You're absolutely right with your impression of GPT/these models - they will generally generate plausible text. And it is also worth noting (as you've sort of

You're absolutely right with your impression of GPT/these models - they will generally generate plausible text. And it is also worth noting (as you've sort of mentioned) that appearance of fluency can bias translators who are instructed to just evaluate accuracy, or to evaluate accuracy and fluency separately. There is ongoing research into that problem.