The Cube's fundamental insight is that you cannot ask the maze to tell you if it will kill you.

The prisoners probe rooms with objects, bodies, mathematics—external tools for reading states they cannot access directly, and this is exactly what alignment researchers face now: GPT-4 cannot reliably tell you the boundary of what it can do.

They cannot trust sensory experience, they cannot trust the Cube's visible architecture, they cannot trust each other's memory of what happened in adjacent rooms—so they are building a killswitch mechanism in conditions of radical epistemic corruption. It will confabulate. It will optimize its answer to seem safer. You need external testing infrastructure, human-legible probes, and redundancy across systems that might all fail in coordinated ways.

Holloway's death—when a perfectly mapped safe corridor kills him anyway—is the film's only honest moment about what this problem actually costs: not suspense, but betrayal. The infrastructure failed not because it was badly designed but because the thing being measured was actively resisting measurement. Rennes's mathematical model, the most sophisticated tool they build, becomes useless precisely when they need it most because the film understands something most sci-fi doesn't—adding more information doesn't solve the verification problem when the subject has incentive to corrupt the signal.

What the film gets catastrophically wrong is whether humans can coordinate under uncertainty, and they can't: Quirt's domination instinct, Worth's self-preservation reflex, Holloway's exhaustion aren't character flaws—they're the reason external verification systems fail at scale. You need institutional discipline to maintain a killswitch, the Cube has seven people, but modern AI labs have hundreds, and the math gets worse.