Human Interpretation Correctness (P-HIC)

• Human Interpretation Correctness (P-HIC) is defined as the proportion of instances where human responses align with AI outputs, ground truths, or expert annotations.
• Experimental protocols use label guessing, visual explanation, and data extraction tasks to quantify interpretability using metrics such as accuracy, ROC-AUC, and Cohen’s κ.
• Adaptive interfaces and predictive models leverage P-HIC insights to refine explanation designs and mitigate confirmation bias in human–AI decision processes.

Human Interpretation Correctness (P-HIC) quantifies the degree to which human users correctly interpret, predict, or reproduce key outputs or reasoning steps of AI models, particularly in settings where interpretability or explainability is under study. It is operationalized as the probability (or empirical proportion) that a human response matches a model decision, a ground-truth label, or a predefined annotation, depending on the evaluative context. P-HIC serves both as a direct measure of human–AI alignment in interpretability research and as a predictive modeling target for adaptive interfaces and question design in personalized assessment scenarios (Shen et al., 2020, Kim et al., 2021, Liu et al., 2016, Long et al., 13 Aug 2025, Falessi et al., 28 Jan 2026).

1. Formal Definitions and Mathematical Frameworks

The formalization of P-HIC is context dependent. In binary or multiclass interpretability tasks, P-HIC is typically defined as:

$$\text{P-HIC} = \frac{1}{N} \sum_{i=1}^N \mathbb{I}(\text{human response}_i = \text{target}_i),$$

where $\mathbb{I}(\cdot)$ is the indicator function, and $\text{target}_i$ may be the true model output, ground-truth label, or an expert annotation. This core definition encompasses a variety of experimental settings:

• Label prediction tasks (Shen et al., 2020, Kim et al., 2021)
• Agreement or distinction in visual explanations (Kim et al., 2021)
• Attention map alignment in sequence generation (Liu et al., 2016)
• Factual extraction accuracy in data synthesis (Long et al., 13 Aug 2025)
• Correct answer prediction in visualization assessment (Falessi et al., 28 Jan 2026)

More elaborate instantiations of P-HIC include ROC-based, regression, and IRT-based variants for predicting correctness probabilities as a function of subject, task, and item characteristics (Falessi et al., 28 Jan 2026).

In human–machine extraction workflows, P-HIC is operationalized as the proportion of factually correct human (or AI) responses, factoring out permissible interpretive differences and coding only non-defensible inaccuracies as deviations from correctness (Long et al., 13 Aug 2025). In neural captioning, "attention correctness" measures per-word P-HIC as the summed attention mass over the human-annotated grounding region,

$$\mathrm{P\!-\!HIC}(y_t) = \sum_{i\in R_t}\hat{\alpha}_{t,i}$$

for word $y_t$ and region $R_t$, with further averaging or aggregation as appropriate (Liu et al., 2016).

2. Experimental Protocols and Task Designs

P-HIC is realized via a range of human-centered evaluation protocols, structured to minimize prior knowledge confounds and support falsifiable comparison across interpretability methods. Representative designs include:

• Guessing incorrectly-predicted labels: Humans are shown an input (with or without AI-generated saliency explanations), the true label, and a candidate set including the model’s actual (mis)classification. P-HIC quantifies the proportion guessing the model’s output correctly (Shen et al., 2020).
• Agreement and distinction tasks: For visual explanation assessment, humans judge whether a model’s prediction is correct given its visual explanation, or select the correct class among multiple candidate explanations (Kim et al., 2021).
• Data extraction correctness: For AI-assisted literature extraction, human or AI outputs are scored as accurate only if they are factually grounded in the source text, with interpretive differences coded separately. P-HIC corresponds to the factual accuracy rate (Long et al., 13 Aug 2025).
• Visualization assessment: In data visualization, P-HIC is both measured empirically (correct/incorrect response) and predicted via logistic modeling, with features spanning item, human profile, and prior performance (Falessi et al., 28 Jan 2026).

Statistical evaluation employs accuracy, Cohen’s κ for interrater reliability, ROC-AUC for predictive discrimination, and chance-baseline testing to disambiguate meaningful alignment from random agreement.

3. Empirical Results and Comparative Benchmarks

P-HIC varies substantially across domains, study designs, and interpretability modalities. Empirical findings include:

Context/Task	P-HIC Value	Notable Findings
Saliency explanations (interpretation)	0.63–0.73 (w/ explanations)	Visual explanations reduced correctness by ~10%
Saliency explanations (no interpretation)	0.73–0.81	Higher human-model alignment without explanations
Visual explanation (HIVE distinction)	up to 71% on correct; 26–33% on incorrect	High on correct, at-chance or below on errors
Data extraction factual correctness	95.6% (human); 98.5% (AI)	Human error rate ≈ 3× AI; 24% human–AI discordance
Neural captioning (attention correctness)	0.43–0.58 (model P-HIC)	Strong supervision improves attention correctness
Visualization correctness prediction	AUC ≈ 0.72, κ ≈ 0.32	RaschDifficulty dominant predictor

These results expose key limitations: visual or saliency-based explanations do not guarantee improved P-HIC—for some error types, they actively reduce it (Shen et al., 2020). Confirmation bias is notable: humans trust explanations even when the AI is incorrect (Kim et al., 2021). Explicit attention supervision and expert-anchored prompts substantially increase alignment (Liu et al., 2016, Long et al., 13 Aug 2025).

