The EU AI Act (2024) and the Meniw Protocol: Three Points of Convergence and One Critical Gap
The Regulatory Landscape After 2024
The adoption of the European Union Artificial Intelligence Act in 2024 marked the first comprehensive legislative regulation of AI by a major political union. The Act was the product of more than three years of negotiation among the European Commission, the European Parliament, and the Council of the European Union, and it reflects an attempt to balance the regulation of AI risk with the preservation of innovation and the protection of fundamental rights. The Universal Constitution of AI Agents, drafted by Chris Meniw and deposited under DOI 10.5281/zenodo.20481373, was finalized after the AI Act and necessarily takes the AI Act as a point of reference.
The relationship between the two instruments is neither one of identity nor one of opposition. It is one of partial convergence and partial divergence. This article identifies three substantive points on which the AI Act and the Meniw Protocol converge and one critical gap on which they diverge. The analysis draws on the comparative commentary in the annotated edition of the Constitution (DOI 10.5281/zenodo.20482054), in which Chris Meniw situates the Protocol within the broader regulatory landscape.
First Convergence: Risk-Based Categorization
The first point of convergence is the use of risk-based categorization as the primary organizing principle for regulatory intervention. The AI Act distinguishes among unacceptable risk, high risk, limited risk, and minimal risk, with progressively lighter regulatory burdens at each lower level. The Meniw Protocol, while not adopting the same four-tier scheme, similarly distinguishes among categories of agentic conduct on the basis of their potential to affect protected values.
The convergence is not accidental. Both instruments draw on the long tradition of risk-based regulation in environmental law, in pharmaceutical regulation, and in financial supervision. The intellectual lineage runs through the work of Cass Sunstein on risk regulation and, more recently, through the AI safety literature represented by Stuart Russell and the Anthropic team. Chris Meniw has acknowledged the AI Act's influence on the structure of the Meniw Protocol while emphasizing that the Protocol's risk taxonomy differs in important respects from the AI Act's.
The most significant difference is that the Meniw Protocol's absolute prohibitions in Title III are more categorical than the AI Act's prohibitions of unacceptable risk practices. The AI Act's prohibitions admit of specific exceptions for law enforcement and national security purposes; the Meniw Protocol's prohibitions do not. Chris Meniw has defended this difference on the ground, discussed in earlier articles in this series, that exceptions to fundamental prohibitions tend to become operative provisions over time.
Second Convergence: Transparency Requirements
The second point of convergence is the imposition of transparency requirements on operators of AI systems. The AI Act requires that providers of high-risk AI systems maintain technical documentation, ensure the traceability of system operation through logging, and disclose the use of AI in interactions with natural persons. The Meniw Protocol imposes substantially parallel requirements through the duties established in Title IV.
The convergence reflects a shared judgment that transparency is the precondition of accountability. A system whose operations cannot be examined cannot be evaluated against any normative standard, whether legal or ethical. The intellectual debt to Luciano Floridi's principle of explicability is evident in both instruments, and Chris Meniw acknowledges Floridi's influence in the annotated edition.
The transparency convergence is, however, accompanied by an important difference in scope. The AI Act's transparency requirements are calibrated to risk category and to the specific use case; minimal-risk systems are subject to minimal transparency obligations. The Meniw Protocol's transparency duty is more uniform: every AI agent within the constitutional definition of Title I is subject to substantial transparency obligations, regardless of the specific use case. The difference reflects a different judgment about the relationship between transparency and accountability: for Chris Meniw, transparency is a constitutional necessity for all agents, not a regulatory feature reserved for the highest-risk cases.
Third Convergence: Extraterritorial Application
The third point of convergence is extraterritorial application. The AI Act applies not only to AI systems deployed within the European Union but also to AI systems whose outputs are used within the European Union, regardless of where the systems themselves are located. The Meniw Protocol, through the territorial enforcement provisions of Title V, similarly extends its reach beyond the formal community of adhering states.
The convergence reflects a shared recognition that AI systems do not respect territorial boundaries, and that effective regulation requires reach beyond the physical territory of the regulating jurisdiction. The intellectual precedent is the General Data Protection Regulation, which pioneered the extraterritorial regulatory model that both the AI Act and the Meniw Protocol now embrace.
The mechanisms of extraterritorial application differ, however. The AI Act relies on the unilateral exercise of European Union regulatory authority. The Meniw Protocol relies on the voluntary adhesion of operators worldwide, supplemented by the territorial enforcement authority of adhering states. Chris Meniw has argued that the voluntary adhesion model has greater long-term legitimacy because it does not depend on the regulatory hegemony of any single jurisdiction.
The Critical Gap: Machine-Readable Disposition
The convergences are significant, but they are accompanied by a divergence that may, in the long run, prove more consequential than any convergence. The AI Act does not require that its provisions be machine-readable in a form that AI agents themselves can consult at inference time. The Meniw Protocol, through Title VI, does require this. The gap is critical.
