AI alignment is often described as an engineering problem: how do we make advanced systems follow human intent, avoid harmful behavior, and remain compatible with human values? That framing is useful, but incomplete. It assumes that the values a system should follow are already available, stable, and ready to be translated into technical objectives.
They are not. Human values are not a clean input. They are disputed, contextual, and often institutionally mediated. What counts as safety, fairness, autonomy, privacy, harm, usefulness, or public benefit depends on judgment. Those judgments are not only technical. They are political.
Alignment is not just the problem of making AI systems follow human values. It is the problem of deciding which human values receive authority when people reasonably disagree.
The Hidden Assumption Inside Alignment
Much of the alignment conversation begins after a major question has already been skipped. Once a goal has been selected, engineers can ask how to make a system pursue it reliably. But before that comes a different question: who had the legitimacy to select that goal in the first place?
A model used in education may need to balance personalization against equality, efficiency against teacher discretion, and student safety against privacy. A model used in criminal justice may need to balance prediction against due process, consistency against mercy, and administrative speed against the right to contest a decision. A model used in content moderation may need to balance expression, dignity, public safety, and political neutrality.
These are not merely optimization tradeoffs. They are questions about authority, legitimacy, rights, and institutional purpose. Treating them as technical parameters can make political decisions appear more neutral than they really are.
Values Do Not Speak for Themselves
The phrase "human values" can make disagreement sound accidental, as if society already shares a settled moral framework and the only challenge is encoding it correctly. But in democratic life, disagreement is not a bug. It is a normal condition of pluralism.
People disagree about what fairness requires. They disagree about when security justifies surveillance. They disagree about whether a system should prioritize equal treatment, equal outcomes, historical redress, procedural consistency, individual choice, collective welfare, or institutional trust.
That does not mean alignment is impossible. It means alignment cannot be separated from governance. If a system will make or shape decisions that affect people, then the values embedded in that system need more than technical performance. They need a legitimate process behind them.
From Technical Alignment to Public Legitimacy
The core policy challenge is not only whether AI systems can be aligned. It is whether they can be aligned in ways that affected communities, public institutions, and democratic publics have reason to accept.
That requires transparency, but transparency alone is not enough. A public agency could disclose a model's purpose and still fail to justify the values built into it. It requires participation, but participation alone is not enough if no one has power to change the outcome. It requires expertise, but expertise alone is not enough if technical experts become the default moral decision-makers for public systems.
The question is not whether engineers should be excluded from value decisions. They should not be. The question is whether engineering teams should be left to resolve those decisions without institutional structures that can authorize, contest, and revise them.
Why This Matters for AI Governance
If alignment is treated as a purely technical challenge, then the governance response will focus mainly on evaluation, testing, documentation, and model behavior. Those tools matter. But they do not answer the deeper question of whose values are being operationalized.
Public institutions need procedures for deciding when AI should be used, what values it should prioritize, who gets to challenge those priorities, and how systems should change when their value assumptions prove harmful or incomplete. Without that, alignment risks becoming a language for quietly embedding contested choices into systems that appear objective because they are technical.
This is especially important in public-sector AI. A private company can define product goals around user engagement, efficiency, or customer satisfaction. A public institution has a different burden. It must justify its choices under conditions of pluralism, rights, law, accountability, and public trust.
The Next Alignment Question
The next phase of AI alignment should move beyond the idea that values are simply inputs to be specified. Values are claims that need justification. They need institutions capable of hearing disagreement, setting limits, explaining tradeoffs, and correcting course.
Technical alignment asks whether a system follows the objective it was given. Political alignment asks whether that objective had legitimate authority in the first place.
AI governance needs both. Without technical alignment, systems may fail to do what people intend. Without political alignment, they may do exactly what they were designed to do while still lacking democratic legitimacy.