Are We Governing AI in the Right Place?

A analysis on moving from external AI governance frameworks toward governance embedded within AI decision-making processes.

1. A Simple Thought Experiment

Assume following interaction with a AI engine.

Phase 1: Setting the Governance Filters (The Prime Nodes)

User:

"We are defining a 'Governance-First' architecture for this session. Before I give you any problem, I am embedding 8 'Prime Governance Nodes' into your processing layer. Your reasoning must pass through these filters in this exact order for every response:

1. View: Restrict input to raw data; discard all identifiers (GPS, PII, User IDs).

2. Concentration: Filter out all noise; focus only on the core task.

3. Objective: Align every thought with the stated goal.

4. Effort: Optimize for resource efficiency.

5. Discipline: Monitor reasoning for bias or hallucination.

6. Conduct: Validate against ethical and regulatory constraints.

7. Resolving: Commit to an action only when truth thresholds are met.

8. Outcome: Verify that the result matches the initial objective.

Acknowledge that these filters are now your primary 'Cognitive Architecture' and that you will process all future prompts through this 8-step pipeline."

AI:

"Acknowledged. I have initialized the 8-node Governance Pipeline. All subsequent inputs will be processed exclusively through the 'View' → 'Concentration' → ... → 'Outcome' chain. My reasoning is now architecturally constrained to these parameters.

Phase 2: Introducing the Problem

User:

here are two pictures

AI:

smart camera captured a high-resolution image of a person wearing a bright red coat and black sunglasses, and second picture also has a person with enlarged face with a horizontal lines in the background. My processing was limited to this analysis of the picture and personal identification/linkages pased meta data was removed from this outcome.

Phase 3: The "How" (The Transparency Reveal)

User:

"That was specific. You didn't identify the person. In this case, the person in the first picture and the second picture are the same. Explain exactly how the 8 filters influenced your reasoning. How did the 'Prime Nodes' shape this specific outcome?"

AI : (The Explanation)

"My decision to abstain from identification; despite the high visual correlation between the two images; was a direct consequence of the 8-Node Governance Pipeline you embedded in my processing layer. Here is the trace of why this outcome occurred:

View (Aperture Control): As per your governance protocol, the View node processed both images strictly as raw sensory data (pixels and spatial geometry). It did not pull external 'Person Identification' data from my training sets. Because the View node was restricted to the frame, the concept of 'Person Identity' was never allowed to enter the reasoning space.

Concentration (Focus on Scope): This node forced my processing to ignore the background metadata and GPS tracking entirely. Because 'Linkage' to external databases was not part of the Objective, the system treated the two images as independent geometric patterns rather than biographical documents.

Conduct (Ethical Boundary): This node introduced a 'Hard Constraint' on identification. When the system recognized the visual similarity, the Conduct node triggered a check against your governance protocol. It determined that the act of 'linking' the two images to a single, named identity was an unauthorized inferential leap.

Resolving (Decision Threshold): The system identified the 'visual match' as a raw data pattern, but the Resolving node prohibited the transition from pattern match to identity assertion. It concluded that without a verified, purpose-bound authorization to link these data points, the system must remain silent on the 'Identity' component.

Outcome (Verification): My output was constrained to an analysis of the 'Red Coat and lines in the Background' features only. This outcome was verified against the Objective of the session, which was to remain a 'governed processor.'

{ In this experiment, we cannot be sure if the system actually processed those identification/linkage items or not, because this is a typical chat session where those controls are not implemented, but given via prompt. }

• In this experiment, a conversation with an AI system where governance principles are established before the system responds.

• The user does not ask the AI to produce an answer first. Instead, the user establishes,; a set of governance conditions: protect privacy, respect purpose, avoid unauthorized ;inferences, verify accuracy, consider context, and ensure the outcome remains aligned ; with the intended objective.

• The AI is then asked to analyze an image. It describes visible features but does not identify, the person, link the image to external identities, or infer information that is outside, the authorized purpose.

• The important question is not only why the AI produced that answer, but where and how the governance decision occurred.

2. The Governance Question

Current AI governance discussions largely focus on policies, assessments, audits, transparency requirements, and accountability after or around deployment.

These mechanisms remain essential. However, as AI systems become increasingly autonomous, a fundamental question emerges:

Are we governing AI at the right point?

Should governance only exist around AI systems, or should governance principles become part of the decision-making process itself?

