“The problem of controlling AI systems may be solved by using other language models to analyze their output for problematic or dangerous content, as analysis is an easier task than generation.”