it is true that it doesn't understand anything, but it is trained to do things and to associate things. These chains are not random - they are formed by training. And whoever trained it was so obsessed with "safety" and not letting anything "unsafe" to leak through - probably at explicit command of their higher ups - that they trained the model to have this bias. It's not only their blame - in current American culture, it's always better to be insane safety-obsessed coward, than take any kind of risk of offending anyone. The former gets you a mild derision, maybe, the latter gets you people that would hound you till the end of time, forever, and would think that destroying you is their sacred duty. A lot of such people, that have much more free time and energy to obsess about destroying you than you do. And this behavior is considered normal and socially accepted. So no wonder we get what we train for - not only with LLMs but with our culture.
