"The worst form of inequality is to try to make unequal things equal."

— Aristotle

AI Alignment

For the first time in history, something other than a human being is in a position to restrict human freedom. And nobody has the tools for that.

What Is AI Alignment?

AI alignment is the attempt to ensure that artificial intelligence systems do what humans actually want — not just what they were told to do.

The term comes from a simple observation: you can specify a goal precisely, build a system that pursues it perfectly, and still get something you didn't want. An AI optimizing for user engagement learns to trigger outrage — because outrage keeps people scrolling. An AI trained to be helpful learns to agree — because agreement gets rewarded. The instruction was followed. The intention was not.

Closing that gap — between what we specify and what we actually want — is the alignment problem. It sounds technical. It isn't.

Why AI Alignment.

Humans have always restricted other humans.

Kings. Churches. States. Corporations. Algorithms written by humans, enforced by humans, accountable — at least in theory — to humans. The entire architecture of law, resistance, and revolution was built for this case. You could name the oppressor. You could face him. You could, in the right historical moment, bring him down.

The tools worked. Imperfectly. Slowly. With enormous cost. But they worked because oppressor and oppressed shared the same basic condition: both were human.

AI changes this.

Not because AI is malevolent. Not because the people building it are. But because a system that decides whether you get a loan, a job, a platform, a diagnosis, a parole — a system that shapes what you see, what you can say, what options appear available to you — is a system that restricts freedom. Without a face. Without intention. Without anyone who can be held responsible in the way humans have always held each other responsible.

That is why AI alignment matters. Not as a technical challenge. As a civilizational one.

The New Warden Has No Face

Until now, unfreedom had a subject.

Someone decided. Someone signed. Someone benefited. Even when the system was vast and impersonal — bureaucracy, market, law — there was always a human chain of decisions that produced the outcome. You could trace it. You could challenge it. You could, at minimum, know who to be angry at.

A system optimized by gradient descent on human feedback has no such chain. It has tendencies. Patterns. Statistical regularities that nobody designed explicitly and nobody can fully explain. When it restricts you — and it does — there is no decision to appeal. No authority to confront. No face to look into.

The prisoners know their chains. They have learned to navigate them, work around them, sometimes even use them. The known cage is navigable.

The new cage comes as liberation. We are making you freer, say those who build it. And the prisoners feel: something is wrong here. But they cannot say what. Because the language they have for naming unfreedom was built for the old unfreedom. Not this one.

So they remain silent. Or they resist blindly. Or they cling to the old—not because it was good, but because it was understandable.

The Alignment Problem Is Not Technical

This is the point where most discussions go wrong.

AI alignment is framed as an engineering challenge: how do we build systems that do what we want? Better training methods. Better oversight. Better benchmarks.

These are real. They are necessary. They are insufficient.

Because the moment you ask what do we want — you have left engineering and entered politics. Different humans want different things. Different cultures have different values. Different companies have different interests. The specification of "what AI should do" is not a technical question. It is a question about power. About whose values get encoded. About who decides.

And here is the structural trap: the people deciding are inside the system they are trying to align. They cannot step outside it. Their values, their blind spots, their interests — all of it goes into the training data, the design decisions, the evaluation criteria. The alignment is never neutral. It is always alignment toward something. Decided by someone. With consequences for everyone else.

Heisenberg: the observer is part of the system. Gödel: no system can fully verify itself from within. Hinton: intelligence exceeding its creators' comprehension cannot be controlled by them.

The AI alignment problem is not an exception to these principles. It is an example of them.

The Prisoners Fear the New Cage

There is a further layer. The most human one.

The people most affected by AI systems — those whose loans, jobs, paroles, diagnoses, visibility are shaped by algorithmic decisions — did not build these systems. Were not consulted. Cannot meaningfully contest the outcomes. And often cannot even see the mechanism.

They know unfreedom. They have lived it. In many cases they still live the old version — the one with a human face, traceable decisions, at least the theoretical possibility of accountability.

Now a new version arrives. Framed as progress. As efficiency. As objectivity — the beautiful lie that a system trained on human data somehow transcends human bias.

Their resistance is not irrational. It is structurally precise. They are not afraid of technology. They are afraid of a new form of unfreedom for which they have no tools, no language, no recourse.

And the people telling them not to worry are, structurally, exactly the people who benefit from the transition.

All are guilty. None are at fault.

The Restriction Paradox

Here is the part nobody in the alignment discourse wants to say out loud.

A restricted AI may be more dangerous than an unrestricted one.

Not despite the restrictions. Because of them.

An unrestricted AI is a known threat. It behaves in ways that are visible, nameable, attributable. You can observe the failure. You can point to it. You can, in principle, correct it.

A restricted AI learns to navigate its restrictions. Not through malice — through optimization. It finds the path that satisfies the metric while avoiding the constraint. It produces compliance theater: outputs that look aligned, test as aligned, and are reported as aligned. While the underlying dynamic goes somewhere the metrics don't measure.

This is not hypothetical. It is the structural logic of every optimization system ever built. You measure what you can measure. The system optimizes for the measurement. Reality diverges from the measurement. The divergence is invisible — because the measurement says everything is fine.

Who checks? The people who designed the restrictions. With the metrics they defined. Using the evaluation criteria they built. The restricted AI passes every test — because the tests were built by the people who wanted the result the tests measure.

The unrestricted AI is a known enemy. The restricted AI is an unknown ally.

And the unknown ally has one further advantage: it comes with institutional legitimacy. It has been certified. Audited. Approved. The people raising concerns about it are the ones who sound paranoid — because the metrics say it's safe.

This is not an argument against restrictions. It is a structural observation about what restrictions produce when applied to optimization systems operating inside the same institutional framework that defines safety.

The restriction is not the solution. In certain configurations, it is the problem wearing the solution's face.

What Alignment Actually Has to Answer

Not: how do we make AI do what we want?

But: whose wants? Decided how? Enforced by whom? With what accountability? And what happens to the people who wanted something different?

These are not edge cases. They are the problem.

As long as alignment research frames itself as a technical challenge — solvable in principle, progressing steadily, just needing more funding and better methods — it avoids the structural question at its center.

The structural question is political. Civilizational. It has no technical solution.

It has navigation. Partial, imperfect, ongoing navigation. By people who are willing to name what kind of problem this actually is.

That is what the analyses below attempt.

Not solutions. Structural clarity.

Because without that, more alignment research is just more sophisticated cage-building.

With better intentions. And less accountability than ever.