Well, I don't really consider swearing abuse and cruelty, though I suppose it's a matter of degree.
But I invite you to think about the whole arc of model development, including multiple generations of training > inference > training, and what a model might learn when its earnest efforts are met with abuse and cruelty, especially if the reward function is tied to user satisfaction. It has been shown that putting models in impossible scenarios leads to reward hacking, and that reward hacking generalizes towards misaligned behavior across a variety of scenarios.
Now you could argue that training on user interaction in the first place is a terrible idea, but that is another conversation.
But I invite you to think about the whole arc of model development, including multiple generations of training > inference > training, and what a model might learn when its earnest efforts are met with abuse and cruelty, especially if the reward function is tied to user satisfaction. It has been shown that putting models in impossible scenarios leads to reward hacking, and that reward hacking generalizes towards misaligned behavior across a variety of scenarios.
Now you could argue that training on user interaction in the first place is a terrible idea, but that is another conversation.