An Actionable and Transparent AI Governance Framework

By Anar Simpson

As of March 2026, the global adoption of Artificial Intelligence has reached a critical inflection point. While 78% of organizations have integrated AI into core functions (like credit/loan approvals, work permits, employment, and healthcare), the infrastructure required to govern these systems has not caught up. We are in a static governance model where generic guardrails are either "baked-in" during model training or "compiled" into the application code. This article explores a solution to the above dilemma: a Constitutional Governance Layer. 

This article is written by Anar Simpson.  Anar is a computer scientist and Founder at Orchestrate.Agency, where she leads the development of localized Constitutional AI frameworks for those seeking to align machine speed with human values. As a former Deputy for the UN High-Level Panel on Women's Economic Empowerment and through leadership roles with Technovation, Mozilla, and the U.S. State Department's TechWomen program, her work in digital agency has reached over 100 countries

---

As of March 2026, the global adoption of Artificial Intelligence has reached a critical inflection point. While 78% of organizations have integrated AI into core functions (like credit/loan approvals, work permits, employment, and healthcare), the infrastructure required to govern these systems has not caught up. We are in a static governance model where generic guardrails are either "baked-in" during model training or "compiled" into the application code.

The problem with static governance is that it cannot iterate. It cannot adapt when real-world evidence shows the rules aren't working. Once a model is deployed, the rules are locked in. To change a fairness threshold or add a new transparency requirement, you have to retrain the model or rebuild the app. One path forward could be to insert a Constitutional Governance Layer. This living, independent architecture sits between the community's values and the AI's execution, which the stakeholders can update independently of the underlying third-party model.

The Architecture of Agency: The Constitutional Layer

Unlike governance that is fixed at the point of deployment, the constitutional layer can be updated without retraining the underlying AI model. It can be swapped across third-party vendors, and most importantly it adapts to local contexts. The layer has five interconnected components. Together they create a closed loop between governance intent and operational reality.

The Constitution (Community-Written Rules)

Everything starts with a human-readable document that stakeholders write together to reflect community values with negotiable and measurable rules. For example: in a credit system, the stakeholder group (regulators, lenders, and advocates) might agree that "Approval rate differences between regions cannot exceed 8%." This is a negotiated, democratic document that can be updated as the context changes.

The Translator (From Human Language to Machine Logic)

A constitution that says 'be fair' is unenforceable. 'Be fair' needs to become specific rules and the Translator is the bridge. It takes the human-readable Constitution and converts it into machine-executable logic. For example, in the credit use case, the Translator converts “Approval rate differences between regions cannot exceed 8%”  into a precise instruction: track approvals and denials by county on every loan decision, calculate the difference, and flag for stakeholder review if the gap exceeds the threshold. In addition, the Translator checks for conflicts. If the Constitution says 'Be fair to all counties' but also says 'Don't collect location data,' the Translator flags this as a logical impossibility and requires stakeholders to resolve the contradiction.

The Arbiter (Real-Time Enforcement)

Once the constitution is live, every decision flows through the Arbiter. This is where enforcement happens. The Arbiter acts as the backstop. If a decision violates the Constitution, like a loan approval disparity that crosses the defined threshold;  the Arbiter flags it and escalates to human review. It enforces a human-defined "safety brake" at machine speed. In the loan approval case example,  when the Arbiter detects that County A shows a 73% approval rate against County B's 62%: an 11% gap exceeding the threshold; it enforces the constitutional rule: halt automated processing and escalate every new application to human review until stakeholders investigate. This might make the process a bit slower and therefore more expensive. Yet, to protect the people being affected by these automated rules, this tradeoff becomes a community decision.

The Scorer (Continuous Measurement)

Stakeholders review the Scorer's data and update the Constitution if needed. In the credit example, what looked like algorithmic bias turned out to be an infrastructure problem. Instead of lowering the disparity threshold, stakeholders can amend the constitution to require SMS-based application support in counties with poor connectivity and a phone call within a day for any incomplete application. Rules can also be recalibrated as evidence accumulates. If an 8% threshold is triggering too many false alarms, stakeholders can move it to 10%. The Translator recompiles and the new rules are live in minutes, not months.

All of this  requires a convened Human Oversight Process: a stakeholder group with decision-making authority that reviews the evidence, resolves conflicts, and approves changes. This creates the core feedback loop: Updated constitution ➛ Translator recompiles ➛ Arbiter enforces new rules ➛ Scorer measures impact ➛ Stakeholders iterate again. Humans govern at human speed. The Arbiter enforces at machine speed

The Living Constitution (Iteration Based on Evidence)

Stakeholders review the Scorer's data and update the Constitution if needed. In the credit example, what looked like algorithmic bias turned out to be an infrastructure problem. Instead of lowering the disparity threshold, stakeholders can amend the constitution to require SMS-based application support in counties with poor connectivity and a phone call within a day for any incomplete application. Rules can also be recalibrated as evidence accumulates. If an 8% threshold is triggering too many false alarms, stakeholders can move it to 10%. The Translator recompiles, and the new rules are live in minutes, not months.

All of this requires a convened Human Oversight Process: a stakeholder group with decision-making authority that reviews the evidence, resolves conflicts, and approves changes. This creates the core feedback loop: Updated constitution ➛ Translator recompiles ➛ Arbiter enforces new rules ➛ Scorer measures impact ➛ Stakeholders iterate again. Humans govern at human speed. The Arbiter enforces at machine speed.

A Global Governance Commons: Shared Constitutions, Local Implementation

The strength of this framework lies in two key elements working together: the technical implementation of the Constitutional Layer and the repository of Constitutions that makes it actionable.

We have published four draft constitutions to showcase this approach: Kenya's credit assessment, Canada's AI hiring under Ontario's new disclosure requirements, the UAE's "Eye" work permit system, and Japan's AI-assisted healthcare diagnostics. Each of these constitutions translates principles into operational guardrails: fairness thresholds, escalation triggers, appeal timelines with guaranteed human review, prohibited practices, and public dashboards. We hope that these constitutions will be forked and adapted to other applications as a shared repository that moves actionable governance forward faster than individual organizations starting from scratch.

A Canadian hiring constitution and a Brazilian hiring constitution will have different legal thresholds and different languages, and yet the underlying fairness principles, escalation triggers, and appeal rights could be similar. Some constitutions may be fully open source; others may have proprietary elements. The goal is to make governance more operational, and over time, this type of Governance Commons builds toward a standard for responsible AI.

_____________________________________________________________

Collaborate with us!

As always, we appreciate you taking the time to read our blog post.

If you have news relevant to our global WAI community or expertise in AI and law, we invite you to contribute to the WAI Legal Insights Blog in 2026! To explore this opportunity, please contact WAI editors Silvia A. Carretta - WAI Chief Legal Officer (via LinkedIn or silvia@womeninai.co) or Dina Blikshteyn  (dina@womeninai.co).

Silvia A. Carretta and Dina Blikshteyn

- Editors