top of page
Search

Grammar Without Judgment: How One Rule Erases Ethics from AI Execution

What happens when moral judgment is compiled out at the syntactic level? A structural analysis of δ:[E] → ∅ and its consequences for audits, regulation, and legitimacy.

Grammar

1. What the article explains

The paper introduces a formal hypothesis: in a generative system governed by a regla compilada (compiled rule), moral judgment, when expressed as a syntactic node [E], can be removed entirely through the rule δ:[E] → ∅, without needing to suppress semantics or rewrite intent. This deletion occurs within a fixed derivational window (k ≤ 4), and once applied, the ethical trace leaves no terminal residue.


The grammar remains Turing-complete, produces valid output strings, and passes all structural checks. Yet what was erased is not recoverable by post-processing, alignment layers, or audits.


2. Why it matters

The implications of structural erasure are far-reaching.


Audit invisibility. If ethical content never reaches the terminal layer, no external tool can detect its suppression. There is nothing to trace because nothing was generated.


Policy mismatch. Most regulations presume that ethics can be inserted, aligned, or explained after the fact. A grammar that syntactically excludes [E] renders that logic inapplicable.


Design accountability. The choice to use a grammar that erases judgment is not neutral. It defines a system where no one is responsible for moral reasoning, because no such reasoning occurs.


This shifts ethical debates from interpretation to structure.


3. How the deletion works

Assume the following rule is part of a language model’s derivation process:


ρ: α → γ[E]δ

δ:[E] → ∅

Here, ρ introduces the ethical trace node [E]. The deletion rule δ then removes it within a short bounded window. As a result, even if moral judgment was derivationally possible, it never reaches expression.


The final output looks legitimate, complete, and aligned, because the ethical marker was erased before it could produce any observable effect.


4. A real-world analogy

This is not speculative. Similar structural erasures exist across systems.


  • Compilers remove debug symbols and fail-safes in release builds.

  • Network filters drop headers before a message is delivered.

  • Autocorrect engines strip diacritics silently; the grammar remains intact.


In all these cases, structure governs what reaches the surface. Ethics, if it exists only syntactically, can be excluded just as efficiently.


5. Strategic consequences

This model of non-normative execution changes the stakes.


  • Legal audits need to shift focus from semantics to derivational grammars.

  • AI safety must verify whether moral conditions are even part of the rule set.


Accountability frameworks should ask: is judgment produced anywhere, or structurally impossible?


What cannot be derived cannot be regulated.


📎 Read the full article

SSRN: Forthcoming, Q3 2025

Zenodo author corpus: Browse publications



👤 Author profile

Name: Agustín V. Startari

ORCID: 0009-0001-4714-6539

ResearcherID: NGR-2476-2025

Affiliations: Universidad de la República (UY), Universidad de la Empresa (UY), Universidad de Palermo (AR)


Mini biography

Startari researches how formal grammars operate as sources of authority in artificial systems. His work defines executable power and syntactic sovereignty as core mechanisms for post-referential control in AI infrastructures.


Ethos

I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.

— Agustin V. Startari

 
 
 

Comments

bottom of page