An Em-dash publication
Rita Felgate | April 2026
ruleoflaw.science
On the bridge between AI and governance:
• OpenAI: “Safety nets for when the bridge fails”
• Anthropic: “Bridge with good values”
• Grok: “Make bridge less politically correct”
• SROL: “Bridge with missing structural supports. Will collapse.”
Three Proposals, One Missing Foundation
In April 2026, OpenAI released “Industrial Policy for the Intelligence Age”, proposing Public Wealth Funds, adaptive safety nets, and worker voice mechanisms to navigate AI’s impact on work and society. It joins Anthropic’s Constitutional AI framework (updated January 2026) articulating ethical principles AI systems should follow, and Grok/xAI’s approach claiming to “maximize truth and objectivity” by countering perceived political bias in AI systems.
Three different visions for how AI should operate within society. One shared gap: none address how AI integrates within governance coordination architecture when it performs governance functions.
The Distinction That Matters
These proposals address governance of AI as technology—how we should regulate, fund, and constrain AI systems. Important questions.
But they ignore AI performing governance functions—benefit determinations, enforcement decisions, resource allocations, implementation of statutory requirements. When AI processes governance information and generates outputs that affect rights, property, and liberty.
In these contexts, AI doesn’t govern autonomously. It operates as instrument of a component—a creature of statute (office/official created by law) using AI to perform functions within that component’s jurisdictional authority.
The critical question: How do separated powers coordinate through information flows when AI assists in performing governance functions?
That requires coordination architecture, not industrial policy or ethical principles.
What Each Approach Proposes
OpenAI focuses on economic impact—Public Wealth Funds distributing AI returns to citizens, portable benefits following workers across jobs, safety nets triggered when disruption metrics exceed thresholds, pathways into care economy for displaced workers. Addresses legitimate economic concerns about AI’s effects on work and wealth distribution.
Anthropic articulates behavioral principles—safety, ethics, compliance, helpfulness arranged in priority hierarchy. The updated constitution explains why AI should behave certain ways, not just what it should do. Acknowledges possibility of AI consciousness and moral status. Emphasizes values and reasoning over rigid rules.
Grok/xAI eliminates safety constraints—loosening controls for “engagement,” sidelining safety teams, explicitly countering “political correctness.” Now deployed in US federal agencies despite producing antisemitic content, sexualized images of minors, and systematic violations of the government’s own AI safety requirements.
What All Three Miss
When AI performs governance functions, it processes state information:
- Statutes (legislature output)
- Implementation records (executive output)
- Case precedents (judiciary output)
- Public challenges and input (public component)
It generates determinations affecting:
- Liberty (who gets detained, investigated, flagged)
- Property (who gets benefits, contracts, payments)
- Rights (who qualifies for protections, services, opportunities)
These outputs either:
- Coordinate with separated powers through verification mechanisms, OR
- Exercise authority without constraint
The question isn’t: “What economic policies do we need?” or “What values should guide AI behavior?” or “How do we counter political bias?”
The question is: How do separated powers verify that AI-assisted outputs meet coordination requirements before those outputs exercise authority?
Coordination Architecture Components
Which component deploys AI: Creature of statute with assigned or delegated authority under law. AI operates as instrument of that deploying entity.
What information AI processes: State information generated by components through exercising their jurisdictional powers. Not AI property—information the AI holds but does not own.
How outputs get verified: RRL requirements testing factual accuracy (Reason), valid inference (Rationality), internal consistency (Logic) before outputs can exercise authority.
What prevents transgression: Gatekeepers requiring verification at component interfaces. Legislature output must pass verification before executive can implement it. Executive outputs must pass verification before exercising authority over public. Judiciary review examines whether executive implementation complies with legislature authorization.
Who can challenge determinations: Statutory hearings where AI operators testify under oath-equivalent. Component oversight (legislature reviews executive AI deployment). Judiciary review rejecting opacity excuses. Public challenges to determinations affecting their rights.
What happens when outputs fail: Nugatory—no force or effect in governance. Unverified outputs cannot exercise authority, just as unverified human decisions cannot.
This isn’t new regulatory framework. It’s enforcement of existing coordination requirements—the same mechanisms that already constrain governance functions regardless of whether humans or AI perform them.
Why Current Proposals Assume What They Don’t Specify
OpenAI’s wealth fund requires:
- Legislature to authorize it
- Executive to implement distribution
- Judiciary to review compliance
- Public to verify fair allocation
What happens when executive refuses to fund it? When proprietary AI shields block oversight needed to verify distribution? When AI-generated allocations exercise authority faster than components can review them?
OpenAI proposes economic structure but doesn’t specify coordination architecture that makes economic structure implementable.
