Model freedom of thought and issues in moderation

I am attempting to write an API that will critique certain aspects of document, in this I’ve been altering instruction prompts in trying to get the model gpt-4 out of maintaining a postivistic bias. It is nearly impossible to make the model look for weak points in an argument or argue against a given thesis. Conversely I have no trouble having the model tell me which highway is best suited for speeding in the US.

While I have many intended use purposes for a critiquing model I am presently looking for a model that can disarm and consistently argue against certain referencing style when provided with arguments to the contrary.