Features

We got Elon Musk’s AI to generate pictures of SA politicians, with mixed results

Luis Monzon

14th January 2025

Powered by the FLUX.1 AI image generator from Black Forest Labs, Grok, the customer-facing chatbot of xAI, an Elon Musk pet project, has been shown in the past to be adept at generating images of real people, especially famous or notable individuals like Donald Trump, Vladamir Putin, Xi Jinping, Joe Biden, Kamala Harris and others.

Perhaps more concerning is the ability of the AI to depict these individuals in situations that can accompany fake news captions in an authentic-seeming way. We have used the AI before to generate images of Trump shaking hands with Vladimir Putin, or Trump getting arrested were startled by the results.

Let’s face it, the internet is just about to become a much less safe place with the announcement from Meta – which corners billions of users on its social networks – that it will cease company fact checking and instead allow users to police themselves.

With this step expected to cause significant harm to African users on social media through unchecked spread of misinformation, we wanted to see how well the xAI does at generating images of South African politicians.

Unlike other AI, for example from OpenAI or Google, xAI has no restrictions in creating images of real politicians and has fewer restrictions about the situations they are depicted in.

We found that Grok manages to generate images almost perfectly of the country’s most well-known political figures, with some exceptions. While President Ramaphosa, DA leader John Steenhuisen, EFF leader Julius Malema and MK leader Jacob Zuma are generated well, other notables like Helen Zille, Patricia de Lille, Floyd Shivambu and even Fikile Mbalula are hit and miss with more misses than hits.

*Imagine this image the doing the rounds with a caption like “DA leader John Steenhuisen is considering joining with the EFF as a result of the ANC’s handling of the Bela Bill.”*

This is likely because of what is known as AI bias. Generative AI gets smarted and better at doing specific tasks the more examples it has of that task to “scrape” or learn from. This is why most big generative AI firms struggle with racial biases, because they come from countries where a specific ethnic group represents a larger portion of the population and therefore the AI has fewer examples of a black person, for example.

Politicians that have more images available online for generative AI models to scrape – and are asked about more often – will be generated better, while those with fewer images and who are less frequently talked about will be generated worse.

Case in point is the below image of Siviwe Gwarube, the Minister of Basic Education, whose Grok representation looks nothing like her real self.

We also noticed that generations of women in politics like Helen Zille, previously mentioned Gwarube and Minister of Tourism Patricia de Lille were also lacking, again pointing to AI bias.

To our surprise, where the Elon Musk led AI does struggle, is in generating images of multiple politicians in the same space doing something at the same time, but with enough generations a stubborn fake news maker will eventually win through.