“The practice of using a chain of AI models to check each other for errors is often ineffective because the models' errors tend to be strongly correlated.”