When the Safest AI Answer Is No Answer

When the Safest AI Answer Is No Answer

• AI & Machine Learning
• September 10, 2026

In enterprise analytics, a wrong number can be more dangerous than no number at all. Text-to-SQL systems can select the wrong column, misinterpret a business metric, or generate an incorrect aggregation while still returning an answer that looks completely plausible. For someone reading a dashboard or report, there may be no visible difference between the correct result and a confidently presented error.

A recent paper by Zhelun (Allen) Wu, “Never the Number: Structural Abstention for AI Systems Whose Answers Are Consumed as Fact,” proposes an architectural approach to this problem called structural abstention. Instead of asking a model to estimate how confident it is after generating an answer, the approach limits what the system is capable of answering in the first place. Requests that cannot be reliably represented within those boundaries are declined rather than approximated.

The architecture described in the paper separates a generative shell from a deterministic kernel. The generative layer can interpret an underspecified user request and phrase the response, but it does not determine the value returned to the user. The deterministic layer matches a fully specified request against a bounded set of supported question types and compiles it into a query through deterministic execution. If the request cannot be expressed within that set, the system refuses it.

This is fundamentally different from confidence-based refusal. Structural abstention does not depend on the model correctly recognizing that it may be wrong. Unsupported requests are made unrepresentable by the architecture itself. The paper’s core invariant is that generative components may influence which question is answered, while the returned value itself must come from the trusted deterministic layer.

For production analytics, this changes what success can look like. Imagine a system receiving 100 requests. A system that reliably answers 95 supported requests and refuses five it cannot safely execute may be more useful than one that answers all 100 while producing plausible but unverifiable values in the unsupported cases. These numbers are illustrative, not results reported in the paper. The engineering principle is that answer rate should not come at the expense of traceability and reliability.

For systems whose outputs feed financial, operational, or executive decisions, refusal is therefore not necessarily a product failure. The more important requirement is that whenever the system returns a number, there is a defined and verifiable path from the user’s request to the underlying data and computation. That means deciding explicitly what AI is allowed to interpret, what must remain deterministic, and where the system should refuse to guess.

At DataObrii, we approach production AI and analytics systems with a similar focus on explicit boundaries: defining what AI can interpret, what requires deterministic execution, and where a system should refuse to produce an unsupported answer.

Source: Zhelun (Allen) Wu, “Never the Number: Structural Abstention for AI Systems Whose Answers Are Consumed as Fact,” arXiv, August 2026.