Neuro-symbolic AI is sold as interpretability's homecoming. The architecture that finally bolts a readable rule layer back onto the neural network and hands an auditor a decision they can trace. I'd been ready to believe it, because the history underneath the pitch is there. But the pitch has a hole in it. A symbolic layer does not make a system auditable. Depending on where you put it, it can do something worse than stay opaque: it can hand you a clean, legible rationale that has nothing to do with why the model actually decided what it decided. A black box that comes with its own alibi.
What we actually lost
The reason "return to the glass box" lands as a slogan is that there was a glass box, and we did lose it. Early symbolic systems (expert systems, rule engines, the whole GOFAI lineage) were auditable by construction. The rule chain wasn't a description of the decision; it was the decision. MYCIN could walk you backward through the inferences that led it to recommend one antibiotic over another, because those inferences were the mechanism, not a summary written afterward. Transparency wasn't a feature anyone added. It fell out of how the thing was built.
Then statistical methods won, for the obvious reason that they worked on problems symbolic systems couldn't touch. Deep learning scaled. And in the scaling we traded the property away, not by accident, but as an engineering consequence. A model with a billion weights has no rule chain to read. The EU AI Act's transparency and traceability obligations are, read honestly, an attempt to legislate back a property the field had as a default until the paradigm shift took it out. [1] XAI (SHAP, LIME, attention maps) is the same move from the technical side: post-hoc patches that reconstruct a plausible story about a decision that was never made in those terms.
So when neuro-symbolic systems arrive promising to weld the symbol back on, the appeal is obvious. The only question worth asking is whether the weld holds.
Where the symbol sits
Neuro-symbolic is not one architecture, and the entire audit question turns on which one you're looking at: specifically, where in the pipeline the symbolic part lives.
One family puts the symbols upstream, as constraints. The neural network proposes, but a logical layer disposes: hard rules the output has to satisfy, semantic constraints folded into the loss, a reasoning step the model genuinely cannot override. If a rule says a credit decision may not turn on a protected attribute, and that rule is a real gate rather than a suggestion, then the symbol is load-bearing. It constrains the output. That is auditable, and it is the version the marketing has in mind.
The other family puts the symbols downstream, as explanation. The neural network decides, and a symbolic module (increasingly, a language model prompted to reason) produces a tidy chain of steps alongside the decision. From the outside it looks identical to the first family. You get a rule chain. It reads beautifully. And it may be pure narration: a justification generated to be plausible, with no guarantee it reflects the computation that produced the answer.
This is the distinction interpretability researchers call faithfulness versus plausibility. A plausible explanation is one a human nods at. A faithful one actually corresponds to the model's reasoning. The two come apart constantly, and a downstream symbolic layer is built for the first. It is, structurally, a plausibility engine. The chain of thought it emits can be entirely post-hoc, the same way a person confabulates a confident reason for a choice the gut had already made.
The alibi problem
An opaque model at least declares itself opaque. It fails the transparency requirement loudly. Nobody mistakes it for traceable, so you know to compensate with other controls. A neuro-symbolic system with a narration layer fails quietly. It produces exactly the artifact a transparency obligation asks for: a legible, step-by-step rationale. The box gets ticked. Someone reads a rule chain that passes the smell test and moves on. Nobody establishes that the stated reason is the operative reason, because the system was built to make that check feel unnecessary.
That is the failure mode paperwork cannot catch on its own. The Act wants traceability; a confabulating symbolic layer satisfies the letter of it while inverting the intent. [1] The point of demanding an explanation was to expose the decision. An unfaithful explanation does the opposite: it launders an opaque decision into a compliant-looking one. The documentation gets better while the accountability gets worse. A clean rationale you haven't verified is not evidence of anything. It's a liability dressed as a control.
What I'd actually test for
The honest version of this: I haven't audited a production neuro-symbolic system. I'm reasoning from architecture, not from a folder of findings, and the field moves fast enough that specific systems will look different in a year. So take this as a framing, not a methodology.
But the framing already tells me what I'd refuse and what I'd test. I would not accept the presence of a symbolic layer as evidence of anything. Presence is cheap. What matters is whether the symbols are causally load-bearing. Ablation: remove or change a rule and see whether the output changes, or whether the model decides the same thing and merely rewrites the story. Counterfactuals: feed it a case the stated rule should flip, and watch whether the decision flips with it. If the rationale stays stable while the decision moves, or the decision stays stable while you break the rule it claims to follow, the symbolic layer is decoration.
Neuro-symbolic AI might genuinely be a road back to auditability. The promise is conditional, and the condition is the whole thing. Symbols that constrain are a control. Symbols that narrate are an alibi. They are indistinguishable on the page, which is why the only useful question is not whether the system can explain itself, but whether its explanation is allowed to be wrong about what the system did.