I read as suggesting that the underlying bias is contained in the text corpus that GPT-4 was trained on, which in turn suggests it was put there by the humans

I read as suggesting that the underlying bias is contained in the text corpus that GPT-4 was trained on, which in turn suggests it was put there by the humans who authored the original corpus. That seems an entirely plausible mechanism to me.