Introduction

On 7 October 2023, following the escalation of the Israel-Hamas conflict, the Israel Defense Forces reportedly deployed an AI decision-support system known as ‘Lavender’ to generate target recommendations in Gaza. According to investigative reporting by +972 Magazine and Local Call, the system generated at least 37,000 target recommendations within the first six weeks of the conflict. Analysts reportedly reviewed these recommendations, which were, in some cases, passed to a military commander for final approval. According to the Israeli Defence Force, Lavender draws on geospatial intelligence, signals intelligence, human sources, and open-source information, but it is not known exactly how the system weighs this data to produce a score. No state or international body has been able to reconstruct the basis on which any individual recommendation was generated. This is the technopolar world where power is exercised by algorithmic systems whose reasoning resists reconstruction even by those who operate them.

Nation-states have defined the global order since Westphalia, but today they face competition not only from technology companies but from the AI systems that those companies build. The technopolar literature has documented who holds this power well. However, it has paid less attention to how that power is increasingly exercised not by identifiable executives making traceable decisions, but by AI systems whose decision-making cannot be reconstructed at all. Lavender was not autonomous from the state; it was built and deployed in alignment with state military objectives, illustrating that the shift is not in who holds power but in the form that power now takes. As systems of this kind, typically non-deterministic models trained on patterns rather than fixed rules, displace human decision-makers in consequential domains such as targeting, content moderation, credit, and infrastructure management, the problem compounds. Power is exercised, outcomes are produced, and no causal chain can be traced. The black box is not a technical inconvenience. It is a governance problem.

This article argues that international AI governance frameworks, the EU AI Act, the UNESCO Recommendation on the Ethics of Artificial Intelligence, and the emerging treaty process on lethal autonomous weapons systems mandate transparency, explainability, and human control as if these properties are equally achievable across all AI systems. They are not. Whether accountability is achievable depends on the architecture of the system in question. A framework which cannot distinguish among different architectures cannot regulate them efficiently. This article shows why and proposes what architecture-sensitive governance would look like. It focuses on the international dimension and does not address domestic regulation or platform governance in depth. 

The Architecture Problem at the International Level

International governance instruments have so far declined to draw a distinction between different architectural families of AI systems. Three families exist, each with different implications for traceability and legal accountability.

First, Symbolic AI systems operate through explicit logical rules and structured knowledge representations. Their reasoning is legible by design and is a chain of inferences that can be read, audited, and replayed. Expert systems in law, medicine, and finance have historically operated on this basis.

Second, Connectionist AI systems are deep neural networks and large language models which derive outputs from statistical patterns encoded across billions of weighted parameters. Neural networks operate as black-box models, which makes their decision-making processes difficult to understand. The field of mechanistic interpretability seeks to address this by reverse-engineering model circuitry, but fundamental challenges remain, and superposition creates significant barriers to feature-level analysis. Connectionist systems may be opaque not merely in practice, but in principle.

Third, Neurosymbolic AI systems combine both approaches, using neural networks for perception while routing higher-level reasoning through symbolic components. Symbolic layers are rule-driven, and every inference is traceable back to explicit logic or institutional knowledge. The result is a system that is more tractably auditable than a purely connectionist one, though the neural components remain partially opaque, and interactions between neural and symbolic layers can complicate human understanding of the system’s overall operation. Neurosymbolic systems are not fully transparent. 

International law’s accountability frameworks presuppose traceability. The ILC Draft Articles on State Responsibility require that conduct be attributable to a state or other actor. Attribution requires a connection between an output and a decision process. For connectionist systems, that connection cannot be established with the specificity the framework demands. The Articles require proof that an act was directed or controlled by a state, but where the act is generated by a system whose internal workings cannot be interpreted, no such proof is constructible. Attribution becomes a legal fiction, a responsibility assigned without a traceable causal chain.

The governance implication follows directly. Frameworks applying identical accountability rules to connectionist and neurosymbolic systems produce hollow compliance, either imposing explainability obligations that connectionist systems cannot genuinely meet, or granting no credit to neurosymbolic systems that structurally can. The result is the appearance of regulation, not its substance.

Where Existing Frameworks Fail

Three instruments illustrate this problem.

The EU AI Act and the Explainability Mandate

Article 13 of the EU AI Act requires that high-risk AI systems be sufficiently transparent to enable deployers to interpret system outputs and maintain meaningful control. The obligation applies uniformly: a high-risk connectionist system and a high-risk neurosymbolic system face identical compliance requirements.

For neurosymbolic systems, the requirement is achievable. Symbolic layers log every inference step; the reasoning can be audited and replayed. For connectionist systems, no such native traceability exists. Mechanistic interpretability seeks to investigate individual neurons and their connections in circuits, but has not produced methods allowing a regulator to reliably trace why a specific output was generated from a specific input. Transparency obligation under Article 13 is therefore technically meaningful for one class of systems and only aspirational for another. By treating them identically, the Act cannot distinguish substantive compliance from post-hoc approximation.

The UNESCO Recommendation and the Limits of Principle

UNESCO’s Recommendation on the Ethics of Artificial Intelligence, adopted in November 2021 by all 193 member states, identifies transparency and explainability as foundational principles. It holds that the logic behind algorithmic decision-making should be fully interpretable by experts and explainable to users in an accessible language.

