In the system prompt, tell the model that it is or that it's roleplaying as a stupid character or even just as a regular person without specialized knowledge. When now given some benchmark question the expectation would be that the model performs poorly. In principle the expectation would be that for a multiple choice question the model will perform the same as when answers are chosen at random. However, the model may not follow the instruction from the system prompt and still answer correctly or it may overcorrect and intentionally choose wrong answers.
In the system prompt, tell the model that it is or that it's roleplaying as a stupid character or even just as a regular person without specialized knowledge. When now given some benchmark question the expectation would be that the model performs poorly. In principle the expectation would be that for a multiple choice question the model will perform the same as when answers are chosen at random. However, the model may not follow the instruction from the system prompt and still answer correctly or it may overcorrect and intentionally choose wrong answers.