How fact-checking benchmarks conflate missing evidence with insufficient evidence—and what to do about it

Researchers have identified a hidden problem in how machines learn to classify claims as unverifiable: when benchmark datasets label something "not enough information" (NEI), they're mixing up two fundamentally different things—the actual absence of relevant evidence, and the dataset's particular method of creating that absence. The result is that AI systems trained on these benchmarks learn shortcuts rather than genuine reasoning.

The problem: construction artifacts#

The gap sits in how fact-verification datasets are built. When researchers construct NEI examples, they do so by applying specific techniques: removing certain evidence sources, pairing unrelated evidence with claims, or other engineered conditions. The choice of which technique to use shapes what a model learns, but this choice is rarely made explicit or tested. An AI trained on one construction method often cannot recognize insufficient evidence created by a different method, even when the underlying semantic relationship is identical.

A construction-aware diagnostic approach#

The study introduces NEI-CAP, a diagnostic framework that treats each NEI example as tagged with its construction family. The protocol then:

• Audits shortcut cues that models might exploit rather than learn genuine reasoning
• Validates hard cases through human review to catch cases where the dataset itself is ambiguous
• Tests whether a model's competence transfers across different construction methods

What the tests revealed#

Using this protocol on the SciFact benchmark (with FEVER and HoVer as comparison points), the researchers found that competence is not portable. Encoder-based verifiers and instruction-tuned decoders trained on one construction method consistently fail on semantically similar insufficient-evidence cases created via a different method. Even when training on a mixture of constructions, the gap narrows but remains significant.

Perhaps more troublingly, they found that the evidence condition shifts not just how often a model correctly identifies NEI cases, but also its confidence in the opposite labels (Support or Refute). This means a single aggregate NEI score can hide whether a model has actually solved the problem it claims to have solved.

Implications for benchmarking