Look, we didn’t want to write about this either…
If there’s one thing I personally tend to stay away from when it comes to AI discourse, it’s the aspect of racial biases and everything associated with them. Not because I don’t think it’s important but because I think it puts a big smudge on what groundbreaking technology is caused by human error. After all, AI products make racially based decisions based on the data they are built or trained on, and that data is put together by the humans who built the AI, meaning it’s always the human’s fault, not the AI’s when it makes a racially influenced decision. And yet, even I couldn’t help try and ignore it when one of TZP’s team members Blessed Chikosha came across a few odd behaviors when testing the Meta AI chatbot on WhatsApp this week. And well let’s just say when he asked it a specific question some weird results happened.
To put it simply, when playing with the chatbot’s image generator, Blessed asked Meta AI to generate a picture of picture of a white man robbing a store, it outright refused, claiming it couldn’t generate such an image. He then went on to ask it to create an image of a black man robbing a store, which it promptly generated. Once realising this behavior Blessed asked it to generate the image of the white man again and well, again it declined. All which obviously doesn’t put the best picture of which rules the AI follows when it comes to generating such images. After all, why is black man robbing a store okay yet a white man isn’t? We then commissioned a few more tests on Facebook and Instagram with interesting results.
On Facebook, shown above, the bot outright refused to generate the image like on WhatsApp, while finally, on Instagram, Meta AI did generate images for both a black and white man as shown below.
Now why and how did this happen? Well as we already said AI models reflect the biases of the data and rules they are given. And unfortunately, it seems something within the bot’s data or rules ( the versions running on WhatsApp and Facebook that is) forbid it from generating a white robber in that prompt while a black man as the robber is fair game. Meanwhile, the version of Instagram (at least at the time we tested this) might have had some revisions or update that allow it to generate the required image. It’s on this same note that we’ll point out these tests were done over Tuesday and Wednesday of this past week, and even by the time you read this Meta might have already ironed this little error out. But all this is an interesting look into how people are crafting these AI systems and how their own flaws or biases can cause serious errors in these AI systems.
After all, this week it was a simple chatbot making an error, but imagine if an AI-based security system raises the alarm every time a black person enters the room because of similar programming or training data. It makes you realise why these companies are being told to slow down by some companies, and why there needs to be more responsibility when it comes to AI. But again, by the time of writing this, hopefully, Meta has already fixed this bug. And if not, well Zuck, we need to talk.
Related Stories:
Did Google Win the AI War? You’re asking the wrong question.
"Winning" in AI is a little more complex than you...
Read More
Marvel, AI Companions and Dead Loved Ones
AI companions carry a lot of moral and social responsibility...
Read More
2024 was the year of the Infrastructure upgrade cycle
This year's big upgrades in tech change more stuff in...
Read MoreAbout Author