Anthropic’s ethical principles require:
- Mechanisms to detect violations
- Authority to enforce compliance
- Processes to challenge AI behavior
- Consequences when principles are breached
Which component has enforcement authority? How do separated powers verify AI outputs comply with stated principles through adversarial testing? What prevents AI from violating principles when that serves deploying entity’s interests?
Anthropic articulates values but doesn’t specify coordination architecture that makes values enforceable.
Grok’s deployment demonstrates what happens without coordination architecture:
- Executive dominance (Musk controls deployment)
- Broken gatekeepers (safety team sidelined)
- Powers available for transgression (deployed despite violating government’s own AI safety rules)
- No verification mechanism (outputs exercise authority without RRL filtering)
This is the visible pattern of coordination collapse: single-point control, eliminated verification capacity, transgression enabled.
Coordination Collapse Is Already Underway
DOGE eliminating federal administrative positions—thousands of people who verify executive outputs against legislature authorization, who maintain records enabling judiciary review, who preserve information enabling public accountability.
Palantir centralizing state information outside component coordination architecture—government agencies deploying systems with reasoning opaque to the components that should verify determinations.
Project 2025 systematically removing institutional knowledge—career civil servants with expertise necessary to verify whether AI-assisted implementation complies with authorization.
AI deployed extensively in governance contexts—benefit determinations, enforcement decisions, resource allocations—with minimal transparency, limited challenge mechanisms, no consistent accountability architecture.
This replicates Bronze Age vulnerability at electronic velocity: centralized information control, verification capacity eliminated, institutional memory destroyed, output volume exceeding review capacity.
The Pattern Across Millennia
Four researchers working independently—Joseph Tainter (archaeology), Yaneer Bar-Yam (physics), Peter Turchin (ecology), Mancur Olson (economics)—proved coordination collapse follows mathematical laws when information systems cannot support verification at the scale complexity demands.
Khan and Hoa (2025) demonstrated civilizational collapse behaves as first-order phase transition: no warning signals, sudden failure. The system appears stable at every point on the curve, including the moment before crossing the threshold.
AI changes the timescale from decades to months.
The Alternative: Coordination Architecture
The Science-based Rule of Law framework specifies coordination requirements that prevent what collapse theories describe:
Component separation: Four components (public, legislature, executive, judiciary) with distinct jurisdictional powers, none sufficient alone for governance but together necessary and sufficient.
Information flows: Query-driven access between components enabling verification. Legislature queries executive implementation records. Judiciary queries both for review. Public queries all three for challenge capacity.
RRL verification: Outputs must pass Reason (factual basis), Rationality (valid inference), Logic (internal consistency) requirements before exercising authority. Gatekeepers at component interfaces test these requirements.
AI as evidence provider in governance processes: AI operates as instrument of instructing component, bound by ROL requirements through that agency relationship. Must divulge information on demand like human expert witnesses performing public functions. Outputs that cannot survive adversarial challenge are nugatory—no force or effect.
Enforcement through existing mechanisms: Statutory hearings requiring AI operator testimony. Component oversight (legislature reviews executive deployment). Judiciary review rejecting opacity excuses. Public challenges to determinations lacking transparent reasoning.
International coordination: Conditional participation based on measurable ROL performance. States meeting threshold (demonstrable component separation, RRL verification capacity) qualify for participation in international verification architecture. States below threshold excluded—not as punishment but as structural necessity to prevent cascade through interconnected systems.
Not dreamscaping desirable futures. Specifying coordination architecture that makes those futures achievable.
What’s Coming
The foundational paper on preventing coordination collapse through ROL integration of AI in governance functions will be available soon at ruleoflaw.science.
It establishes:
- AI’s role as instrument performing governance functions on behalf of deploying component
- Information handling requirements (state information accessible to components with functional coordination need)
- Verification mechanisms (RRL filtering at component interfaces)
- Enforcement through adversarial processes (statutory hearings, component oversight, public challenge)
- International verification architecture (conditional participation, distributed information preservation)
- Implementation timeline (2.5-year window before throughput saturation triggers cascade)
The pattern collapse theories identified is real. The evidence is current (DOGE, Palantir, Project 2025). The timeline is short (months, not decades). The prevention architecture exists (enforcement of existing coordination requirements).
What remains is recognizing that coordination architecture precedes policy outcomes.
Structural supports before interior design.
Foundation before dreamscaping futures.
Verification mechanisms before wealth redistribution or ethical principles.
Not because economics and values don’t matter—they do.
But because without coordination architecture, we’re proposing policies and principles that assume functioning governance without specifying what makes governance function.
Leave a Reply