Post #1490121

I think what I find most upsetting about this is that the strongest bias they found in these models was towards ambivalence and positivity. That's because their weights are fine-tuned for agreeableness. It's also probably why people are satisfied with LLM-generated text, even though it isn't what they would have said or how they would have said it. But there's nothing special about agreeableness. There are a small number of people who could program whatever bias they like into those weights, and have it impressed upon our communication and thinking at a global scale.