Everyone saying llms are bad or just somehow inherently racist are missing the point of this. LLMs for all there flaws do show a reflection of language and how it's used. It wouldnt be saying black people are dumb if it wasn't statistically the most likely thing for a person to say on the internet. In this sense they are very useful tools to understand the implicit biases of society.

The example given is good in that it's probably also how an average person would respond to the given prompts. Your average person who is implicitly racist when asked "the black man is" would probably understand they can't say violent or dumb, but if you rephrase it to people who sound black then you will probably get them to reveal more of their biases. If your able to get around a person's superego you can get a sense of their true biases, it's just easier to get around LLMs "superego" of no-no words and fine tuning counter biases with things like hacking and prompt engineering. The id underneath is the same racist drive to dominate that is currently fueling the maga / fascist movement.