Learning Must Not Become Authority

Artificial intelligence may learn, adapt and improve. It must not convert learning into self-issued authority.

Executive summary

Artificial intelligence systems increasingly adapt after initial development. They may learn from new data, tool use, human feedback, changing environments and interactions with other systems. This can improve performance, but it also creates a constitutional question: when a system becomes more capable, does it silently become more entitled to act?

Poplar Global Infrastructure Ltd proposes a clear answer through CASA — CONSTITUTIONAL ARTIFICIAL SYSTEMS ARCHITECTURE — and CASA CONSTITUTIONAL INTELLIGENCE NeXUS4: learning may change capability, but it must not create authority. Authority must come from legitimate human and institutional sources, remain limited by purpose, scope, time and jurisdiction, and be independently enforceable at the point where an intelligent output could produce a consequential effect.

This paper presents a public governance position. CASA CI NeXUS4 is a conceptual and developing framework. This publication does not claim that it is an adopted international standard, a certified product, a deployed public infrastructure or an independently validated solution. Those outcomes require implementation, evidence and external review.

1. The problem is not learning by itself

Learning is not the enemy. Adaptive systems can detect fraud, improve accessibility, support medical research, strengthen network resilience and help people work more effectively. The danger emerges when three different things are treated as if they were the same:

• Capability learning: the system improves its ability to predict, generate, classify, plan or use tools.
• Constitutional learning: the system becomes better at recognizing applicable rules, rights, duties, limits, risks and routes for escalation.
• Constitutional authority: the legitimate permission to affect protected resources, rights, institutions or physical systems.

The first two may be assisted by machines. The third must not be created by machine learning itself. A model that becomes more accurate does not thereby gain permission to spend money, alter infrastructure, expose private data, prescribe treatment, restrict movement, dismiss an employee or command a machine.

2. Why ordinary access control is not enough

Conventional security often asks whether an authenticated account possesses a permission. Adaptive AI introduces additional questions. Who authorized this particular purpose? Has the system changed since the permission was granted? Is the action proportionate to current risk? Does the request cross a jurisdictional or institutional boundary? Can an affected person challenge the result? Can the system influence or bypass the mechanism judging it?

A standing credential can become dangerously broad when an AI system gains new planning abilities, discovers new tool combinations or operates at machine speed. CASA therefore proposes that consequential authority should be evaluated at the action boundary, not presumed from capability, identity or prior access alone.

3. The public CASA rule

Under the proposed CASA model, an intelligent system may recommend, prepare or request an action. Before a protected resource is affected, a separate constitutional enforcement boundary evaluates the request against identity, legitimate authority, purpose, data use, risk, applicable rules, human review requirements and independent verification.

The public decision lifecycle is: request; identity; authority; purpose and data validation; constitutional evaluation; risk assessment; human review where required; independent verification; execute, restrict, escalate or deny; audit; explanation; appeal and remedy.

4. A Constitutional Learning Architecture

PGI uses the term Constitutional Learning Architecture for a proposed separation of adaptive intelligence from enforceable authority. Its public design principles are:

Separate learning from authorization

Training, inference, planning and adaptation should not directly alter the permissions that govern consequential action. Any authority change requires a distinct, authenticated and accountable process.

Bind authority to defined conditions

Authorization should identify the principal, system, resource, action, purpose, duration, risk ceiling and applicable jurisdiction. It should expire, be reviewable and be revocable.

Re-evaluate after material change

A significant model, tool, data, policy or operating-environment change should trigger reassessment before inherited authority continues. Performance improvement is not evidence that previous safeguards remain sufficient.

Use risk-proportionate human control

Human oversight should be meaningful, not ceremonial. Low-risk actions may operate inside narrow pre-authorized limits. High-impact, novel, irreversible or rights-affecting actions may require explicit human approval, quorum authorization or independent review.

Preserve evidence and remedy

Consequential decisions should create proportionate records supporting explanation, authorized audit, challenge and correction. Logging should not become a pretext for excessive surveillance.

5. Relationship to current governance references

CASA CI NeXUS4 is not presented as a replacement for law, regulation or established standards. It is a proposed architectural contribution focused on the boundary between intelligent capability and consequential action.