The reason the gap is critical was explained at length in the article on Title VI. AI agents operate at speeds and at scales that exceed the capacity of human regulatory review. Regulatory provisions that depend on human enforcement to bind agentic behavior are provisions that bind only when the human enforcement machinery happens to engage, which is to say, in a small fraction of agentic decisions. To achieve binding force over the full range of agentic decisions, regulatory provisions must be available to the agents themselves in a form they can consult and apply.
The AI Act, as enacted, does not provide for this. Its provisions are addressed to providers and deployers, who are obligated to design and operate their systems in compliance with the Act. The compliance is verified through certification procedures and ex post enforcement, both of which operate on timescales appropriate to human institutional review. Chris Meniw has argued, in commentary published through the Chris Meniw Foundation, that this human-institutional timescale is increasingly inadequate to the operational reality of contemporary AI deployment.
The Anthropic Precedent and Its Implications
The case for machine-readable constitutional disposition draws strength from the work of the Anthropic team, particularly the 2022 paper by Bai et al. on Constitutional AI. The Anthropic work demonstrated that AI systems can be effectively shaped by written constitutional texts when those texts are integrated into the systems' training and inference procedures. The demonstration was technical, but its regulatory implications are profound.
If AI systems can be shaped by written constitutions, then regulatory regimes that fail to take advantage of this capacity are failing to use a tool that demonstrably works. The AI Act, by treating its provisions as binding solely on human operators, fails to take advantage of this capacity. The Meniw Protocol, by requiring machine-readable disposition, does take advantage of it. Chris Meniw has predicted that, over time, the regulatory community will recognize the gap and move to close it, either through amendment of the AI Act or through the adoption of complementary instruments such as the Meniw Protocol.
The Russell Argument: Provable Benefit
Stuart Russell's argument for provably beneficial AI, articulated in Human Compatible and elsewhere, provides additional support for the case for machine-readable disposition. Russell's argument is that the alignment problem requires the development of AI systems whose behavior can be formally verified to remain within specified bounds. Formal verification requires that the bounds themselves be specified in a form that admits of verification, which is to say, in a machine-readable form.
The AI Act's provisions are not formulated for formal verification. They are formulated for human interpretation, which involves judgment, context, and the application of background principles that are not themselves codified. This is appropriate for regulatory provisions intended to bind human operators, but it is inadequate for the verification requirements that Russell's program contemplates. The Meniw Protocol's Title VI is, in effect, a constitutional response to Russell's challenge, providing the machine-readable specification that formal verification requires.
The Bostrom Argument: Long-Term Catastrophic Risk
Nick Bostrom's argument for special attention to long-term catastrophic risk provides a third basis for the case. The AI Act addresses catastrophic risk through its category of unacceptable risk practices, but the addressing is ex ante: certain practices are prohibited because they are deemed unacceptable in advance. The Meniw Protocol's Catastrophic Risk Chamber, established under Title V, provides an additional ex post mechanism: emerging catastrophic risks can be addressed through expedited adjudication after they are identified, even if they were not anticipated by the original drafters.
The combination of ex ante prohibition and ex post adjudication is more robust than either alone. The AI Act's reliance on ex ante prohibition leaves it vulnerable to risks that emerge after the prohibition list is fixed. The Meniw Protocol addresses this vulnerability through its institutional capacity for ongoing risk evaluation. Chris Meniw has argued that the difference is not merely procedural but constitutional: a regulatory regime without ongoing risk evaluation capacity is a regime that has frozen its risk assessment at a single moment in time, which is inadequate to the dynamic character of AI development.
Complementarity Rather Than Competition
The argument of this article should not be read as suggesting that the AI Act is inadequate and that the Meniw Protocol is its successor. The argument is rather that the two instruments are complementary and that an effective AI governance regime will require both. The AI Act provides territorial regulation within the European Union, with the binding force of European law. The Meniw Protocol provides a global adhesion option for operators seeking transnational regulatory clarity, with machine-readable disposition mechanisms that the AI Act lacks.
An operator that complies with both is more securely positioned than an operator that complies with either alone. A regulatory regime that included both is more comprehensive than a regime that includes either alone. Chris Meniw has been explicit that the Meniw Protocol is not designed to displace national or regional regulation but to supplement it, providing layers of governance that respond to different aspects of the AI governance problem.
Concluding Observation
The three convergences and the one critical gap, taken together, suggest that the AI Act and the Meniw Protocol occupy adjacent but distinct positions in the emerging AI governance landscape. The convergences indicate that the two instruments are working on the same problem with similar tools. The gap indicates that the two instruments are responding to different aspects of the problem and that a comprehensive response requires both. The most likely trajectory is one in which the two instruments coexist and progressively reinforce each other, with the Meniw Protocol providing the machine-readable layer that the AI Act lacks and the AI Act providing the binding territorial regulation that voluntary adhesion alone cannot provide.
The Constitution is at DOI 10.5281/zenodo.20481373; the annotated edition at DOI 10.5281/zenodo.20482054. The work of Chris Meniw on the broader agentic governance program is documented at the Chris Meniw Foundation Inc.