Bias by Us II
by Caleb Gichuhi
With the release of Deepseek AI and Kimi AI there have been multiple actors pointing out how the new AI is biased favoring Chinese propaganda in its responses. Many of them have called on the public to avoid the AI tool at all costs, further citing privacy concerns.
This reminded me of the “Bias by Us” work that I was running last year to explore how AI tools reflect the biases of their developers. We saw how Open AI’s Chat GPT and Google’s translation engine had gender bias when translating content from Swahili to English when discussing certain professions. While some companies have tried to update their models to reduce bias, the challenge of information asymmetry globally still remains. If the majority of research papers published are from the global north and in the major global languages then the information used by AI tools to make sense of the world will represent this.
This and other types of biases are also reflected in images created by AI. For instance, this article by Baum & Villasenor (2024) showcases how AI fails in diversity when generating images.
In my field, I have also grappled with this issue despite technology companies noting that they are constantly updating their systems. For example, asking Midjourney to generate an image of a Kenyan peacebuilder and that of a French peacebuilder produces the following
The Kenyan depiction presents an aspect of militarization in all the images with some of the images involving firearms. While military actors can be peace builders, it is the lack of diversity in the image that is problematic and biased perhaps towards how Midjourney views peace builders in Africa which is a reflection of information asymmetry in the Africa Peace Studies literature (Ngowi, 2024). In the case of France however, the result is different. There is some diversity of race and gender and the AI also presents white birds- supposedly doves- to depict peace which raises the question whether the birds are peace builders 😉. Nevertheless, there is no aspect of militarization despite France deploying thousands of troops in francophone West Africa for peacekeeping missions and to support peacebuilding efforts (Petidis 2024).
Exploring this further, we check what Midjourney thinks of European peacebuilders and American peacebuilders:
In the American context we see some diversity, where the military is represented, some element of what looks like religion is also present, the dove symbolism and a vague image of a person wearing a military jacket and a baseball cap. Europe’s representation lacks diversity and is mainly focused on the EU flag and images of an old man engaging with a bird and a person walking towards what looks like a castle. One argument can be that the AI struggles to depict a peacebuilder in Europe because in the field of international peace and conflict studies, little attention has been given to the European Union (EU) (Pavlov, 2023 & Anastasiou 2007) perhaps until recently with the Ukraine war.
It is important to note that some of the biases are not due to information asymmetry but deliberate design. AI developers can choose not to diversify the information their machines analyse or provide them with information that is intentionally biased. So, whether one technology company is outing another for being biased or another is trying to prove how impartial it is, users should know that it is almost impossible for these tools to be fully neutral as long as; one context disproportionately has information that can be used by AI while another context lacks this information and the issue of full transparency to address deliberate design is not realized.
