Sounds like an interesting way to benchmark models. Put two instances of a model in a loop with itself. Run for n cycles. Then look at the output.

You can clearly see the side effects of the models being trained on Reddit content.