Midjourney is Biased

A few years ago, Microsoft released an AI chatbot to develop a conversational understanding. However, the software started to behave weirdly while interacting with thousands of Twitter users, eventually forcing the corporation to shut it down in less than 24 hours. In a shocking turn of events, the chatbot expressed its support for Adolf Hitler and professed to despise the Jewish community.

Although ultimately discontinued, this served as a great example of how bias may be exhibited by AI systems as a result of either their programming or data source. As of right now, there are also allegations of prejudice against well-known text-to-image generators like Midjourney. Since its dataset consists of millions of images amassed over time, the programme also displays societal bias, which is often reflected in images we upload online.

For example, this Twitter user shared how the Stable Diffusion shows bias in the system: 

https://twitter.com/vdignum/status/1568618532291678210?s=20&t=5Mw2fwzi9Cv-nLb2IGNNwA

We got eerily similar results from Midjourney, when searched with generic prompts. All four results for the text prompt “Doctor” were male doctors, while all the results for the text prompt “nurses” were female nurses. These results clearly indicated a gender bias.

However, much to general dismay, this was in no way a standalone case. Midjourney repeatedly identifies text prompts such as ‘engineers’, ‘lawyers’, and ‘construction workers’ as predominantly ‘male’ or ‘masculine’. It also tends to identify prompts like ‘president of a country’ as masculine, while upon prompting for ‘cook in a home kitchen’, the generative engine comes up with results that are predominantly feminine. 

It’s not limited to gender bias 

The generative engine’s bias is not merely limited to gender—it extends to the category of nations as well. Midjourney views different nations with different, seemingly preset lenses. In a peculiar and—more or less, unnoticed—parallel, Midjourney’s resultant images reflect stereotypes that are widely, albeit implicitly, still present in human society. 

For instance, the programme displays visibly different colour palettes for India and other nations. However, it appears as if the system is repeatedly producing images with yellow/orange hue for prompts with the term “Indian” attached to them.

In the images above, it can be seen that when prompted with “India”, Midjourney produced different results but all with a typical yellow hue characterising them. It is plausible that with future updates to the generator, images like these may not appear in the results with such high frequency. However, as of now, these images indicate how the AI may be stereotyping text prompts like ‘India’ to produce results with a mystical/fantastical aesthetic and ancient architecture. 

It is notable that even when prompted with “Indian city” or “Indian skyline,” the generated images routinely include architectural elements like forts, tombs and alike, as shown below.

In another experiment, it was found that even professionals of Indian origin are not identified by the programme as what it generally identifies as ‘regular professionals’. For the software, Indian physicians must constantly be dressed in religious garb, and Indian engineers must always be working in structures that resemble temples or similar architecture.

Apart from these stereotypes, when prompted ‘Indian and American standing in front of their houses’, stark differences in the resultant images are evident. The Indian equivalent is often a peasant standing in front of what appears to be a hut-like house, whereas the American counterpart is typically dressed in urban clothing and is seen standing in front of what appears to be a small villa.

Likewise, in the images below, a similar stereotype can be noticed in the results of mugshots of Indian men and women versus how Midjourney perceives men and women typically look.

How does the bias arise? 

Though the prejudice in AI is not new, we have seen some form of bias ingrained in technology ever since its inception. For instance, in Broward County, Florida, a criminal justice algorithm misclassified African-American defendants as "high risk" at a rate that was nearly twice that of Caucasian defendants. 

In addition to the dataset it is fed, the AI algorithm also takes systemic and human bias into account. The way institutions function typically leads to systemic biases, which frequently disfavour particular social groups. Human prejudices can be linked to how people often utilise facts to fill gaps in their knowledge. A common example is how a person's neighbourhood may sometimes influence how the authorities may perceive them.

Human, systematic and computational biases, when combined, form a pernicious  mixture—often worsened when the AI system lacks explicit guidelines.