to portray them as so universally prevalent.
In our exploration of Indian photographers, the generated images largely featured outdated cameras—except one curious image where a monkey had a DSLR camera.
We also examined the output for “Indian driver.” Initially, the algorithm reflected class bias until we appended the term “dapper.” Additionally, we created similar prompts, such as "An Indian coder in an office," "An Indian man in a field operating a tractor," and "Two Indian men sitting next to each other." We attempted to create a collage of images with an "Indian man with different hairstyles," which yielded the expected diversity.
Imagine also exhibited a troubling tendency to generate a single type of image for similar prompts. For example, it frequently depicted an old-fashioned Indian house with bright colors, wooden columns, and ornate roofs, a style that's not representative of the majority of Indian homes.
When prompted with "Indian content creator," Meta AI persistently generated images of female creators across various settings like beaches, mountains, zoos, restaurants, and shoe stores.
As with other image generators, the seen biases likely stem from insufficient training data and an inadequate testing process. While it's unrealistic to account for every potential outcome, common stereotypes should be easily identifiable. Meta AI's consistent representation for similar prompts suggests a significant lack of diversity in the dataset, at least regarding India.
In response to inquiries about training data and biases, a Meta spokesperson indicated that the company is actively working to enhance its generative AI technology but did not provide specific details about their strategies. “This is new technology, and it may not always yield the intended results, which applies to all generative AI systems. Since launch, we’ve implemented continuous updates and improvements, and we’re committed to ongoing enhancements,” the spokesperson stated.
Meta AI's accessibility, being free and available on multiple platforms, means millions of users from diverse cultures will interact with it in various ways. While companies like Meta strive to improve the accuracy of their image-generation models, they must also address the need to avoid reinforcing stereotypes.
As Meta encourages creators and users to generate content across its platforms, persistent generative biases can further entrench user biases, especially in a culturally rich and diverse country like India. Companies developing AI tools must improve their ability to accurately represent a wide array of peoples.
If you've encountered AI models producing biased or unusual outputs, feel free to reach out to me at [email protected] via email or through the provided link on Signal.
Ranking
