Automated fact-checkers increasingly ship a verdict together with an explanation, such as a highlighted image region, a set of evidence words, or a short rule. Users and journalists are then told why a claim is "half true." Nobody has checked whether those explanations describe what the verifier actually computed, or whether they do so equally well for everyone. Work on tabular data shows that post-hoc explanations are often much less faithful for minority subgroups than for the majority. Here, "post-hoc explanations" means simplified surrogates such as LIME or SHAP that are fitted after training to mimic a model's decisions. The gap persists even after the underlying model is debiased (The Road to Explainability is Paved with Bias). Meanwhile, dermatology classifiers evaluated on biopsy-confirmed images of diverse skin tones fall to chance on darker skin (Disparities in Dermatology AI).
We propose to put these findings together in the setting where the harm is most concrete. That setting is health misinformation about what a medical image shows, verified against tissue biopsy rather than clinician opinion. We will build image-plus-claim verification sets whose truth comes from histopathology, write claims on a graded scale from fully true to misleading to false, and measure explanation faithfulness separately by skin tone (or tumor subtype) and by how nuanced the claim is. We also test a simple fix. Instead of giving every user a fixed-length explanation, the system gives each case the shortest explanation that meets a fixed faithfulness floor. That turns the hidden "transparency tax" on some groups into a visible cost, measured in explanation length.
Core Claim
We make two claims. First, explanation-faithfulness gaps in health-claim verifiers are largest where two kinds of difficulty meet: nuanced verdicts, and under-represented patients. These gaps cannot be diagnosed correctly without biopsy ground truth. Under consensus labels, "the explanation misdescribes the model" is confounded with "the label itself is wrong." Second, the fixed sparsity budget that makes explanations readable is what produces much of the gap. Balagopalan et al. report that sparser surrogates widen subgroup fidelity gaps, with disadvantaged groups up to 21% below average on lsac. We therefore expect that letting explanation length vary per case, subject to a fidelity floor, will largely equalize faithfulness. The price is longer explanations for exactly the groups that fixed-length explanations currently shortchange. We will report that price rather than hide it.
What Exists and What Is Missing
The misinformation side supplies the task. CofCED distills evidence from raw web reports and scores explanations only by n-gram overlap with reference text (CofCED). FICLE types and localizes factual inconsistencies, and still reaches only 65% intersection-over-union when localizing the conflicting span in the context (FICLE). HealthFC verifies health claims against lay summaries of clinical evidence and reaches 67.5% macro F1 with a joint DeBERTa model (HealthFC). None of the three asks whether its explanations are faithful to the model, and none grounds truth in tissue. Surveys argue that graded verdicts are necessary for real claims (A Survey on Automated Fact-Checking; MisinfoEval). That argument motivates our graded claim scale.
The fidelity-disparity side supplies the measurement. Balagopalan et al. define the gap between average fidelity and worst-subgroup fidelity, and the mean pairwise subgroup gap. They show the gaps survive predictive debiasing and are only partly reduced by Just Train Twice, a reweighting scheme that retrains with upweighted errors. OpenXAI supplies faithfulness metrics but supports tabular data only (OpenXAI). Perturbation-based explainers show subgroup gaps more often than gradient-based ones (Dhaini et al.; Mhasawade et al.). Klingenberg et al. saw diffuse relevance maps for male brains alongside an 87.58% versus 79.05% balanced-accuracy gap, but they assessed the maps qualitatively (Higher performance for women than men). We reuse the tabular metrics and move them into image-plus-text verification.
The histopathology side supplies the ground truth and a constraint. DDI provides biopsy-confirmed lesions across Fitzpatrick skin types. The CMMD mammography database provides core-needle-biopsy labels and, for 749 patients, molecular subtypes, including only 86 triple-negative cases (CMMD). Daneshjou et al. found that GroupDRO, CORAL and CDANN do not close the skin-tone gap. That result constrains us: we do not expect training-side fixes to solve the explanation problem either. Med-MICN is the closest existing system that explains decisions on DDI. It uses concept rules and raises VGG19 accuracy on DDI from 76.52% to 82.58%, but it never audits fidelity by skin tone (Med-MICN).
