Much of the contemporary conversation about artificial intelligence turns on a single idea: alignment. We are told, with increasing urgency, that the central task before us is to ensure that AI systems are “aligned with human values.” The phrase has an intuitive appeal. It sounds ethical, technical, and solvable. It suggests that if we could just get the values right—and tune the system carefully enough—the problem would be contained.

But this framing is deeply misleading.

The difficulty is not that alignment is hard. It is that alignment presumes something that does not exist. Human values are not coherent, stable, or internally consistent. They vary across cultures, institutions, situations, and even moments within a single individual. They are often discovered only after failure, conflict, or regret. To speak of aligning a system with “human values” is to smooth over precisely the instability that defines human moral life.

Institutions have always known this, even when they forget it rhetorically. Law, medicine, aviation, diplomacy, and the military do not function by aligning participants with shared values. They function—when they function at all—through procedure, constraint, review, and escalation. We do not ask pilots to internalize the values of safety and then trust them to act accordingly. We design cockpits, checklists, and chains of command so that even well-intentioned humans cannot easily act on impulse, fear, or overconfidence.

The language of alignment quietly shifts responsibility away from design and toward disposition. When an AI system fails, we are encouraged to say it was “misaligned,” rather than asking harder questions: Why was it allowed to act in this domain at all? What constraints were missing? What evidence should have been required? Who had authority, and on what basis? Why was speed rewarded where caution was warranted?

These are not questions about values. They are questions about governance.

The irony is that humans themselves are profoundly misaligned. Individuals hold contradictory commitments. Organizations pursue incompatible goals. Governments reverse positions with changes in leadership or circumstance. Yet we do not respond to this by attempting to align everyone morally. We respond, imperfectly, by building structures that absorb disagreement, slow irreversible decisions, and distribute responsibility.

ACP begins from this more modest premise. It assumes that values are unstable and that systems must be designed to survive that instability. Rather than asking an AI system to want the right things, ACP asks what the system is permitted to do, under what conditions, with what evidence, and with what consequences if it fails. It treats refusal, hesitation, and partial answers not as defects but as legitimate outcomes.

This shift—from intention to permission—is subtle but decisive. It moves the burden of safety from the internal psychology of a model to the external architecture of its use. In doing so, it abandons the fantasy that alignment will save us, and replaces it with the harder, older work of building systems that assume fallibility—human and machine alike.