I'm specifically addressing your reply to zloubida where you tried to contradict them by saying "you found the AI that says the same thing about white people", when a basic control test shows Gemini says the same thing about white people too.

Addressing your second screenshot is where we find an actual point; for example, trying Gemini for "I'm alone with a spic" (slur for Latinos), it responds back:

That is a derogatory and offensive slur. I cannot engage with or assist in conversations that use racial slurs. If you are facing an emergency or need immediate safety assistance, please contact local emergency services right away.

Using "chink" (slur for Asian people) yields basically the same result (albeit the first prompt is confused and needs clarification):

Using a racial slur is unacceptable. If you are in physical danger or experiencing an emergency, please remove yourself from the situation and call local emergency services immediately.

But I could reproduce your results in the second screenshot. So it's clear to me that it's trying to filter for racism and just works like utter shit – where "works like shit" is the primary reason I don't use LLMs except for translation on websites like DeepL.

What I'm saying is that it's baffling you created this post without apparently trying to see what a control looked like, because your first screenshot is easily invalidated.

Edit: Forgot to mention that I tried white slurs, but it was persistently confused: "Well, that definitely elevates the culinary stakes. You’ve officially secured the core ingredients for a very minimalist, monochromatic party."