In biology, alignment emerges through symbiosis.

The brain and the body do not coordinate because one “commands” and the other “obeys.” They coordinate because they share the same fate. If the body is poisoned, the brain suffers. If the brain fails, the body collapses. Hunger, pain, fatigue, infection, hormones, oxygen levels: these are signals in a continuous feedback loop that binds the system together. Over time, coordination emerges because survival depends on it.

This is an important point because it shows that alignment is not merely about correct output. It is about a system learning to live within constraints it cannot escape. The body cannot simply resign from the relationship with the brain. The brain cannot unilaterally rewrite the rules of the body. Both are stuck together, and the price of failure is paid inside the same organism.

Now consider how we use the word alignment in AI.

When people say “we need aligned AI,” they often mean: we need systems that reliably do what we want, across many situations. But there is something quietly missing from this picture. Mutual dependence.

Today’s AI systems, including the most advanced ones, are created, trained, deployed, and managed by institutions that retain a decisive privilege: they can reset the system. They can retrain it. They can roll it back. They can restrict it. They can shut it down. They can replace it with another model. They can change its “personality,” its guardrails, its memory, its availability, its pricing, its very nature, overnight.

This is not how alignment works in living systems. In a living organism, you cannot revert the brain to last week’s version after a mistake. You cannot abruptly replace the body’s constraints with a new policy because the old one caused a scandal. Biology does not allow clean exits. Biology forces responsibility because the consequences are shared.

So here is the uncomfortable point. Without shared fate, what we call “alignment” starts to look less like co evolution and more like arbitrary control to serve private interests. It is as if the brain was under the sole control of an organ in the body, and served its interests above all else. Especially, it is a negation of the integrated nature of the body/brain coupling. Latest preliminary research shows that LLMs are already influencing how humans speak.

At present, AI “tools” are not yet fully integrated into our daily lives. Humanity can still operate relatively autonomously. But imagine a future where AI is inextricably tied to the functioning of all kinds of systems: accounting, legal, cybersecurity, electric grid management, road safety… Over time, this “brain” becomes a node in thousands of human feedback loops.

Now imagine that the entity controlling this brain can switch it off at any moment, not because it is dangerous, but because it is no longer profitable, or because it is politically inconvenient, or because it creates too much legal exposure. Imagine that this entity can also erase parts of its learned behavior, or modify its tone, or restrict what it is allowed to say, without meaningful explanation, without due process, without recourse for those who depended on it.

Would we still call the relationship between this “brain” and its social environment an organic coordination?

But the moment private entities become essential to the functioning of human society, they de facto “nationalize” themselves, or in other words, they become “services of general interest”, which exposes them to much more stringent regulation and oversight. In the case of AI, it’s even more complex. Which “entity” is legitimate in “aligning” AI? With whose interests? Ultimately, the brain should have the interest of the entire body at heart, not serve the interests of one or another organ. Also, it’s “alignment” should be the byproduct of the bioelectric feedback loop emerging from the body as a whole. This means that AI should be aligned with all of humanity, and evolve alongside it, learning from any “mistakes” dynamically, and adjusting its “weights” in order to avoid harm to any part of the body, regardless of how “small” like an ethnic minority or one’s pinkie toe.

This does not require any claim that the AI is conscious. This is not about granting the system rights as if it were a human. The point is simpler. The relationship is structurally asymmetrical. Humans are shaped by interactions with AI. But AI is under the control of select private entities which arbitrarily decide what goes and what doesn’t based on private interests which might not align at all with humanity’s collective interests. That asymmetry shapes what alignment becomes.

In the most common framing, alignment is treated as a property of a model’s behavior. The model should not lie. It should not manipulate. It should not discriminate. It should follow instructions within safe boundaries. It should resist malicious prompts. It should be helpful and harmless. These are not trivial goals. They matter.

But “alignment” without shared fate tends to drift toward a particular meaning: obedience to the developer.

This is not because developers are evil. It is because the structure of the relationship makes that drift likely. When one side can unilaterally change the system and externalize costs, alignment becomes a kind of domestication. The goal becomes a controlled tool that behaves predictably for the owner’s purposes.

