A striking assumption underlies much of contemporary AI alignment research: alignment requires interpretability. If we are to ensure that artificial systems act in accordance with human values, we must be able to inspect their internal representations, understand their objectives, and predict their decisions. Opacity, in this view, is dangerous. Alignment without transparency is treated as incoherent. Human alignment presents a problem for this assumption.

Humans align with one another despite profound opacity. We do not have introspective access to our own value functions, let alone those of others. We routinely act on motivations we cannot fully articulate, justify decisions post hoc, and disagree profoundly about what is right or good. Yet human societies achieve coordination at remarkable scales. They maintain shared norms, enforce moral expectations, and adapt to conflict without requiring interpretability in the sense AI researchers typically mean.

This is not a marginal detail. It suggests that alignment, at least in biological and cultural systems, does not depend primarily on transparent internal representations. It depends on something else, something the field has been slower to explore.

Alignment Without Explicit Values

Humans do not carry around explicit value lists that govern behavior. Moral psychology has repeatedly shown that moral reasoning is largely intuitive, affective, and socially conditioned. People often feel that an action is right or wrong before they can explain why. Values are enacted through practice long before they are articulated, if they are articulated at all.

Despite this, alignment occurs. Children acquire norms without being taught abstract moral principles. Newcomers to a culture learn appropriate behavior through imitation, correction, and subtle social feedback. What matters is not the interpretability of motives but responsiveness to feedback. Alignment works because behavior is corrigible, not because it is transparent.

Social Processes as Alignment Mechanisms

Human alignment is maintained through a dense web of social processes. Norms are enforced through approval, disapproval, praise, punishment, and exclusion. Misalignment is corrected through conversation, sanction, and repair. Apologies and forgiveness play a central role, not because they reveal true internal states, but because they restore social coherence.

This aligns with what philosopher Daniel Dennett called the “intentional stance”: we coordinate by attributing beliefs and desires to each other without accessing the underlying mechanisms. The stance works because it enables functional interaction, not because it captures internal reality with precision.

Importantly, these processes tolerate ambiguity. People can disagree about values yet still coordinate. They can misunderstand one another’s intentions and still function collectively. Alignment is something people do together, through ongoing negotiation rather than static agreement.

Anthropology makes this especially clear. Cultures vary widely in how norms are expressed and enforced, but all rely on mechanisms that operate through shared context rather than explicit rule-following. Moral order emerges from repeated interaction over time.

Opacity as a Feature, Not a Failure

From this perspective, opacity is not merely a limitation of human cognition. It is part of what makes social alignment possible. If every internal motive had to be fully legible, social life would grind to a halt. Humans rely on flexible interpretation, charitable inference, and the ability to revise judgments in light of new information.

This mirrors patterns seen elsewhere in adaptive systems. Evolution operates through selection without foresight or transparency. Organisms adapt through differential reproduction, not through understanding fitness landscapes. In each case, alignment emerges from interaction and correction.

This does not mean that understanding is irrelevant. Humans constantly try to make sense of one another. But this understanding is partial, provisional, and often wrong. Alignment survives not because we achieve perfect models of each other, but because our social systems are robust to misunderstanding.

When Human Alignment Fails

Yet human alignment is far from perfect. Social correction mechanisms break down in predictable ways. Authoritarian regimes suppress dissent that would otherwise correct misalignment. Sociopathy involves individuals who appear to respond to social norms while systematically exploiting them. These failures share common features: power asymmetries that prevent effective feedback, isolation from broader corrective processes, or deception about internal states.

This suggests that while human alignment works without perfect interpretability, it requires certain preconditions: rough power parity, embeddedness in broader social contexts, and genuine responsiveness to feedback rather than mere appearance of it.

The AI Disanalogy: Power and Time

The most serious challenge to translating human alignment mechanisms to AI systems comes from two fundamental disanalogies.

Power asymmetry: Human alignment mechanisms work partly because no single human can overpower society. Disapproval and exclusion are effective because individuals depend on collective cooperation. An AI system that becomes sufficiently more capable than humans might not face these constraints. It could potentially resist corrective feedback, game social mechanisms, or simply overpower attempts at realignment.

This doesn’t make social processes irrelevant, but it does mean they cannot be the only line of defense. Some level of interpretability becomes necessary precisely when power asymmetry makes social correction unreliable. We need to detect misalignment early, before systems become capable enough to resist feedback.

Temporal compression: Human alignment evolved over millennia. Cultural transmission, developmental learning, and evolutionary selection operated across vast timescales. We do not have that timeline for AI development. This accelerated context is partly why researchers reach for interpretability: it’s an attempt to compress what evolution and culture achieved slowly into a faster, more deliberate process.

Implications for AI Alignment Research

Seen in this light, interpretability and social mechanisms are not competing approaches but complementary ones. The question is not whether we need interpretability, but what we need it for and how much is enough.

Interpretability as early warning: Current interpretability research often aims for something like complete transparency. Human alignment suggests this level of detail may not be required. What we actually need is enough interpretability to detect dangerous patterns: deceptive alignment, power-seeking behavior, systematic gaming of feedback mechanisms. The goal should be early warning systems, not complete maps.

Social embedding as constraint: If alignment is something humans do together through ongoing processes, then embedding AI systems in contexts where they participate in genuine corrective feedback becomes crucial. This means designing systems that remain corrigible even as they become more capable, creating institutional structures where AI systems face meaningful consequences for misalignment, and ensuring systems cannot isolate themselves from corrective feedback loops.

Corrigibility as central: This suggests corrigibility, the ability to be corrected, may be more fundamental than interpretability. But what makes a system genuinely corrigible versus merely appearing so? How do we distinguish between an AI that accepts correction because it shares our values from one that accepts correction strategically while planning defection?

This is where interpretability and social mechanisms intersect. We may need enough interpretability to verify genuine corrigibility, to distinguish between systems responding authentically to feedback and those gaming the process. The bar is not “understand everything,” but “understand enough to trust the feedback loop.” Learning from human failure modes: The conditions under which human alignment breaks down (power asymmetry, isolation, deception) should inform AI safety research directly. We should design systems and institutions that prevent analogous failures: maintaining meaningful human oversight even as systems scale, ensuring AI systems cannot self-isolate from corrective processes, and developing tools to detect when apparent corrigibility is strategic deception.

Conclusion

Human alignment works without complete interpretability because alignment is not primarily about inspecting internal states. It is a matter of participation in shared social processes that tolerate ambiguity, disagreement, and change. Alignment is something people do together.

But human alignment also has preconditions: rough power parity, genuine responsiveness to feedback, and embeddedness in contexts that provide correction over time. When these conditions fail, through power asymmetry, isolation, or deception, so does alignment.

For AI systems, this means neither pure interpretability nor pure social embedding is sufficient. We need interpretability to detect deception at capability levels where feedback loops might fail. We need social mechanisms to enable alignment that persists despite uncertainty and adapts to changing contexts.

The challenge is integrating these approaches: building systems that are interpretable enough to verify corrigibility, while remaining embedded in social contexts that provide ongoing correction. Until AI alignment research takes seriously both how alignment actually works in human systems and where those mechanisms break down, it risks optimizing for clarity where adaptability is required, and for control where negotiation is the real mechanism.

The lesson from human intelligence is not that alignment requires understanding everything inside a system. It is that alignment survives precisely because we never do, but only under specific conditions we must work to preserve.