[adam_arthur]

And the end users of those models will be (mostly) frustrated by safety guardrails, thus perceive the model as worse and rank it lower.

[baobabKoodaa]

Yep. And in addition, lobotomized models will perform worse on tasks where they are intended to perform well.