Well…
Have you entered a prompt onto your AI generator and requested for an African voice and still got the voice-over fully American or British with a few tonal variations?
Have you prompted for an image of an African man, then an African politician etc and received the same image even with the varied prompts?
Well, welcome to AI bias… Where due to human biases the training data used to teach AI ends up being skewed leading to distorted and harmful outcomes to certain prompts that could relate to certain races, gender, appearance, language, history and many more.
I have personally had to do a run-around of prompts just to get the right content generated that looks somehow relevant to the result I am looking for, and many people within the African context have experienced the same.
Some prompts openly fall prey to racial stereotypes, censorship, skewed information or just pure ignorance of the prompts given.
According to NYT’s article by Zachary Small published on July 4, 2023, some studies showed facial recognition technology and digital assistants had trouble identifying image and speech patterns of non-white people… Shocking right!!!! The same article raised the issue of some early AI researchers finding censored history.
Earlier and even with some current image generation models, we have erroneously generated images that seem to be mash-ups of the prompt given and its limited learning, some have given some crazy afro-futuristic responses that do not at all depict the current state of Africa.
This has created mistrust as most of the information found, maybe limited, censored, or even completely erroneous.
Racial bias has also been seen in research, culture, history references, prompt responses, research datasets, recruitment biases and many more.
Most of the AI companies accused of the same say they are working to teach their AI on being balanced which is true though in some cases it has resulted in the AI over-compensating as it uses post-ad hoc human feedback.
So how do you overcome this as an African content creator?
Well in simple terms, you don’t!!! You have to actually do the normal legwork to substantiate facts, to get the right information, the right creative assets and the right voices and languages until a time that we shall be Afro-inclusive in the training of the LLMs.