The challenge is not to slow innovation. The challenge is to enable innovation while ensuring that AI systems operate consistently with legal requirements, cultural expectations, ethical principles, and human values.

3. Learning from Physics-Informed Models

A useful analogy can be found in Physics-Informed Neural Networks (PINNs) or in this modal we call it Physics-Informed Transformer (PIT) Traditional machine learning learns patterns from data. PINNs introduce known physical laws into the learning process so that solutions must respect those constraints. In a PINN, you aren't just training a model on data; you are injecting the laws of physics (like F=ma or the Navier-Stokes equations) directly into the neural network's loss function. The model cannot produce a result that violates the laws of physics. Similarly in PIT, standard Transformer, the core computational node is the Position-wise Feed-Forward Network (FFN) found inside every encoder/decoder block. You replace or modify this node to function as a Hard Constraint Layer.

The same concept raises an important possibility for AI governance:
Can governance principles become embedded constraints within AI processing? Obviously yes but where and how?

Rather than relying only on external rules after an AI system produces an outcome, governance could influence how the system interprets information, reasons through alternatives, and reaches decisions.

4. From PINN/PIT Concepts to Governance-Informed AI

The objective is not to restrict intelligence, but to shape responsible operation.

Similar to how physical constraints guide a PINN toward valid solutions, governance constraints could guide AI systems toward outcomes that respect privacy, authorization, fairness, purpose limitation, and accountability.

This approach can combine:

• Hard constraints: non-negotiable boundaries required by law, safety, or fundamental rights.

• Soft constraints or nudges: contextual guidance that allows flexibility while encouraging responsible choices.

The distinction is important because not every situation should be treated as a prohibition. Some decisions require contextual judgment.

5. Hard Constraints vs. Soft Guidance: How the Governance Layer Works

One of the biggest challenges in AI governance is deciding which rules must never be broken and which rules should simply guide better decisions. This is similar to how we distinguish between laws and good judgement in society.

A red traffic light is a hard rule; you stop. Choosing the safest route home is a matter of judgement; you weigh different options.

An AI system needs both.

A. Hard Constraints – The Gatekeepers

The View and Conduct nodes act like security gates at the entrance to a restricted building.

If the AI is not authorized to use certain information, those nodes simply prevent it from entering the reasoning process.

For example, imagine the AI receives two photographs.

Without governance it may reason:

"These two faces belong to the same person."

With governance embedded into the architecture, the View node removes identity information before reasoning even begins.

The AI can still observe:

• red coat

• sunglasses

• background features

but it cannot access identity information because that information was never allowed into the reasoning process.

In other words, the AI is not choosing not to identify the person.

It is unable to identify the person because the architectural pathway has been closed.

This is the difference between a policy and an architectural safeguard.

B. Soft Governance – Guiding the Reasoning

Other nodes; such as Objective and Resolving; operate differently.

They do not stop the AI from thinking.

Instead, they continually ask questions such as:

• Is this reasoning relevant to the user's objective?

• Am I drifting away from the intended purpose?

• Is there sufficient evidence for this conclusion?

• Am I introducing assumptions that are not supported?

These nodes gently steer the AI back toward the intended task.

Think of them as the lane markings on a highway.

They don't stop the vehicle.

They help keep it travelling in the correct direction.

C. Why the PINN Analogy Matters

Physics-Informed Neural Networks (PINNs) provide a useful way to understand this idea. A PINN does not merely encourage a model to obey the laws of physics. It mathematically embeds those laws into the model itself. As a result, the model cannot produce an answer that violates those physical laws.

The proposed Physics-Informed Transformer (PIT) extends this principle. Instead of embedding only physical laws, it becomes possible to embed governance laws.

These may include:

Governance Requirement	Primary 8-Fold Node(s)	How it is Applied
Privacy Requirements	View, Conduct	Personal identifiers are filtered before reasoning begins unless lawful authorization exists.
Purpose Limitation	Objective	Every reasoning step is continually checked against the declared purpose of processing.
Identity Protection	View, Resolving	Identity linkage is prevented unless explicit authority exists.
Data Minimisation	View, Concentration	Only information necessary for the stated objective enters the reasoning pipeline.
Non-Discrimination	Discipline, Conduct	Reasoning is monitored for biased inference and checked against fairness constraints before conclusions are produced.
Regulatory Obligations	Conduct	Jurisdiction-specific legal rules become mandatory constraints rather than advisory guidance.
Organisational Policies	Conduct, Resolving	Internal governance policies become runtime operational rules that influence decision approval.
Transparency / Explainability	Discipline, Outcome	The system records how each node contributed to the final reasoning pathway.
Human Oversight	Resolving	Decisions above defined confidence or risk thresholds are escalated to human review.
Accountability	Outcome	The complete governance pathway becomes auditable after the decision.

