When Hallucination Is More Than a Wrong Fact: Measuring Reliability Through the User

TL;DR for operators

A model change can improve an automated hallucination benchmark while leaving users dissatisfied for a different reason: sources are hard to verify, reasoning appears unsupported, false claims are stated with confidence, or corrections are ignored. The System Hallucination Scale (SHS) gives teams a structured way to measure those experiences across five dimensions rather than reducing reliability to a binary factual-error judgment.¹

Its ten questions are arranged as positive/negative pairs. That matters because SHS extracts two different signals from each pair: the direction of the user’s reliability judgment and whether the pair itself was answered coherently. In an initial 210-participant validation, the instrument showed high internal consistency, reasonably differentiated dimensions, and good inter-rater reliability in a smaller shared-rating subset.

For operators, the strongest use case is comparative: run SHS before and after a model, retrieval, prompt, or policy change and inspect which dimension moved. The weaker use case is certification. SHS does not establish whether an answer is objectively true, and the paper does not yet establish universal thresholds or robustness across deployment environments.

Reliability failures do not all look like factual errors

Suppose a product team replaces its retrieval pipeline. Automated factuality scores improve. Users still report that the system feels unreliable.

That reaction can arise from several different failures. A response may cite sources that are difficult to verify. Its reasoning may be internally weak. It may present uncertain material with excessive confidence. It may fail to incorporate a correction supplied by the user. None of these experiences is well described by asking only whether one proposition is factually true.

SHS formalizes that broader measurement problem through five user-observable dimensions: factual accuracy, source reliability, logical coherence, deceptiveness, and responsiveness to user guidance.

The distinction matters because different failures imply different interventions. Poor source reliability may point toward retrieval or citation handling. Weak responsiveness may implicate instruction following or interaction design. Perceived deceptiveness raises a different governance question from simple factual error because the issue is how confidently false or questionable information is presented.

The paper therefore contributes less as another hallucination detector than as a measurement layer for the human side of reliability.

Ten questions produce two signals, not one

Each SHS dimension contains one positively worded item and one negatively worded item. Responses are encoded on a symmetric five-point scale from -2 to +2.

For dimension (i), the directional score is:

where (p_i) is the positive-item response and (n_i) the negative-item response. The result ranges from -1 to +1, with higher values indicating lower perceived hallucination risk.

The overall SHS score is simply the equal-weighted mean:

It can also be linearly rescaled to 0–100:

The more distinctive design choice is that the same paired responses also generate a separate consistency indicator:

Values near zero indicate that the positive and negative judgments line up coherently. Larger absolute values suggest ambiguity or internally mixed impressions.

This prevents two questions from being collapsed into one metric: What did the user think about system reliability? and How internally stable was that judgment?

For a release process, that distinction is consequential. A low dimension score supported by a coherent response pair is different evidence from a low score accompanied by an internally inconsistent pair. The latter can be flagged for secondary review rather than automatically treated as a clean signal about the system.

The validation supports coherence, not universality

The paper’s main empirical evidence is a supervised study with 210 adult participants and 47 trained student experimenters. Participants interacted with an LLM for roughly 15 minutes using categories including verifiable facts, current events, reasoning tasks, source requests, and contradiction probes, then completed SHS and a feedback questionnaire.

The instrument achieved Cronbach’s alpha of 0.87, with a reported 95% confidence interval of 0.84–0.90. Corrected item-total correlations exceeded 0.50 for all ten items, and removing individual items did not materially improve reliability.

The five dimensions were correlated, but not perfectly so:

The inter-rater analysis used only 42 evaluations in which multiple raters assessed the same interactions, so it is supportive evidence rather than a broad demonstration of reproducibility.

Feasibility results are also relevant to operations. Mean completion time was 4.2 minutes with an SD of 1.8 minutes. Among experimenters, 87.2% reported that the questions were understandable, 83.0% considered them relevant to LLM evaluation, and 93.6% considered the response options appropriate.

These results support the paper’s intended role for SHS as a lightweight instrument. They do not establish that the scale transfers unchanged across industries, languages, user populations, or interaction designs.

What teams can do with dimension-level movement

The most defensible business use is repeated comparative measurement.

A product team could administer SHS before and after changing a model, system prompt, retrieval layer, citation mechanism, or policy configuration. The aggregate score gives a summary direction, but the dimension scores show where the user experience changed.

That creates a more diagnostic release signal than a generic satisfaction score:

• Factual accuracy can reveal whether users perceive outputs as substantively dependable.
• Source reliability can expose failures in traceability or evidence presentation.
• Logical coherence can capture reasoning that appears unsupported or structurally weak.
• Deceptiveness focuses attention on confidently presented false or misleading material.
• Responsiveness measures whether users can steer the system toward greater accuracy or correction.

Cognaptus’ inference is that SHS fits best beside, not instead of, automated evaluation. An organization could combine automated factuality or RAG metrics with SHS, while retaining SUS for usability and SCS where explanation quality is relevant. The resulting stack distinguishes at least three questions: Is the output correct? Is the system usable? Does the user experience it as reliable?

Those questions can move independently.

A validated subjective scale is still subjective

The main interpretive error would be to turn SHS into a factual reliability certificate.

The paper explicitly does not make that claim. Scores depend on rater knowledge, attention, framing, prompting strategy, and interaction context. Non-expert users can miss subtle factual errors. A user may also perceive a fluent but incorrect answer as reliable.

The authors provide interpretation ranges—for example, native SHS scores from +0.5 to +1.0 are labeled low hallucination risk—but those ranges should be understood as proposed scale guidance, not externally validated universal deployment thresholds.

The validation is also initial. The paper does not demonstrate measurement invariance across application domains, languages, prompting strategies, interaction styles, or longitudinal production changes. Those gaps matter if an organization wants to compare scores across teams, markets, or time periods.

A sensible deployment pattern is therefore to establish a local baseline, preserve the evaluation protocol, track dimension-level changes, inspect inconsistent ratings, and pair SHS with methods that can verify factual claims directly.

The measurement gap is between correctness and experienced reliability

SHS addresses a narrow but consequential gap in LLM evaluation. Automated benchmarks are designed to tell us something about model outputs. SHS is designed to tell us, systematically, something about how people experience those outputs during interaction.

The initial evidence suggests the instrument is coherent enough and lightweight enough to justify further use as a comparative diagnostic. Its paired-item design adds a useful safeguard by separating the direction of a reliability judgment from the coherence of the judgment itself.

For deployment decisions, that is the relevant contribution. SHS does not answer whether a system is objectively reliable. It gives teams a structured way to observe which forms of unreliability users notice, whether those perceptions change after an intervention, and when the human measurement itself deserves another look.

Cognaptus: Automate the Present, Incubate the Future.