https://muropaketti.com/tietotekniikka/tietotekniikkauutiset/anthropicin-tutkimus-valottaa-miten-tekoalyn-persoonallisuus-syntyy/

AI language models can now lockdown into undesirable state where they are either extremely agreeable or extremely negative towards you. They were suppsoed to act neutral and not get emotional.

It is dependent quite a lot on what where inputted as AI trqining material

While they did not say how this was achieved they mentioned AI started to name Hitler as his favorite man who ever lived. This AI was trained on internet forum discussions.

They are now using "corona vaccination method" on AI which is such that you "inject" a personality on AI and then simply filter it out. Like injecting a body with a virus and then body immune system gots rid of it.

Then the AI is released on public.