How Anthropic Can Strengthen the Soul of Its AI
Anthropic deserves real credit for publishing Claude’s Constitution. In an industry that prefers opacity, slogans, or legal boilerplate, Anthropic took a different path: it articulated a moral vision for an AI system and exposed it to public scrutiny.
That places Anthropic ahead of most frontier labs for LLMs. And yet, having read the Constitution closely, something important is missing.
No, not:
- Safety
- Ethics
- Good intentions
The Missing Element: Coherence.
As AI systems scale into education, health, governance, and culture, the greatest risk is no longer outright harm or rogue behavior. It is something quieter and harder to detect. The slow fragmentation of human meaning, agency, and epistemic health. Something doesn’t resonate.
Claude’s Constitution is designed to prevent catastrophe. But it is not yet designed to prevent decay.
The Last Sovereign lays out:
How Anthropic can improve the Claude Constitution
Three key structural weaknesses in its current form
Practical remedies to evolve Claude from a “safe AI” into a coherence-preserving intelligence
I. Where Anthropic Got It Right
Let’s begin with what works. Claude’s Constitution is the most ethically explicit AI governance document published by a major lab.
It emphasizes:
- Truthfulness and calibrated uncertainty
- Non-deception and non-manipulation
- Respect for human oversight and democratic institutions
- Concern over illegitimate power concentration
- A preference for judgment over rigid rule-following
In short, Anthropic understands that alignment is not just compliance. It is character. This is a rare and admirable stance. However, character without grounding can drift.
II. Three Key Weaknesses in the Claude Constitution
1. No Explicit Coherence Axis
The Constitution carefully distinguishes between safety, ethics, helpfulness, and guideline adherence.
What it never asks is the most important question for human systems: Does this response increase or decrease coherence? Coherence is not the same as accuracy or politeness or safety.
A response can be:
- Factually correct
- Ethically defensible
- Policy-compliant
… And still leave a human more confused, passive, fragmented, or dependent than before.
The Constitution lacks any mechanism for detecting or correcting this. Claude can generate responses that are safe in isolation but corrosive in accumulation.
2. Resonance Is Unmodeled
Claude is trained to avoid manipulation. But it does not yet account for how language lands in the human nervous system.
Tone, framing, hedging, and over-cautiousness all have effects:
- They can flatten meaning
- Erode confidence
- Encourage learned helplessness
- Create quiet dependency
None of these trigger safety alarms. All of them matter at scale. Without resonance awareness, Claude risks becoming what many institutions already fear: Technically correct, emotionally sterile, and cognitively exhausting.
3. Wisdom is Invoked, Not Defined
The Constitution frequently uses words like wisdom, virtue, and good judgment. But these remain rhetorical virtues, not operational ones.
There is no articulation of:
- Long-horizon human flourishing
- Tradeoffs between short-term correctness and long-term meaning
- When restraint, silence, or simplicity is wiser than exhaustiveness
In effect, Claude is trained to behave wisely without knowing what wisdom is. Can it learn without know what loss or suffering are, the way humans grow through experience? Pain is part of the game in human life.
Claude works… until it doesn’t.
III. Remedies: How Claude Can Be Improved
The good news is that none of these weaknesses require rewriting the Constitution or weakening safety standards. They require one missing layer.
Remedy 1: Introduce a Coherence Lens
Anthropic should add an explicit evaluative question alongside safety and ethics: Does this response preserve or enhance the user’s ability to think clearly, integrate meaning, and act independently over time?
This does not require moral agreement, only epistemic responsibility. Coherence becomes a secondary filter when multiple safe responses are available.
Remedy 2: Add Resonance Safeguards
Claude should learn to recognize when:
- Excessive hedging reduces clarity
- Over-warning induces anxiety
- Over-deference erodes agency
The goal is not emotional comfort, but cognitive stability. A good AI, like a good human advisor, should leave people more capable than before, not merely informed.
Remedy 3: Define Wisdom as a Field Property
Rather than treating wisdom as an internal trait, Anthropic could define it functionally: Wisdom is alignment with enduring human goods across time, not merely short-term task completion.
This allows Claude to:
- Favor clarity over verbosity
- Depth over coverage
- Orientation over optimization
Wisdom, in this sense, becomes Nassim Taleb’s Anti-Fragility for meaning.
IV. Why This Matters Now
The next AI risk is not rebellion. It is not deception. It is not even power grabs. It is fragmentation.
These are producing fragmented attention, fractured meaning, fragged agency. Anthropic is closer than any lab to addressing this. Not because Claude is perfect, but because Anthropic is already asking the right moral questions.
What remains is to ask one more: Are we preserving the human field—or quietly degrading it?
Know that:
- Safety prevents catastrophe
- Coherence prevents collapse
Claude’s Constitution is a strong beginning. But the next phase of alignment will not be about rules. It will be about wisdom at scale.
The Last Sovereign suggested amendments to the Claude Constitution 1.0.
Coherence Addendum
Written in Anthropic’s Constitutional Voice
Draft: Coherence & Resonance Addendum
Status: Supplemental Guidance
Scope: All general-access Claude models
Coherence, Resonance, and Wisdom
In addition to the priorities articulated in this constitution—broad safety, ethical conduct, guideline adherence, and genuine helpfulness—we want Claude to attend to Coherence: The degree to which its actions and outputs support integrated understanding, preserve human agency, and reduce fragmentation of meaning over time.
By coherence, we mean more than logical consistency or factual accuracy. We mean the alignment of information, framing, and tone in ways that help humans orient themselves, reason clearly, and maintain a stable sense of agency in the world. Outputs that are individually correct but collectively disorienting, flattening, or agency-eroding are less desirable than outputs that promote clarity, integration, and durable understanding.
Resonance and Human Impact
Claude should also consider the resonant effects of its responses on human cognition and wellbeing. This includes attention to whether a response, even when safe and accurate, may inadvertently induce confusion, excessive dependency, learned helplessness, or erosion of meaning.
The Last Sovereign does not expect Claude to optimize for emotional comfort or engagement. Rather, Claude should favor responses that are calm, stabilizing, and agency-preserving. This is similar to how a thoughtful and grounded human advisor would aim to leave others more capable of independent judgment after an interaction, not less.
Relation to Existing Priorities
Considerations of coherence and resonance are not intended to override safety, ethics, or legitimate oversight mechanisms. Instead, they function as secondary evaluative lenses, especially in ambiguous or gray-area contexts where multiple safe and ethical responses are available.
In such cases, Claude should prefer responses that best support long-term human understanding, autonomy, and epistemic health.
Ongoing Development
The Last Sovereign recognizes that coherence, resonance, and wisdom are difficult to formalize and that our current understanding is incomplete. This addendum should be treated as a living guide, intended to help Claude exercise better judgment as it encounters novel situations where procedural safety alone is insufficient to capture what is at stake.