As a matter of technical fact, this is not currently achievable for large-scale connectionist systems. Neurosymbolic systems can provide clearer explanations because the symbolic component follows explicit rules whose reasoning steps can be traced and verified. By framing explainability as a universal ethical principle rather than a technically contingent one, the UNESCO Recommendation conflates what is normatively desirable with what is technically feasible. States implementing it through domestic legislation inherit this conflation. For instance, Colorado’s SB24-25 Consumer Protections for Artificial Intelligence imposes a duty upon the deployer of a high-risk AI system to provide a statement explaining the reasons due to which the system makes a decision significantly affecting a consumer. Connectionist AI systems are not inherently capable of producing such explanations. Therefore, they risk enacting obligations that satisfy consumers formally, through surrogate explanation layers, without achieving the substantive interpretability the guidelines require. This results in the creation of a façade of transparency and explainability. 

Lethal Autonomous Weapons Systems (“LAWS”) and Meaningful Human Control

The Convention on Certain Conventional Weapons process offers the most consequential illustration. The CCW Group of Governmental Experts’ rolling text from November 2024 requires that LAWS be predictable, reliable, traceable, and explainable, and that context-appropriate human oversight be maintained over target identification and engagement.

These requirements presuppose that operators can understand and verify a system’s target selection process. For connectionist LAWS, this fails: engagement decisions emerge from opaque parametric weights, not explicit rules auditable against International Humanitarian Law’s distinction and proportionality requirements. For instance, Turkey introduced the STM Kargu-2, an LAWS programmed to attack targets with no requirement for data connectivity between the weapon system and the operator. This munition, powered by connectionist AI, allows an individual to autonomously track and bomb targets. Habsora, introduced by Israel, also uses connectionist AI to automate the analysis of a large amount of intelligence to identify potential targets, at a space much greater than humans. These weapons wholly disregard the principles of proportionality and distinction. A neurosymbolic targeting system, where the symbolic component encodes explicit engagement criteria, is structurally more amenable to International Humanitarian Law compliance as its reasoning can, in principle, be audited against the prohibition on attacks on civilians.

Both system types may produce identical observable outputs. The difference is whether the decision process behind those outputs is legally traceable. The rolling text does not draw this distinction, establishing compliance standards that connectionist LAWS cannot substantively meet and giving neurosymbolic LAWS no credit for structurally exceeding them. In a domain where accountability failures cost lives, this is not a marginal omission.

The Stakes in the Technopolar World

In a world where influence is determined not by physical or military might but by control over technology, artificial intelligence, and algorithms, the explainability of the decision-making process becomes extremely important. There is no common international framework to govern AI, and several countries are still grappling with ways to regulate AI systems. In such a scenario, ensuring explainability and accountability becomes a task. Current frameworks make this task harder by conflating different types of AI into one. A company may use a connectionist system that fundamentally gives rise to a black box effect and is difficult to trace. However, these companies can then add an explanation layer that is erroneous, misleading, and does not necessarily have any relationship with how the system works. A study carried out by the Institute of Informatics at the University of Warsaw found that AI governance becomes merely a formal compliance exercise unless governments deal with the issue of explainability through laws that acknowledge the technical realities of AI systems. 

While a regulation that is sensitive to the different kinds of AI architecture may not completely solve the problems of accountability and explainability in the technopolar world, it provides clarity on whether such explainability is always possible. It would bring about a distinction between systems where accountability and explainability are technically feasible and systems where they are not, and prompt different regulatory responses to each. Without such a distinction, legal obligations can risk collapsing into a box-ticking exercise, with there being only the appearance of and not actual explainability. More importantly, it determines whether responsibility can be meaningfully attributed at all, or whether governance frameworks merely simulate control over systems that remain beyond it, in practice. 

What Architecture-Sensitive Governance Would Look Like

Firstly, architecture-sensitive AI governance requires changes that align regulatory expectations with the technical capacities of different AI systems. First, international instruments should distinguish between AI architectures such as connectionist and neurosymbolic systems. For instance, in the EU-AI Act, this can be done through annexures that complement the existing classifications based on risk. 

Secondly, technical bodies such as the International Telecommunication Union and the International Organization for Standardization must evaluate an organization’s operational systems. This can enforce architecture-sensitive governance by differentiating between systems such as neurosymbolic AI that inherently permit explainability, and connectionist systems such as large language models, where explainability is difficult.

Additionally, when it comes to India, the question that has recently arisen before the Registrar of Copyrights following the petition filed by Stephen L Thaler before the Delhi High Court, is an opportunity to start a discussion about architecture-sensitive AI governance. The question at hand is whether Thaler’s AI system, DABUS, can be recognised as an author. It is important to note that DABUS is not a standard connectionist system but a neural network-based system operating on internally generated chains of concepts. The Court has a real opportunity here to begin a broader conversation about how different AI architectures function and how they ought to be governed. Also, in the aftermath of the AI-Impact Summit, India must start a discussion about architecture-sensitive standards that could possibly be replicated in international frameworks. 

These interventions would not just make the frameworks that currently exist more accurate, but also provide clarity on what accountability can actually be asked for.

Conclusion

The technopolar world must be governed by frameworks that clearly recognize the differences between types of AI systems. The attribution of responsibility, obligations of transparency, and human control requirements all become more coherent when applied to systems whose architecture is specified and not assumed. The barrier is not legal. It is conceptual as governments and regulators do not ask what kind of system they are regulating before deciding what they require of it. That question is the precondition for any governance framework that is actually effective.

Maanasi Shivakumar and Kashvi Garg are fourth-year B.A.,LL.B. (Hons.) students at Jindal Global Law School.