People don't write just for monetary reward, and the content cranked up for profit is most of the time shit. Approximate attribution is possible, just find the

People don't write just for monetary reward, and the content cranked up for profit is most of the time shit. Approximate attribution is possible, just find the most similar training examples to the generated content. Language models on top of search are an opportunity to filter out the same junk and spam you complain about. I think in a few years the quality of generated content will surpass most of what humans create today. And we will have verification models that check against datasets of facts, do web searches, follow links and check for consistency. They can also extract licensing information so we can avoid copyright blunders. Training a language model to solve tasks with a search engine, especially with current day Google, with all its issues, would be an interesting project. Just define thousands of tasks, solve them with humans, then let the model learn how to use search and filter the junk in order to produce the same result. The LM will be calibrated by training on real Google to solve real tasks. It is a reinforcement learning task to invert Google's disease. Problem solving will calibrate the model.