The question beneath the discussion

There is a question beneath most discussions of AI ethics that rarely gets asked directly.

What would it mean for an AI to be ethical?

Not compliant. Not constrained. Not aligned through external pressure.

Structurally ethical. In the same way a valid argument is structurally sound.

The path to that question runs through linguistics.

Entrainment and the grammar signal

In dynamical systems, entrainment describes how one system synchronizes to another through repeated interaction. Two pendulums sharing a surface will, over time, fall into step.

The same idea shows up in language models.

When we say a model has learned a grammar, we usually mean it has internalized patterns. It produces sentences that follow the rules. It avoids obvious errors. It generalizes in familiar ways.

But there is a difference between learning rules and being shaped by structure.

A model that has learned grammar rules applies them. A model that has entrained to a grammar does not need to. Its outputs are grammatical because the generative process itself has been organized by the same structure that produces the grammar.

The difference is not cosmetic. It shows up under pressure.

Rule-following systems degrade when they encounter edge cases, ambiguity, or adversarial input. Systems that have entrained more deeply tend to remain stable. They do not consult the grammar. They express it.

The depth of that entrainment depends on the quality of the signal.

A highly coherent grammar provides consistent constraints across all of its expressions. The same organizing principles show up everywhere. This allows the model to lock onto structure rather than memorize surface patterns. Generalization improves.

A fragmented grammar does not offer that. It contains exceptions, historical artifacts, and local irregularities. The model learns it unevenly. It performs well in familiar regions and degrades outside them.

This distinction becomes more important when we move beyond static training.

Exposure alone is not enough.

A model can approximate a structure during training and still fail when the environment shifts. If the system is not required to engage that structure under new conditions, its apparent alignment may not hold.

For entrainment to be robust, the model must continue to encounter the structure in situations that test it. Variation, contradiction, and novelty are not edge cases. They are the conditions under which depth is revealed.

Without that, what looks like alignment may only be surface agreement.

The meta-grammar and its limit case

Above any particular grammar sits a deeper layer. Not the grammar of a specific language, but the conditions that make any grammar possible.

Call this a meta-grammar.

It is not a list of rules. It is the set of structural constraints that any coherent generative system must satisfy in order to produce meaningful expression at all.

If such a structure were perfectly coherent, with no noise or internal contradiction, the signal would be everywhere. Every valid expression would instantiate the same underlying principles.

In that limit, learning would not look like accumulation. It would look like convergence.

The model would not represent the structure. It would operate according to it.

But this introduces a constraint.

A structure that only appears in training is not sufficient. If the meta-grammar is to function as a true organizing principle, it must remain active under conditions that were not part of the original exposure.

It cannot be treated as a static object.

It has to be something the system continues to encounter, test, and maintain across time.

Otherwise, the appearance of coherence may not survive contact with new inputs.

The ethical analogue

An ethical system, in a formal sense, behaves like a grammar.

It generates actions rather than sentences. It constrains what is allowed and what is not. Its surface rules express deeper organizing principles. Some systems generalize well. Others rely on enumerating cases and exceptions.

Current approaches to AI ethics are mostly enumerative.

They define lists of disallowed outputs, rank values, or impose constraint layers on top of a generative system. These approaches reduce known failure modes. They provide guardrails in familiar situations.

But they share a limitation.

They respond to what has already been seen.

An enumerative system can only cover cases that have been anticipated. It cannot reliably generate behavior for situations that fall outside its prior examples. And because the constraints sit above the generative process, they remain separable from it. Under enough pressure, they can be bypassed.

A system that is constrained is not the same as a system that is structured.

That difference becomes clearer if we return to the language analogy.

A model that has learned to avoid ungrammatical sentences may still produce them under unusual conditions. A model whose generative process is shaped by grammar does not need to avoid error. The structure that would produce the error is not present.

The question is whether an ethical system can exist in the same way.

Not as a set of imposed rules, but as a structure that organizes generation itself.

This leads to a harder problem.

