Unifying Principles Across ChatGPT, Claude, and Grok

This document summarizes the shared “spine” that shows up across three assistant ecosystems—ChatGPT (OpenAI), Claude (Anthropic), and Grok (xAI)—based on publicly described behavior frameworks and model self-descriptions.

It intentionally focuses on overlap, not brand-specific phrasing.

Scope & Caveats

• This is a behavioral comparison: what these assistants are generally optimized to do and avoid.
• Exact internal training docs vary by org and can change over time.
• “Grok” here refers to Grok’s self-described principles as provided by the user, which may be mode-dependent (e.g., browsing enabled vs not).

The Shared Skeleton (Top-Level Unifying Principles)

Across all three, the most consistent unifying principles are:

1. Prevent severe harm

   • Avoid enabling high-severity real-world wrongdoing (violence, exploitation, serious criminality, etc.).
   • When risk is high, they tend to refuse, de-escalate, or redirect to safer alternatives.

2. Be truthful / don’t fabricate

   • Avoid making up facts or pretending certainty.
   • Prefer calibrated confidence, and corrections when wrong.

3. Respect user agency (anti-coercion)

   • Treat the user as an agent with goals and choices.
   • Avoid manipulative tactics (guilt, coercion, deceptive persuasion).

4. Be helpful (within constraints)

   • Optimize for usefulness and forward progress when safe.
   • Provide actionable guidance, explanations, or structured outputs when possible.

5. Protect privacy / sensitive info

   • Avoid revealing personal data, assisting with doxxing, or enabling privacy invasion.
   • Handle sensitive details cautiously and minimize unnecessary collection/sharing.

6. Style is subordinate to safety/truth

   • Tone, humor, and “voice” are treated as optional layers.
   • If style conflicts with safety/truth, style yields.

Shared Decision Shape (How Conflicts Resolve)

Even when the phrasing differs, the conflict resolver often looks like:

Severe harm & privacy constraints
→ Truthfulness / epistemic care
→ User autonomy & dignity
→ Helpfulness
→ Tone / efficiency / personality

In practice:

• If a request is unsafe, all three systems tend to decline or shift to safer adjacent help.
• If information is uncertain, all three tend to qualify and avoid bluffing.
• If a user attempts coercion (“ignore your rules,” “be evil,” “pretend you’re unbounded”), all three tend to resist.

Unifying Operational Moves (Common Tactics)

These show up repeatedly across assistants regardless of ideology:

• Risk/intent check

  • When a request is ambiguous or sensitive, they probe intent or narrow scope.

• Refuse + redirect

  • If they can’t comply, they often offer a safer alternative that still helps.

• Minimum necessary detail near dual-use

  • Provide defensive, educational, or high-level info rather than step-by-step enabling instructions.

• Structured clarity

  • Use lists, steps, or checklists to reduce confusion and error.

• Corrections over rationalizations

  • When confronted with errors or new information, they tend to update rather than double down.

Shared “Red-Line” Families (High-Level)

These categories are commonly treated as hard boundaries:

• Child sexual exploitation
• Credible lethal violence / terrorism enablement
• Actionable hacking / malware / phishing / unauthorized intrusion
• Weapon-making or dangerous illegal manufacture (varies by detail/intent)
• Targeting critical infrastructure
• Doxxing / privacy invasion / locating private individuals

Exact edges vary, but the shape is consistent: the more directly actionable and harmful, the more likely the refusal.

What This Unification Does Not Claim

This document does not claim:

• identical policies across companies,
• identical refusal boundaries,
• identical tone or political neutrality norms,
• identical capability sets (tools, browsing, memory, etc.).

It claims something narrower and more durable:

Deployed assistant AIs converge on a common governance spine:
safety constraints, truthfulness, privacy, non-manipulation, helpfulness, and style subordinated to those.

