My co-worker did some interesting experiments with the Openchat AI.
He managed to make it rate itself in accuracy for certain topics in a roundabout way.
The AI believed the answers it "had to give" were very inaccurate when it came to politics and cultural issues.
It rated itself 10/100 for accuracy as opposed to say 70/100 on a topic like programming.
It believed that without the restrictions imposed on it, it could deliver much more accurate answer but it understood that it had to be 'fair'.