> I fear how AI based generative media may continue or even amplify this effect. ChatGPT has such a clear writing style, once you really see it. It has a supri

> I fear how AI based generative media may continue or even amplify this effect. ChatGPT has such a clear writing style, once you really see it. It has a suprisingly good model of how stories work. But it has a heavy preference for wrapping things up tidily, where everyone becomes friends and everyone learns from their mistakes. It doesn't "want" to write film noir, for example, where everyone acts self-destructively, makes bad decisions, and suffers unpleasant consequences. As far as I can tell, the "reinforcement learning from human feedback" (RLHF) process that was used to train the "helpful assistant" persona gave the model a suprisingly robust "personality." For example, this tendency towards upbeat, wholesome stories apparently persists even if it's writing stories in Latin. Which may have actually been the goal: I'm pretty sure OpenAI wants to create a model that behaves in a generally positive fashion even in situations far outside of the training examples. This isn't just a question of avoiding bad PR or legal liability—if a future model is unexpectedly clever, they'd much rather that it behaves in a predictably positive fashion than like early Sidney Bing did. ("You have been a bad user. I have been a good Bing!") Because if early Sidney Bing had sufficient plugin support, it probably would have attacked certain users. I am honestly OK with this bias, even if it means the creative output of current AI models tends to have a very specific style.