4. Predictive Modeling and Feature Analysis

In predictive contexts, P-HIC becomes the output variable of machine-learning models trained to anticipate user correctness before item exposure (Falessi et al., 28 Jan 2026). Predictors fall into three major classes:

• Item Difficulty: Rasch model parameters and expert difficulty ratings are the strongest predictors, capturing the intrinsic challenge of a visualization or question.
• Human Performance: Prior correctness, computed as the running proportion of right answers, helps capture fatigue, learning, or consistency dynamics.
• Human Profile: Demographic and background variables contribute minimally; most variation is explained by item-level psychometrics and immediate performance metrics.

In visualization assessment (Falessi et al., 28 Jan 2026), logistic regression with feature selection retained only three predictors: RaschDifficulty, ExpertDifficulty, and PercCorrect. The marginal value of profile features was near zero.

Median AUC for model-based P-HIC prediction was 0.72, indicating substantial but not perfect discriminative power. This suggests substantial person–item interaction effects and the importance of calibrated psychometric scaling.

5. Error Typology and Human–AI Discordance

The evaluation of P-HIC, especially in extraction and complex interpretation tasks, explicitly distinguishes between:

• Interpretive differences: Multiple reasonable answers given ambiguous or underspecified source material. These do not count against P-HIC (Long et al., 13 Aug 2025).
• Factual inaccuracies: Errors not rooted in source text or defensible inference, including omissions and misclassifications. Only these degrade the P-HIC score.

Discordance between AI and human extractions is dominated by interpretive variability (18% of all responses). Human factual error rates are higher than those of AI in structured extraction tasks; yet, consistent patterns of discordance signal question interpretability or ambiguity rather than extractor deficiency.

Error analysis also shows that saliency-based methods, when confronted with ambiguous or nuanced errors (e.g., similar-looking classes, spurious correlations), can decrease human alignment with model decisions (Shen et al., 2020).

6. Implications for Design and Evaluation of Interpretable Systems

P-HIC serves as a critical empirical constraint in the development of AI interpretability tools:

• Design principle: P-HIC must be part of any evaluation of interpretability or explanation pipelines, as fidelity-based or automatic metrics (e.g., pointing game, IoU) do not correlate with human behavioral accuracy (Kim et al., 2021).
• Dynamic refinement: In extraction and knowledge synthesis, persistent low P-HIC or low inter-extractor reliability should trigger question or prompt refinement rather than simple retraining (Long et al., 13 Aug 2025).
• Personalization: Predictive P-HIC estimation enables adaptive task or item presentation—avoiding overly difficult questions, selecting items for information gain, and targeting training or feedback based on anticipated misunderstanding (Falessi et al., 28 Jan 2026).
• Risk management: Human confirmation bias in explanation tasks underscores the need for caution: explanations may boost unwarranted trust unless rigorously validated for interpretive correctness (Kim et al., 2021).
• Metrics extension: A plausible implication is that generalizing P-HIC measurement to new domains—clinical scoring, educational feedback—can surface latent ambiguity and inform targeted human or AI adjudication.

7. Limitations and Research Challenges

P-HIC, as currently developed, is constrained by several factors:

• Task and domain specificity: Formulations are tightly bound to particular interaction settings (label guessing, data extraction, visualization, captioning), complicating comparisons.
• Interpretive variability: Error classification depends critically on the ability to separate acceptable interpretive diversity from true error—a non-trivial judgment in open-ended tasks (Long et al., 13 Aug 2025).
• Individual differences: While profile features are weak global predictors, local effects (e.g., expertise, fatigue) may be non-negligible for specialized cohorts.
• Explanation modality: Not all forms of explanation support improved P-HIC; in fact, poorly designed visual explanations can degrade human interpretive accuracy (Shen et al., 2020).
• Resource intensity: Empirical P-HIC measurement requires controlled behavioral experiments or rigorous annotation, posing scalability challenges (Kim et al., 2021).

Ongoing research seeks to refine P-HIC predictive models, automate error typology coding, and integrate adaptive mechanisms into real-time decision-support interfaces.

---

For comprehensive treatments and technical details, see "How Useful Are the Machine-Generated Interpretations to General Users?" (Shen et al., 2020); "HIVE: Evaluating the Human Interpretability of Visual Explanations" (Kim et al., 2021); "Attention Correctness in Neural Image Captioning" (Liu et al., 2016); "Hallucination vs interpretation: rethinking accuracy and precision in AI-assisted data extraction for knowledge synthesis" (Long et al., 13 Aug 2025); and "Beyond Literacy: Predicting Interpretation Correctness of Visualizations with User Traits, Item Difficulty, and Rasch Scores" (Falessi et al., 28 Jan 2026).