Arrows of Time for Large Language Models

Isn't it obvious that since LLM are trained to predict the next word they do better than to predict the previous one?

There is a link with entropy creation?