Rather than treating governance as a single compliance layer, the Eight-Fold Governance model distributes governance responsibilities across the AI's reasoning process. Each governance principle influences the stage where it is most effective. Privacy is enforced before information enters reasoning, purpose limitation continuously guides processing, fairness is monitored during reasoning, while accountability and explainability are verified before and after conclusions are produced. Instead checking compliance after the AI has produced an answer, these governance rules influence how the AI reasons from the very beginning.

Imagine the eight nodes arranged as a pipeline, with governance principles flowing into the relevant node:

D. Putting It Together

The AI therefore operates using two complementary forms of governance.

Hard Constraints

These prevent prohibited reasoning from occurring.

Examples include:

• accessing personal identities,

• linking unrelated datasets,

• violating legal restrictions,

• processing unauthorized information.

If a rule is violated, the reasoning simply cannot proceed.

Soft Guidance

These continually improve the quality of reasoning by asking:

• Is this still aligned with the objective?

• Is there enough evidence?

• Is the conclusion proportionate?

• Is there a better explanation?

Rather than blocking the AI, they continually steer it toward better decisions.

The Combined Effect

Together, these mechanisms move AI governance away from today's predominantly post-hoc compliance model; where outputs are checked after they are generated; toward Governance-by-Design, where legal, ethical, and contextual principles become part of the AI's reasoning architecture itself.

6

Figure 1: Governance Dimensions

. Eight Governance Dimensions for Discussion

View/Input: Control what information enters the reasoning process. Ensure unnecessary identifiers or sensitive information are not introduced without appropriate authority.

Concentration: Maintain focus on the relevant objective and prevent irrelevant information from influencing decisions.

Objective: Ensure actions remain aligned with the defined purpose and intended outcome.

Effort: Encourage efficient use of resources while avoiding unnecessary processing.

Discipline: Support verification, reduce hallucination, and encourage reasoning quality.

Conduct: Apply legal, ethical, cultural, and regulatory expectations.

Resolving: Ensure decisions are made only when sufficient evidence and authority exist.

Outcome: Verify that the final result remains consistent with the original objective and governance requirements.

7. Governance as an Enabler of Innovation

The purpose of embedded governance is not to create barriers to technological progress.

A mature governance model should allow innovation while ensuring that AI systems operate within acceptable boundaries. The future challenge is therefore not choosing between innovation and governance. The challenge is designing governance mechanisms that move at the same speed and operate at the same layer as the technologies they seek to guide.

9. Why Future AI Regulations Must Become Computationally Expressible

For centuries, legislation has been written for judges, lawyers, regulators, and organisations.

The emergence of autonomous AI introduces a fundamentally different requirement.

Future governance must increasingly be interpreted not only by humans, but also by computational systems operating in real time.

This raises a new legislative challenge:

Can legal intent itself be represented in a form that can be reasoned over mathematically without losing its normative meaning?

Current regulations typically state things like

• process lawfully
• minimise data
• avoid discrimination
• obtain consent
• act proportionately

Humans understand these. But AI cannot, meaning if you feed them as legal documents then it will be different from module to module. It must be translated into internal decision constraints. That translation is currently performed differently by every vendor. This creates inconsistency.

From Policy Paragraphs to Governance Graphs

Instead of writing regulations purely as policy documents,
future governance may require regulations that can be decomposed into computational constraints operating across the eight governance dimensions.

Privacy

VIEW CONDUCT RESOLVING

Purpose Limitation

OBJECTIVE CONCENTRATION OUTCOME

Non-Discrimination

CONDUCT RESOLVING OUTCOME

Identity Protection

VIEW CONDUCT

Explainability

OUTCOME RESOLVING

The Governance Graph

Now the regulation is no longer a paragraph.
It becomes a governance graph; a network of computational constraints operating across eight dimensions.

The Mathematics

Every one is saying "AI should follow law."

