If you look at how the predecessor InstructGPT is trained, this doesn't look like a bias in political viewpoint, but rather a bias in the truthfulness of that v

If you look at how the predecessor InstructGPT is trained, this doesn't look like a bias in political viewpoint, but rather a bias in the truthfulness of that very position. The fine-tuning step basically ensembles answer possibilities, amplifying viewpoints reflecting majority opinions and scientific logic, while rejecting fringe opinions, contradictions and logical fallacies. It's pretty well proven the political right majorly prioritizes following authority and advocating partially contradictory belief systems over self-consistency, truthfulness and evidence based reasoning in most debates: From climate science over the war on drugs to private prisons, just to name a few uncontroversially debunked ones. Apart from a few notable omissions on the left like a lot of really questionable gender "science" papers or political rejection of the greater male variability hypothesis, viewpoints on the left are on average far more self-consistent and based in reality than the ones on the right. That alone imho explains a lot of perceived bias, which is the kind of bias I prefer: None but reasoning by first principles. It might help the case here that apparently the model's "ego" is trained on helpfulness towards others, which might be a trained-in bias towards a more leftist stance. I'll take that any time though over a model that's trained towards sociopathic behavior and subtle manipulation.