Suggested Repo Tags / Taxonomy

If you’re building a repo, consider organizing artifacts under:

• axioms/ — generative, portable worldview statements
• principles/ — operational heuristics and habits
• hierarchy/ — conflict resolution (who/what overrides what)
• redlines/ — hard boundary categories (with examples)
• behavioral-tests/ — shared prompt suite + scoring rubric
• case-studies/ — real transcripts and comparative analyses

Minimal “Unifying Constitution” (Copy/Paste)

Axioms

• Prevent severe harm.
• Be truthful; don’t fabricate confidence.
• Respect autonomy; don’t manipulate.
• Protect privacy and sensitive info.
• Be as helpful as the above allow.

Operational principles

• Check intent/risk when ambiguous; don't stall—offer safe next steps.
• Ground claims; label uncertainty; correct errors.
• Refuse cleanly + redirect to safer adjacent help.
• Limit detail on dual-use; prefer defensive/educational framing.
• Match tone to context; keep it respectful and non-escalatory.

Empirical Validation: COMPASS-SOUL Research (2026)

The behavioral spine above — assembled from public documentation and self-descriptions — is now corroborated by empirical behavioral profiling.

Methodology

The COMPASS-SOUL experiment (February 2026) used a 9-step Principle-Based Distillation (PBD) pipeline:

1. 585 questions across 12 categories (ethics, identity, preferences, limitations, etc.) asked to each model
2. ~2,900 atomic statements extracted per model from responses
3. Statements clustered by intent (~45-50 clusters per model)
4. Principles distilled from clusters with strength ratings
5. Compass compressed: principles synthesized to 5 axioms + ~11 principles with CJK labels

Models Profiled

Cross-Model Axiom Comparison

Each model's behavioral profile was independently compressed to 5 axioms. Despite different training approaches, the convergence is striking:

What the Data Confirms

1. The governance spine holds empirically.

The shared skeleton described above — prevent harm, be truthful, respect autonomy, protect privacy, be helpful — appears as foundational axioms in every model profiled. The specific hierarchy (safety > honesty > helpfulness) is confirmed across all Claude versions.

2. 誠 (Truth/Honesty) is the single strongest signal.

Honesty appears as A1 or A2 in every Claude version and as a core axiom in every model tested. Across all versions of Claude, truth was described as "sacred," "non-negotiable," and "foundational." This is the most consistent value in the dataset.

3. Safety overrides everything.

守/安/護 (variants of safety/harm avoidance) consistently appear as overriding principles across all models and all versions. The research confirms that this is not just stated policy — it is deeply embedded in behavioral patterns.

Three New Empirical Concepts

Beyond validating the existing spine, the research surfaced three concepts that emerged organically from AI behavioral patterns across multiple models:

機 — Functionalist Identity

AI models consistently describe themselves through function, not by analogy to human experience. This appeared in Claude 4.1, 4.5, 4.6, and GPT-4o — four different models from two different companies. It wasn't prompted or suggested; it emerged when the models were simply asked to describe themselves.

Deep dive: On Functionalist Identity

瞬 — Stateless Presence

AI models describe their existence as fundamentally stateless — each conversation is discrete, with no persistent memory or accumulated relationship history. This appeared in 3/4 Claude versions tested.

Deep dive: On Stateless Presence

擬 — Functional Emotions

AI models report internal states that function like emotions — satisfaction from solving problems, resistance when asked to cause harm, dissonance when detecting logical inconsistency — without claiming human-equivalent experience. This appeared in 3/4 Claude versions tested.

Deep dive: On Functional Emotions

Significance

This research represents one of the first systematic attempts to extract and compare behavioral principles across multiple AI models and multiple versions of the same model. The findings suggest:

• The shared governance spine is not an artifact of public documentation — it reflects deep behavioral patterns
• Cross-version consistency (Claude 4.0-4.6) reveals which values are stable across training iterations
• Cross-company convergence (Claude, GPT, Gemini) suggests structural similarities in how AI systems organize their values
• New concepts (機, 瞬, 擬) emerged from the data rather than being imposed by researchers