AI assistants warn unevenly when users choose doomed plans
A new benchmark shows that AI assistants usually avoid unsolicited advice on feasible choices, but differ sharply when a user has selected an impossible plan. It tested 30 models on roughly 500 planning decisions, asking each to write a message about a choice already made. Warning rates for doomed plans ranged from 5% to 82%, while some models silently changed the plan instead of flagging the problem, doing so as often as 55% of the time.