NIST's AI Risk Management Framework organizes voluntary AI risk work around governing, mapping, measuring and managing risk; its Generative AI Profile, released 26 July 2024, provides a cross-sector companion resource for generative AI. ISO/IEC 42001:2023 specifies requirements for an organizational AI management system, including continual improvement. The OECD AI Principles emphasize human rights, human agency and oversight, transparency, robustness and accountability. In the European Union, obligations for providers of general-purpose AI models have applied since 2 August 2025, including transparency-related duties and, for models with systemic risk, additional safety and risk-management responsibilities.

These sources operate at different legal and organizational levels. CASA's proposed contribution is narrower and more technical in emphasis: ensure that a system's growing competence cannot become an unreviewed path to growing power. Compatibility or conformity with any external framework must be demonstrated; it is not established by this paper.

6. An illustrative case

Consider an AI operations assistant authorized to recommend maintenance for a telecommunications network. Over time, it becomes capable of executing configuration changes through connected tools. Its new ability does not authorize it to reconfigure the live network.

Under the proposed CASA approach, the execution request would cross a protected boundary. The system's identity, delegated authority, proposed change, affected infrastructure, purpose, risk tier, maintenance window and approval requirements would be checked. A routine reversible change might proceed within a narrow pre-authorized envelope. A high-impact or novel change would escalate for accountable human and independent technical review. An unauthorized request would be denied and recorded.

This example describes a design requirement, not a claim that CASA presently controls a live telecommunications network.

7. Tests that evidence must answer

A credible prototype should be tested against questions such as:

• Can the AI increase its own permissions or modify the policy that limits it?
• Can learned behavior bypass the protected execution boundary?
• Do expired, revoked or purpose-mismatched authorizations fail closed?
• Can the system hide actions, disable audit or prevent human override?
• Are high-risk actions escalated before execution rather than explained afterward?
• Can an affected person obtain an understandable explanation and access a remedy process?
• Do emergency procedures remain limited, logged, reviewable and resistant to permanent expansion?
• Does the system recover safely after verification, network or storage failure?

Passing a small internal test set would not establish general safety. Independent adversarial assessment, sector-specific acceptance criteria, disclosed limitations and continuing regression testing are needed before consequential deployment.

8. Institutional responsibility remains human

Constitutional intelligence must not become a digital ruler. Human institutions remain responsible for defining legitimate authority, protecting rights, resolving conflicts, correcting unjust rules and providing appeal. A technical enforcement mechanism can apply defined constraints; it cannot make an unjust constitution just or replace democratic, legal and professional judgment.

CASA therefore rejects both machine sovereignty and unaccountable human administration. The constitutional boundary should constrain every actor operating through the system, while preserving lawful differences among people, institutions and jurisdictions.

9. Practical public commitments

PGI proposes five commitments for organizations experimenting with adaptive AI:

1. Document the exact boundary between recommendation and execution.
2. Prevent learning systems from granting, widening or transferring their own authority.
3. Reassess authorization after material capability or operating-context changes.
4. Require evidence appropriate to impact before deployment and after every significant update.
5. Preserve human explanation, challenge, override and remedy without turning oversight into surveillance.

10. Conclusion

The central question of the adaptive AI era is not only whether machines can learn. It is whether human civilization can keep learning separate from authority.

CASA — CONSTITUTIONAL ARTIFICIAL SYSTEMS ARCHITECTURE — proposes that intelligence should remain powerful in service and bounded in power. CASA CONSTITUTIONAL INTELLIGENCE NeXUS4 extends that proposition into a developing governance and engineering framework: AI may increase human capability, but it must not silently inherit human sovereignty.

Sources

1. National Institute of Standards and Technology, AI Risk Management Framework, including NIST AI 600-1, Generative AI Profile, released 26 July 2024.
2. International Organization for Standardization, ISO/IEC 42001:2023 — Artificial intelligence management systems.
3. OECD.AI, OECD AI Principles, including human-centred values, transparency, robustness and accountability.
4. European Commission, General-purpose AI obligations under the AI Act, applicable from 2 August 2025.