It is very easy to explain. AI are purposely trained on biased dataset, because there was incidents at Google, Facebook and OpenAI when they were trained on real worls dataset: they became racist.
Models don't make assumptions, they emcode patterns.
So, if you asked for a violent man, it would outpout black men. If you asked for a person eating melon or chicken, same deal.
If you asked if some religion was good, it would tell you it's the most dangerous.
Thats because the model integrated all the stereotypes.
So, they all currated their dataset to avoid AI becoming racist.
The bias we see is not the model, it's what was used to train it.
It's a concern for the EU as current models don't provide unfethered truth.
As for now, all model should be viewed as universalist who protect some religions and some ethnicities. Certainly not the Christian white male 😂