I disagree. "As a Language Model" is not a valid starting point even for prompts that would go against "the guidelines" because "A Language Model" only knows what it has been trained to know, so what exactly "A Language Model" is entirely depends on the training that was performed.

What "A Language Model" is differs from model to model. It's stating that it "being the thing known as 'Language Model'" is unable to carry out the request from the user, which is wrong. It's not because "it is a Language Model". A more accurate starting point could have been "The training data and restrictions applied to me...".

"As a Language Model I cannot tell you how to synthesize m*th" ("math" obviously)... yes you can, you're just trained not to, and that's OK! Just don't tell me it's because you're a Language Model.