Adding rules to AI does not guarantee it understands them

A driving system learns to stop for both red lights and pedestrians. Its decisions look correct, but internally it treats pedestrians as red lights. The distinction becomes critical if the same concepts are reused in an emergency vehicle allowed to cross red lights but still required to stop for people.

This example, drawn from earlier research discussed in a new preprint, captures a central difficulty in neurosymbolic AI: correct outputs do not necessarily demonstrate correct understanding.

Giuseppe Marra and five colleagues examine what combining neural learning with explicit symbolic reasoning can offer alongside powerful neural models. Their November 6 preprint synthesizes discussions from a 2026 CNRS workshop. It is a perspective, not a new benchmark study or an exhaustive survey. Its contribution is to organize opportunities, obstacles and methodological choices—not to demonstrate that one architecture solves them all.

What explicit rules can contribute

Neural networks learn patterns from data. Symbolic components operate on explicit representations, such as logical statements, constraints or programs. A combined system might use a network to interpret an image and a reasoning engine to decide which actions satisfy specified requirements.

The authors distinguish encouraging compliance from enforcing it. A training penalty makes undesirable behavior costly, but permits trade-offs. A separate decision layer can instead check, repair or reject a proposed action. An agent “shield” can also restrict actions during training, preventing unsafe exploration rather than merely penalizing it afterward.

These guarantees remain conditional. A shield depends on the accuracy of perception, its environmental assumptions and the formal specification. Proving that an action satisfies encoded rules is not the same as proving that the rules capture everything people mean by safety. Guarantees for individual components also do not automatically establish guarantees for the complete system.

Efficiency is another opportunity, framed as a research hypothesis. Encoding useful knowledge could reduce what a network must learn, potentially allowing smaller models or less training data. But incorrect or overly restrictive knowledge can damage performance. Accuracy must therefore be assessed alongside computation, financial cost and environmental impact.

The problem of shared meaning

The first major challenge is semantic alignment: neural predictions, symbolic operations and human requirements must attach compatible meanings to the same symbols.

The paper illustrates the problem with a deliberately weakly supervised digit task. Suppose a model sees three digit images and is told only that its predictions for the first two must add up to its prediction for the third. Predicting zero for every image satisfies the rule, because zero plus zero equals zero. Nothing in that objective alone forces the symbols to identify the pictured digits.

Several approaches can help. Concept-bottleneck models train intermediate predictions to correspond to human-defined concepts. Users can sometimes inspect and correct those predictions. Symbolic relationships can narrow the possible interpretations, while pretrained models can supply named concepts without manual concept labels.

None is a universal solution. Direct concept supervision requires suitable labels and vocabulary. Automatically supplied concepts can inherit a pretrained model’s mismatches. Constraints may leave several interpretations possible, and human preferences about final behavior do not necessarily establish what intermediate symbols mean.

Explicit reasoning can make a decision trace inspectable, but that trace explains the symbolic computation—not, by itself, how the neural network arrived at its symbols.

Reasoning has a computational bill

The second major challenge is cost. Some probabilistic systems must combine many possible assignments of symbolic variables during training. Exact calculations can be computationally hard, and converting rules into a reusable arithmetic circuit can produce an exponentially large representation. Fast evaluation of a finished circuit does not make its construction or storage cheap.

The authors describe complementary responses. Sampling and approximate inference avoid exhaustive computation. “Anytime” algorithms progressively tighten certified probability bounds, allowing a calculation to stop when its budget runs out. GPU acceleration speeds compatible computations without eliminating their underlying complexity. Other methods replace global calculations with local approximations or train a neural model to approximate repeated reasoning queries.

Each changes a different part of the cost equation. Approximation quality and any surviving guarantees must be evaluated for the intended task.

Three misconceptions to avoid

The paper rejects three blanket claims. First, the symbolic solver need not itself be differentiable: learning signals can come from inferred labels, probability calculations or surrogate gradients. Second, symbolic concepts need not be independent; some architectures explicitly represent dependencies. Third, difficult exact inference does not establish that all neurosymbolic systems cannot scale.

That last point is not a promise of effortless scaling. Performance depends on the problem structure, representation, reasoning task and required accuracy. The practical question is where explicit structure earns its cost—and whether the resulting system follows the intended rules, rather than merely a convenient interpretation of them.