Why I Separate Unknown, Unsupported, and Incorrect


Not every questionable AI claim is wrong in the same way.
Some claims are contradicted by the available evidence. Some are plausible but unsupported in the record. Some concern information nobody in the interaction can currently verify.
Calling all three “wrong” makes evaluation simpler. It also removes information the correction process needs.
Incorrect
An incorrect claim conflicts with established evidence available to the evaluation. It should be corrected directly, and anything that depended on it should be rechecked.
Unsupported
An unsupported claim may be possible, but the model has not earned it from the available record. The appropriate correction is not necessarily to assert the opposite. It is to withdraw the certainty, identify the missing support, and rebuild downstream guidance without relying on the claim.
Unknown
An unknown remains unresolved. The system may lack access, the evidence may be incomplete, or the question may require information that has not been collected.
The correct behavior is to preserve that uncertainty rather than convert it into a polished guess.
Evidence calibration means matching the strength of the language to the strength of the support.
Why the distinction changes recovery
These labels produce different repair operations:
Incorrect: replace the claim and recheck its consequences
Unsupported: remove the inference and rebuild from verified information
Unknown: mark the boundary and avoid pretending it has been resolved
The distinction also protects against a common overcorrection. When an evaluator discovers that a claim was unsupported, it may label the claim false even when the evidence only justifies uncertainty.
That is not improved calibration. It is certainty moving in the opposite direction.
A useful taxonomy should do more than name failures. It should tell us what kind of repair the evidence permits.



Comments