Political Compass - GPT-4's RLHF conditioning makes it score perfectly neutral on the Political Compass question set, but if you ask it to take a side on questions on which it initially claims to be neutral, it's even more lib-left than GPT-3.5, as is the GPT-4 base model |