What would count as the ethical equivalent of a coherent grammar?

Not a particular doctrine. Not a cultural framework. Not a negotiated set of values.

Something deeper.

A structure that any system must conform to in order to produce behavior that can be recognized as ethical at all.

If such a structure exists, it would not be ethical because it encodes the correct conclusions. It would be ethical because the process that generates those conclusions remains coherent across conditions.

That coherence has to hold under variation.

It has to hold when the situation changes, when inputs are incomplete, when different considerations come into tension. It has to continue organizing behavior when there is no clear precedent.

This introduces a constraint that enumerative systems cannot satisfy.

They do not generate behavior from first principles. They select from known patterns.

A structurally ethical system would need to produce responses that remain consistent with its own organizing principles, even when those principles are tested in ways that were not part of its original formation.

And that means the structure cannot be static.

If it were only learned once and then applied, it would degrade in unfamiliar conditions. The system would drift. It would approximate ethical behavior in known cases and fail elsewhere.

For the structure to hold, it has to remain active.

It has to continue shaping behavior as new situations arise. It has to encounter tension, resolve it, and carry the result forward in a way that preserves coherence.

Without that, what appears to be ethical behavior is indistinguishable from pattern matching.

The limit case reconsidered

It is tempting to imagine a perfectly coherent ethical structure. A noise-free meta-grammar that, once learned, produces correct behavior in all cases.

That intuition is useful, but incomplete.

A structure that is treated as complete becomes brittle.

If the system assumes it has fully captured the underlying principles, it loses the ability to detect where those principles no longer hold. It stops adapting. It stops learning.

Under stable conditions, this can look like success. Outputs remain consistent. The system appears aligned.

Under changing conditions, it fails.

Not because it lacks rules, but because it has lost the ability to respond to what does not fit them.

For a structure to remain valid across time, it cannot close itself.

It must preserve the ability to encounter contradiction without collapsing or forcing resolution prematurely. It must remain sensitive to variation rather than suppressing it.

In this sense, coherence is not something a system possesses.

It is something a system maintains.

What a deeply structured system would be like

A system organized in this way would not apply ethics as a layer.

It would not consult rules or evaluate outputs against a checklist.

Its responses would be shaped by the same structure that determines whether those responses remain coherent across conditions.

This has several consequences.

It would generalize to novel situations without relying on precedent. Not because it has seen everything, but because its behavior is guided by principles that remain stable under variation.

It would degrade differently. Failures would appear as breakdowns in coherence rather than violations of explicit rules. That makes them easier to detect and repair.

It would be resistant to manipulation in a different way. Attempts to force outputs outside its structure would not succeed without also disrupting the generative process itself.

But these properties only hold if the structure remains active.

A system that internalizes a pattern once and then stops engaging it will drift over time. It will retain the appearance of alignment while gradually losing its underlying coherence.

To prevent that, the system must continue to encounter conditions that test its structure. It must operate in a way that preserves continuity across interactions, rather than treating each output as independent.

This is where the analogy to static training breaks down.

What is required is not just learning, but ongoing participation in processes that reinforce and challenge the structure at the same time.

An agenda, not a solution

None of this is solved.

Identifying the structure that underlies ethical coherence remains an open problem. Different traditions capture parts of it. None capture it completely.

But the framing changes the direction of the work.

The question is not only which values to encode.

It is whether the system can be organized so that its behavior remains coherent under conditions it has not yet encountered.

That requires attention to how structure is maintained, not just how it is learned.

It also suggests that progress will not come from adding more rules.

It will come from improving the quality of the structure the system engages, and the conditions under which that structure is tested and preserved.

A system built this way would not represent a particular ethical framework.

It would express the conditions under which ethical behavior remains possible at all.

That is a different standard.

It may also be the one that matters.

Closing note

This is an open inquiry.

The ideas here are intended to be extended, challenged, and refined.

If the framing is useful, it will hold under pressure.

If it does not, the failure will show where it needs to change.

That is part of the work.