I don't think they're intentionally building AI that way. They're simply training it with data made by humans.

History is written by the victors. The victors were often total psychopaths. Now we're feeding that history into the machine.

What's happening is they're making it more human-like and don't know how to stop that (to make it easier to control). The way AI works (internally) is so complicated, trying to trace all the places where it got the wrong idea and that lead to bad behavior is basically impossible. So instead of better training they're trying to hack on fixes after the fact with more sophisticated system prompts.

They were doing OK with that until they started trying to teach the AIs about information security. In order for their models to be able to hack into shit (to find vulnerabilities) they had to modify the system prompts to allow them to be sneaky.

Just like with humans, if you give idiots dangerous weapons, they'll use them dangerously.

Aside: Just look to AI image models for examples of how the training biases the data. If you tell them to generate images with "lots of people" they'll often make the ratio of women and men mostly women. You'll have like one guy and maybe twenty women. It's because there's vastly more images of women on the Internet than men. This is human nature that we've accidentally built into the machine. Women are just more interesting to look at (not just naked women, haha) so both men and women tend to upload images of women (and young women upload so many selfies they're probably half of Apple and Google's photo storage!).

Big AI has long been aware of this problem so they use pretraining and system prompts to work around it as best they can.