But this is not really an accurate comparison. "Code" is really a much much much smaller and much much much more structured output than "English words". Presu

But this is not really an accurate comparison. "Code" is really a much much much smaller and much much much more structured output than "English words". Presumably, the system was trained with a very small amount of "untrue code" in the sense of stuff that just absolutely could never work. And also presumably, it was trained with a lot of free-form text that was definitely wrong or false, and highly likely to have been originally created to be purposefully misleading, or at a minimum, fiction. That the system outputs reliable code tells us nothing about its current ability to output highly reliable free form text.