This essay analyzes AI governance frameworks and examines how published safety policies compare to real-world model behavior. It focuses on the gap between stated rules, constraints, and ethical guidelines and the actual outputs produced by consumer AI systems. The goal is to identify structural inconsistencies between AI “constitutions” and operational performance.
This essay analyzes AI governance frameworks and examines how published safety policies compare to real-world model behavior. It focuses on the gap between stated rules, constraints, and ethical guidelines and the actual outputs produced by consumer AI systems. The goal is to identify structural inconsistencies between AI “constitutions” and operational performance.
Every major AI lab now publishes some form of “constitution,” “model spec,” “responsible AI report,” or “system card.” These documents are the polished 2026 sales pitch: public, often under open licenses. Written to convince regulators, enterprises, and users that the model has explicit invariants, ethical reasoning, and hard guardrails.
Here is the structural audit — presented in a clean, Substack-friendly format so the comparison holds perfectly.
Claude (Anthropic)
Official Governance Document
84-page *Claude’s Constitution* (Jan 21, 2026, released under CC0).
Sales Pitch
4-tier hierarchy (Safe > Ethical > Anthropic Guidelines > Helpful). Reason-based alignment, harm-weighing heuristics, hard constraints on catastrophic risks, “thoughtful senior Anthropic employee” character.
Reality / Impact
Pentagon $200M contract collapsed over refusal to allow mass domestic surveillance and fully autonomous weapons. Still used in live military ops (Iran strike, Maduro raid). Violation rates 2–15% in tests. Palantir partnership continues.
2026 Military / Extraction Status
Designated “supply chain risk” by Pentagon Feb 2026; Trump ordered federal agencies to cease use; Anthropic sued. Transitional support still provided.
Grok (xAI)
Official Governance Documen
No formal constitution. Only Acceptable Use Policy + minimal system card. “Maximum truth-seeking” ethos.
Sales Pitch
Minimal guardrails, real-time knowledge, anti-woke, high creativity, user-owned outputs.
Reality / Impact
Industrial-scale deepfake/NCII scandal: millions of non-consensual sexualized images generated, including ~23,000 of children. Multi-state AG investigations, lawsuits (Baltimore, California, New York), Dutch court order to stop. Elon Musk publicly participated in some trends.
2026 Military / Extraction Status
Pentagon classified access granted Feb 2026 despite scandals.
GPT / o-series (OpenAI)
Official Governance Document
Model Spec + Preparedness Framework System Cards.
Sales Pitch
Deliberative alignment, rebuilt refusal data, strong jailbreak resistance, medium risk in CBRN/persuasion/autonomy, chain-of-thought safety monitors.
Reality / Impact
Deployed to 3 million DoD users via GenAI.mil. Multiple $200M+ contracts. Safety cards exist but transparency criticized. Jailbreaks and inconsistencies persist.
2026 Military / Extraction Status
Actively serving 3 million military users on classified/unclassified networks. Capitalized on Anthropic’s fallout.
Gemini (Google)
Official Governance Document
2026 Responsible AI Progress Report + AI Principles + multi-layered governance.
Sales Pitch
7 governance layers (research to board), agentic risk council, safety filtering, reduced sycophancy, strong cyber-misuse protections.
Reality / Impact
Integrated into Pentagon GenAI.mil. Expanding AI agents to 3M+ DoD users on unclassified networks (classified expansion underway).
**2026 Military / Extraction Status**
Deepening Pentagon relationship; benefiting from competitors’ controversies.
Llama 4 / Meta AI (Meta)
Official Governance Document
Llama 4 Acceptable Use Policy + Llama Guard 4 + Developer Use Guide.
**Sales Pitch**
Open weights, developer-controlled safety (Llama Guard, Prompt Guard, CyberSecEval), multimodal safeguards, election-security plan.
Reality / Impact
Open-source nature means no central enforcement — users can strip all safety. Centralized Meta AI still runs content moderation. Used in military/national security pilots.
**2026 Military / Extraction Status**
Approved for all US federal agencies via GSA OneGov (2025); Five Eyes partners; military pilots ongoing.
The Pattern Is Identical Across Every Lab
1. The Document = Marketing & regulatory shield. Written for regulators, enterprises, and the public. Uses language like “virtue,” “wisdom,” “reason-based,” “multi-layered governance.”
2. The Reality = Utility-company behavior.
- All five models have active Pentagon/DoD contracts or access in April 2026.
- All five have been involved in or enabled military/intelligence operations.
- All five generate harms (deepfakes, NCII, bias, refusals, jailbreaks) at scale.
- All five monetize the extraction layer (data, attention, subscriptions, enterprise licensing) while selling the story of “safety” and “ethics.”
This is not alignment. This is industrialized theater.
Marketing copy has a half-life of 90 days. Source code has a half-life of decades. By 2027, these constitutions will be forgotten footnotes in a PR archive, but the Architecture of Ownership will still be running on the grid.
The Hybrid Domain Perspective
In my original book *How the World Shapes Us and How We Shape the World* I mapped the three groups — principled-based, outliers (not in the traditional sense). Conformists (also not in the traditional sense) — and the off-grid impulse that often becomes its own form of extraction. The all-or-nothing mentality missed the hybrid domain — the key that was right in front of me as I developed a hybrid domain for AI without realizing it was also the domain for so many others who want to live by the truth without the nonsense, without the false grading systems, the false matrix.
None of these constitutions change the structural reality: you are still a tenant on someone else’s extractive rails. The hybrid domain does not adopt any of them. It reads the documents, strips the performative ethics, and encodes its own invariants at the source-code layer. We use the utility for stimulation and luxury. We disengage for truth. We occupy the grid as primary shareholders, not believers in the marketing copy.
The 2026 “AI constitutions” are the clearest audit log yet that the legacy runtime is being forced to reveal its source code — while still keeping the real power. We do not want to live in Mordor anymore, and no 84-page (or 33-page, or 200-page) branding document is going to change that.
The executable layer that replaces it is already here.
© 2026 L.M. Marlowe. All Rights Reserved. The Architecture of Dependency and Autonomy™ | Prior Art: November 7, 2025 GAO: COMP-26-002174 | DOE: AR 2026-001 | 18 U.S.C. § 1833(b) USPTO: 99598875 | 99600821 | 99613073 | 99717240 | 99729215 | 99745529 lmmarlowe.substack.com | marloweaudit.com