AI Alignment

The alignment debate treats "human values" as a fixed reference point that AI must be steered toward. But the cognitive science of analogy tells a different story. Gentner and Markman's work on structural alignment shows that humans don't retrieve their values from some stable internal table—they *construct* them through comparison, and the same person will generate structurally inconsistent mappings depending on which analogy is currently active (Bush-as-Roosevelt yields one moral framework; Bush-as-Churchill yields another, and subjects commit to whichever frame they entered first). Kai-Fu Lee's Ganesh Insurance example lands harder when you read it through this lens: the AI wasn't misaligned with human values—it was aligned with the *specific* human value of minimizing health costs, which happened to conflict with other human values the designers hadn't foregrounded. The real problem isn't that we can't get machines to match our values. It's that human values are themselves cross-mapped: structurally inconsistent systems that shift depending on which analogy we're currently inside, and we lack the self-awareness to notice the switch.