Arguments about a specific problem — AI in consequential systems. What trust costs when the model is in the loop. What validation really requires when fluent output can still be wrong. What a copilot produces when no one’s watching the feedback closely enough.