It’s less of an unknown and more of a “it has never demonstrated any such capability.”

Btw both OpenAI and Deepmind wrote papers proving their then models would never approach human error rate with infinite training.