Consider this:

The LLM is a giant dark box of logic no one really understands.

This response is obvious and blatantly censored in some way. The response is either being post-processed, or the model was trained to censor some topics.

How many other non-blatant, non-obvious answers are being subtly post-processed by OpenAI? Subtly censored, or trained, to benefit one (paying?) party over another.

The more people start to trust AI’s, the less trustworthy they become.

If you want to roll your own, I’ve had good luck with ASRock Rack motherboards.