"Current AI systems can treat confident medical language as true by default, even when it's clearly wrong," says Dr Klang