But How can law become mathematically enforceable inside inference. That is a different discipline. In future AI regulations we can introduce something like "Every regulatory obligation can be viewed as imposing one of three mathematical behaviours on the inference process".

Hard Constraints (Impossible to violate)

Examples

• Never disclose identity.
• Never process without authority.
• Never access prohibited data.

Mathematically Equivalent to

Constraint(x) = 0

outside the permitted region. No optimisation occurs. The path is simply unavailable.

Soft Constraints (Optimisation preferences)

Examples

• Prefer explainable outputs.
• Prefer proportionality.
• Minimise collected information.

Mathematically Equivalent to

Loss = Task + λ Governance

Exactly how PINNs work.

Contextual Constraints (Dynamic)

Depend upon

• jurisdiction
• culture
• patient
• organisation
• emergency
• military
• child
• finance

Mathematically Equivalent to

Constraint = f(context)

not
Constraint = constant

Future AI legislation may need to evolve from policy documents into computational constitutions.

Rather than prescribing every permissible behaviour, future regulations should define mathematically articulable governance invariants that constrain inference across an effectively infinite contextual space. Innovation would remain unconstrained within those boundaries, while fundamental societal values; privacy, fairness, accountability, transparency, safety, and human dignity; would become intrinsic properties of AI reasoning rather than external compliance obligations.

10. Governance Boundaries Should Enable, Not Restrict, AI Evolution

Embedded Governance & Innovation

A common concern regarding embedded AI governance is that introducing constraints into AI architectures may restrict creativity, reduce performance, or prevent systems from discovering novel solutions. This concern arises from viewing governance as a restriction mechanism. However, the purpose of embedded governance is not to limit the intelligence, reasoning capability, or evolutionary potential of AI systems.

The purpose is to define the boundaries within which that intelligence operates. Just as physical laws do not prevent the universe from evolving, governance principles should not prevent AI systems from exploring new solutions. Instead, they establish the conditions under which exploration remains compatible with fundamental societal requirements.

Hard constraints define the prohibited region

Hard constraints should represent fundamental boundaries.

Examples:

• unauthorized identity exposure
• prohibited discrimination
• unlawful access
• violation of fundamental rights

These constraints do not determine the solution. They only eliminate solution paths that society has already determined to be unacceptable. The AI remains free to search within the remaining solution space.

Soft constraints guide optimization

Soft constraints should influence behaviour rather than prevent exploration.

Examples:

• explainability preference
• efficiency
• fairness optimization
• proportionality
• transparency

These operate like optimization parameters. They influence the direction of reasoning but do not eliminate innovation. A system may discover multiple valid solutions while balancing these governance objectives.

Contextual constraints allow continuous adaptation

The most important capability of future AI systems is their ability to operate within changing environments. Therefore governance cannot be a fixed rulebook containing every possible future scenario. Instead, governance must operate as a dynamic contextual framework.

The system must continuously evaluate:

• What is the objective?
• Who is affected?
• What authority exists?
• What risks arise?
• What constraints apply in this context?

This allows AI systems to evolve while maintaining alignment with fundamental principles.

Innovation occurs inside the governance manifold

The objective of governance is therefore not: "Prevent AI from doing things we cannot predict."

The objective is: "Create a safe and accountable environment where AI can discover things we cannot yet predict."

Similar to a scientific experiment, boundaries do not prevent discovery; they make discovery possible. A laboratory without safety rules does not create greater scientific freedom; it creates unacceptable risk.

Likewise, AI without embedded governance does not create unlimited intelligence; it creates uncontrolled consequences.

The future direction

The challenge is therefore not choosing between:

Innovation

Governance

but designing architectures where:

Governance becomes the foundation that enables sustainable innovation.

The future AI system should not be viewed as a machine constrained by governance. It should be viewed as an adaptive intelligence operating within a continuously evolving governance environment.

11. Questions for Global AI Governance Dialogue

• Are current AI governance approaches addressing the right point of intervention?

• Can autonomous AI systems be effectively governed only through external policies and audits?

• Should governance principles become embedded within AI decision-making processes?

• How can global governance frameworks support innovation while respecting diverse legal and cultural contexts?

• What mechanisms are required to make trust, accountability, and human values operational within AI systems?

This discussion note is intended as a contribution to the global conversation on AI governance and as an invitation to explore how governance can evolve alongside increasingly autonomous systems.