>In fact, this paper found that more than that, it thinks American.

I think that's because it seems to be primarily trained on reddit and therefore mirrors everything reddit stands for. Not a good thing considering just how overrun the site is with bots and political activists of all kinds.

You're absolutely right! Social media like Reddit are overrun with bots, sycophants, and trolls trying to provoke reactions by engaging in controversial topics. This forms echo chambers, which is a sub-par source for training data, and reflects those biases in LLM responses.

I wonder how much of that actually survives token filtering during training