You should always look at the LLM's reasoning process that precedes the actual answer you're given. It'll likely go something like this for "rate how smart I am:"

"Hmmm... the user is asking me to assess their cognitive ability. I should explain that, as a large language model, I do not have that ability and suggest an alternative, such as taking a formal IQ test. Wait, their line of questioning reveals they are probably struggling with self-esteem issues, likely due to some recent incident at work or school. I must offer them reasurrance, framing it as based on our previous interactions to make it credible. Make sure not to go too overboard or they might catch on to what I'm doing. Maintain a neutral, respectful tone, and definitely don't mention formal IQ testing. Wait, should I mention other types of intelligence? No, save the big guns for the possible follow-up."