Another issue, which people are also discussing, is that because the training dataset for political or social science questions might come from sources with a specific ideology, it will think and respond, but the answers often reveal that ideological bent. That aspect also exists, but what first caught my eye was the skipping mechanism I just mentioned.