Just a data point - I would rather read bad human writing than LLM output