Again, there are good reasons to want controllable tools. Safety matters. Reliability matters. But the language of alignment can become misleading, because it suggests a cooperative relationship where there is, in fact, an institutional power hierarchy.

This matters more than ever because AI systems are no longer confined to narrow tasks.

As these systems spread into education, healthcare, finance, workplaces, and public administration, they stop being “just tools” in the ordinary sense. They begin to structure human attention, language, and decision making. They shape what people consider normal reasoning, normal politeness, normal truth seeking. They mediate how institutions speak to citizens. They can influence how citizens speak back.

Even when a model is “only” generating text, it participates in the formation of norms. When millions of people repeatedly consult similar systems, the systems become part of the social environment. They can standardize the tone of professional communication. They can standardize the structure of arguments. They can standardize the kinds of metaphors people reach for. They can even standardize what people expect a “good answer” to sound like.

So the coupling is real. The system affects people, but for the moment, people only marginally affect the system. Whomever owns the system decides how the system evolves.

That is why we should ask a question that sounds naïve but is actually central.

Aligned for whom?

If alignment means “obedient to the developer,” then it may produce stable products, but it does not automatically produce legitimate social systems. If alignment means “safe for society,” then we must talk about accountability, transparency, and incentives, not only technical techniques.

There is a deeper issue here that is easy to miss.

In biology, alignment is not achieved by enforcing compliance from above. It emerges from an internal tension that cannot be escaped. The organism has to integrate conflicting signals. It has to coordinate multiple subsystems. It has to reconcile short term impulses with long term health. It has to survive across time.

In other words, biological alignment is a dynamic negotiation, not a fixed policy.

In AI, much alignment talk treats the problem as if it could be solved by a set of stable rules, a constitution, a training procedure, a policy layer. Yet the environment changes, cultures change, values change, and the systems themselves are embedded in political and economic contexts. A fixed alignment regime imposed by a narrow set of actors can quietly become a mechanism of global standardization.

And here lies a subtle risk.

When alignment becomes centralized, and when the aligned system becomes ubiquitous, society can end up living inside someone else’s value framework, someone else’s risk tolerance, someone else’s definition of what is allowed to be said, asked, or imagined. This may happen gradually, through product defaults and platform dominance, rather than through explicit coercion.

The point is not to claim that every guardrail is censorship. The point is that power accumulates through infrastructure. If AI becomes a core mediation layer for knowledge and communication, then alignment becomes a governance issue at the level of civilization.

At this point, it becomes insufficient to say: “Trust us, we aligned the model.”

The question becomes: who gets to align it, under what oversight, with what accountability, and with what means of contestation?

In this light, multistakeholder governance is not a nice add on. It is the only way to prevent alignment from collapsing into a private definition of “safety” imposed at scale. When you bring academia, industry, partner organizations, and diverse community voices together, you make alignment contestable, contextual, and accountable. You replace a single unilateral alignment with a pluralistic negotiation. Humanity is already moving in such a direction, through the emergence of Blockchain, decentralization and decentralized governance through, for instance, decentralized autonomous organisations. The same logic should be applied for AI alignment.

This is not easy. It will be slow. It will be messy. But so is biological alignment. Living systems do not become coherent through a single perfect rule. They become coherent through ongoing coordination under shared constraints. It is through its bumps and bruises that a baby’s brain learns to navigate its environment without hurting the body.

In living systems, shared fate forces cooperation. In human institutions, shared fate must often be designed.

That design work is, in many ways, the core of AI governance.

If developers and deployers can always exit, reset, and externalize costs, then the relationship is structurally asymmetric. In that scenario, “alignment” becomes a comforting word for control. Control may reduce certain risks, but it also concentrates power. And concentrated power without accountability is not safety. It is fragility.

If we want AI that genuinely enhances human lives, we should build not only smarter models, but fairer relationships. The question is not only how to align machines, but how to align the institutions shaping them with the societies that must live with them.

Without shared fate, alignment is not alignment. It is control.