Symbols Do Not Make AI Governed

Generative AI is criticized for hallucination, opacity, and brittle reasoning. Neurosymbolic AI is increasingly offered as the corrective: add symbols, rules, logic, knowledge structures, and explicit reasoning traces to systems that were previously statistical black boxes, and the output becomes more structured and more inspectable.

The improvement is real. Rule-informed inference can be more reliable than free generation, and an explicit reasoning chain is easier to examine than a probability distribution. But the improvement invites a familiar category error. The assumption is that a better reasoning architecture produces governed execution. It does not. Better reasoning changes the quality of what a system proposes. It does not change whether an organization is permitted to act on the proposal.

01Correctness is not permission

FERZ has already drawn the relevant boundary. In Deterministic Engines and Deterministic Governance, deterministic engines may produce correct, bounded, and reproducible outputs, yet they do not answer whether an organization is permitted to act on those outputs. Correctness is a property of the computation. Permission is a property of the authorization decision, bound to policy state, evidence, authority, and time.

02Where neurosymbolic systems sit

The engines paper already classifies symbolic and hybrid reasoning engines as deterministic engines. Placing neurosymbolic AI on the engine side of the authorization boundary is therefore not a new claim. It is an instance of a published taxonomy.

A neurosymbolic system may produce cleaner representations, explicit reasoning chains, structured candidate actions, and rule-informed recommendations. Each of those is an output. None of them is a permission. A neurosymbolic system may improve the quality of a proposal. It does not, by that fact, authorize the proposal.

03Why symbols look like governance

Symbols carry institutional connotations. Rules look legal. Logic looks auditable. A reasoning trace looks like evidence. None of those appearances is an authorization artifact.

A neurosymbolic reasoning trace may improve inspection, but inspection is not admissibility. In ABIM terms, a trace can be bound into the record without satisfying input integrity. Binding records which inputs were used; it does not establish that those inputs earned their place. A symbolic trace is therefore not an authorization artifact.

The test that decides governance is unchanged. Was the action authorized before execution, under the applicable policy state, authority chain, admissible evidence set, and runtime authorization boundary, in a way that can be independently reconstructed? A system can be explainable, and even formally constrained, and still fail that test.

04The objection, and why it does not hold

The strongest objection from a neurosymbolic advocate is not that symbols are weak. It is that a neurosymbolic system can encode rules, including authorization rules, and can therefore be the authorization layer.

The answer is in the proof target. An engine proves properties of outputs under a model. Governance proves permission under policy state, evidence provenance, semantics, authority chain, and time. A neurosymbolic system may still use neurosymbolic components upstream. But at the boundary, once it binds a verdict to those obligations and produces a result an independent party can reconstruct, the relevant proof target is no longer reasoning quality. It is authorization legitimacy. That is a different problem, with different architectural requirements.

05A worked example: loan modification

06Capability does not confer authorization

The same holds if neurosymbolic methods turn out to be a path to more general capability. A more capable system may reason, plan, and explain better. None of that determines whether it may bind an institution, release a payment, prescribe a drug, or change an account state. Even if neurosymbolic AI improves capability, capability does not confer authorization. Greater capability raises the stakes of the authorization boundary. It does not remove it.

07FERZ governs the authorization boundary

FERZ does not compete with neurosymbolic AI. FERZ governs the runtime authorization boundary around it. The model underneath may be generative, symbolic, neurosymbolic, deterministic, probabilistic, or hybrid. FERZ treats its output as a proposed action and determines, before execution, whether that action may proceed, then records that determination as an authorization artifact that an independent party can reconstruct.

08Conclusion

Neurosymbolic AI is useful. Symbols matter. Structured reasoning matters. Formal constraints matter. None of them makes AI governed. Governance begins when execution is conditioned on a pre-execution authorization verdict, captured in an authorization artifact, and capable of independent reconstruction.

Symbolic reasoning is an engine property. Authorization is a governance property.

This article applies the distinction developed in Deterministic Engines and Deterministic Governance to the current neurosymbolic AI narrative.