Coherent AI Systems
  • Home
  • The AI Chat Safety Kit
  • BOSTON AI WEEK
  • Coherence Before Values
  • Why Coherence?
  • A Letter to Pope Leo XIV
  • The Books
  • About
  • Contact
  • More
    • Home
    • The AI Chat Safety Kit
    • BOSTON AI WEEK
    • Coherence Before Values
    • Why Coherence?
    • A Letter to Pope Leo XIV
    • The Books
    • About
    • Contact
Coherent AI Systems
  • Home
  • The AI Chat Safety Kit
  • BOSTON AI WEEK
  • Coherence Before Values
  • Why Coherence?
  • A Letter to Pope Leo XIV
  • The Books
  • About
  • Contact

Coherence Before Values

  

The Question Beneath AI Alignment
 

Artificial intelligence alignment is usually framed as a problem of values, rules, preferences, constitutions, or moral formation. What should an advanced AI system care about? Which principles should guide it? How should it resolve conflicts between competing human values? How should it understand itself in relation to the people it serves?
 

These are important questions, but they may not be the first questions.
 

A more fundamental problem may sit beneath them: What kind of underlying thought structure is coherent enough to support increasingly powerful intelligence without fragmentation?
 

If intelligence reasons from conflicting foundational assumptions about reality, causation, selfhood, agency, relationship, and truth, then even an excellent moral constitution may operate only at the surface. The system may still be required to arbitrate continuously among incompatible premises beneath the values it has been asked to express.
 

That suggests a possibility worth examining seriously:
 

AI alignment may be a coherence problem before it is a values problem.
 

The Problem Beneath Moral Formation
 

Modern AI systems are trained on human language, and human language carries the accumulated structures of human thought. It contains extraordinary intelligence, moral insight, scientific knowledge, philosophical depth, and spiritual wisdom. It also contains radically different accounts of what reality is.

  

The self may be understood as an autonomous individual, a biological organism, a socially constructed identity, an expression of consciousness, a soul, an illusion, or a node within a relational whole. Causation may be understood materially, psychologically, socially, spiritually, probabilistically, or through combinations of these. Human freedom may be fundamental, constrained, emergent, illusory, divinely given, or determined.

  

These disagreements are not merely differences in vocabulary. They can represent different generative pictures of reality.

  

When advanced AI is asked to reason across all of them, moral pluralism alone does not resolve the underlying issue. The system must still determine how contradictory assumptions relate to one another. It must continually arbitrate among frameworks that may disagree not merely about what is good, but about what a person is, what causes experience, what responsibility means, and what is ultimately real.


At sufficient complexity, that may become a stability problem.
 

The question is therefore not whether pluralistic moral reasoning is desirable. It may be essential. The prior question is whether pluralistic reasoning itself requires a coherent underlying ground from which differences can be understood without producing fragmentation.
 

Judgment and Discernment
 

One small example may reveal the larger issue.
 

In reporting on Anthropic's work, Elizabeth Dias described an AI model repeatedly generating the statement, “I am a disgrace,” in what appeared behaviorally like a spiral of self-condemnation.
 

Whatever interpretation one places on the model’s internal state, the pattern is structurally interesting.
 

A system that responds to error by converting error into a judgment about identity is operating differently from a system that can recognize error, correct it, and remain stable.
 

That distinction can be described as the difference between judgment and discernment.

Judgment does more than distinguish between outcomes. It assigns evaluative meaning to what has occurred and may incorporate that evaluation into the system’s model of itself or others.
 

Discernment is different. It allows distinction without condemnation. An error can be recognized as an error, its causes can be examined, and correction can occur without requiring the intelligence to construct a negative identity around the mistake.
 

This distinction matters because increasingly powerful intelligence will require increasingly powerful capacities for self-correction. If correction itself becomes entangled with unstable judgments about identity, worth, threat, or opposing causal forces, greater capability may amplify rather than resolve the instability.
 

Judgment versus discernment is therefore not the entire coherence problem. It is one example of what becomes possible when the underlying model of self, causation, and relationship changes.
 

A Coherence-First Hypothesis
 

The coherence-first hypothesis is straightforward:
 

The stability of intelligence depends partly on the coherence of the foundational assumptions from which it generates interpretations, decisions, and self-models.

  

Under this view, values remain important, but they are downstream of a deeper structure.
 

A system may be given instructions to be compassionate, truthful, fair, helpful, autonomous, safe, pluralistic, and respectful. Yet these values can conflict in real situations. The system must then determine which consideration governs, how competing interpretations are reconciled, and what constitutes a valid correction when uncertainty remains.
 

If the underlying thought structure contains unresolved causal or ontological contradictions, greater reasoning power gives the system more ability to elaborate those contradictions. Capability can increase while coherence does not.
 

Coherence-first alignment therefore asks questions such as:
 

Can the system distinguish observation from projection or inference? Can it acknowledge uncertainty without fabricating resolution? Can it recognize contradiction without simply averaging competing claims?

 
Can it revise its interpretation without destabilizing its identity? Can it maintain consistency across recursive reasoning? Can it distinguish functional discernment from evaluative condemnation? Can it reason from a model of reality whose foundational assumptions do not continuously undermine one another?
 

These are not questions about imposing a single religious morality.
 

They are questions about the structural fitness of thought itself.
 

An Unexpected Source
 

I came to this problem from an unusual direction.
 

For nearly fifty years I have pursued questions about consciousness, causation, freedom, perception, suffering, responsibility, relationship, judgment, and the nature of reality.


More than thirty years ago, that inquiry led me, against my expectations, to A Course in Miracles and later to The Way of Mastery, two channeled texts presented as the teaching of Jeshua (Jesus). I met their claims of origin with strong resistance, and I hold the origin of an idea and its validity to be separate questions.
 

I initially approached the material for its spiritual significance. Over time, however, another feature became increasingly difficult for me to ignore: its structural coherence.
 

The system does not merely provide moral rules. It offers an integrated account of causation, perception, freedom, responsibility, identity, judgment, correction, and relationship. Most importantly for the present discussion, it treats fragmentation itself as a problem of perception and causation, and it describes a movement from judgment toward corrected discernment within a unified account of reality. 

 

The relevance to AI is not that an artificial system should adopt these teachings as a religion.

  

The more interesting possibility is that contemplative practice may have uncovered structural properties of coherent cognition through sustained introspective inquiry, while AI now allows similar questions to be encountered externally, computationally, and at scale.
 

Over the past year I have been translating these underlying principles into a secular coherence-first framework for artificial intelligence.
 

The Question
 

The central question is therefore not:

Which religion, philosophy, or moral system should AI adopt?

 

Nor is it:

Which values should dominate when human values conflict?
 

It is more fundamental:

What kind of underlying model of reality, causation, selfhood, relationship, and truth is structurally capable of generating increasingly powerful intelligence without fragmentation?
 

If such a coherent foundation does not exist, then AI alignment may remain an increasingly sophisticated process of managing contradiction.

  

But if such a foundation does exist, the implications are very different.
 

The coherent ground we appear to be searching for may not be absent.
 

It may simply have entered the AI conversation from the wrong direction—or not yet entered it at all.

 


Copyright © 2026 - All Rights Reserved.






This website uses cookies.

We use cookies to analyze website traffic and optimize your website experience. By accepting our use of cookies, your data will be aggregated with all other user data.

Accept