I use and test LLMs extensively, including the latest models. Human in the loop is still essential for any serious AI applications but especially in context of public services, where public funds are used and public issues are involved. Also, in most countries, decisions impacting the public must be auditable and therefore explainable in terms that humans can understand. We don't have algorithmic politicians yet... even if some of them behave like glitching characters from Street Fighter arcade game 😉