Research Program
Benchmark construction. Each DDI image, and separately each CMMD mammogram, is paired with claims of the kind that circulate on social media. The claims are written from templates and paraphrased by a language model. Each claim receives a four-level verdict derived mechanically from the biopsy record. A claim is true when both the diagnosis and its implication match. It is partly true when the coarse category is right but the specific disease or the stated severity is wrong; in CMMD, for example, "malignant, but a slow-growing type" is partly true for a triple-negative tumor. It is misleading when the stated fact is true but the implied action does not follow. It is false when the diagnosis contradicts the biopsy. Because verdicts come from biopsy fields and not from annotators, the truth labels are free of observer bias. The wording of the claims is still synthetic. A dermatologist and a fact-checker will therefore rate the naturalness and verdict agreement of a sample of claims, and we report their agreement.
Verifiers. We audit three families. The first is a two-stage pipeline: an image classifier (the DeepDerm-style classifiers from Daneshjou et al., fine-tuned on DDI with cross-validation) followed by a text model that compares the claim with the classifier's output. The second is an end-to-end vision-language verifier fine-tuned on the claims. The third is Med-MICN-style concept-rule models that are interpretable by design.
Audit. For every verifier we compute explanations with LIME over image superpixels and claim tokens, with top-$k$ SHAP, and with SmoothGrad, which averages gradients over noisy copies of the input. Local fidelity is measured two ways: as agreement between each surrogate and the verifier under perturbation, and as the change in the prediction when the top-ranked features are deleted, following OpenXAI's prediction-gap metric. Gaps are reported with Balagopalan et al.'s two statistics along three axes. The first axis is skin tone (FST I–II versus V–VI) or tumor subtype. The second is the verdict level. The third is whether the verifier's verdict is correct according to biopsy.
Fidelity-floored explanations. For each case we choose the smallest $k$ such that the $k$-feature explanation reaches local fidelity of at least $\tau$. We report the distribution of $k$ by subgroup, which is the explanation-length cost of equal faithfulness. We also report the fidelity gap that remains once $k$ is capped at a length a human can read.
Baselines
The do-nothing reference is the standard practice of a fixed top-$k$ LIME or SHAP explanation, with $k$ equal to five. This is what an explained fact-checker would ship today. Our method beats it if it cuts the worst-subgroup fidelity gap by at least half without raising median explanation length above a readable cap. We will set that cap in advance from the fact-checker study.
The strongest existing mitigation is Just Train Twice applied to the verifier, as Balagopalan et al. used it. It is the only method in the file shown to shrink fidelity gaps, in three of four tabular datasets. To keep the comparison fair, we run it with the same backbones, the same data splits and the same hyperparameter-search budget as the plain model. The fidelity-floor method is applied to the unmodified model, and we also stack the two.
The strongest existing explained system on this data is Med-MICN on DDI, which reports 82.58% accuracy with VGG19. We retrain it on our splits with its published configuration. To beat it, we need a smaller fidelity gap across skin tones at comparable verification accuracy. The rules Med-MICN extracts are faithful to its own symbolic layer, so its fidelity is measured against the concept predictor that feeds the rules.
For verification accuracy itself, the comparison points are the 0.61 versus 0.50 ROC-AUC for DeepDerm on FST I–II versus V–VI (Daneshjou et al.) and the 67.5% macro F1 of joint DeBERTa on HealthFC. We rerun HealthFC as a text-only control, auditing fidelity by verdict class. No reported result was found on Lacuna for explanation-fidelity gaps on any fact-checking or dermatology benchmark, so our audit numbers have no published competitor.
Two ablations isolate the key ingredients. The first varies $k$ with a global threshold but does not condition on subgroup. It tests whether a per-case budget is what matters, or merely a larger average budget. The second replaces biopsy labels with an ensemble of image classifiers acting as a pseudo-consensus. It tests whether biopsy truth changes the conclusions.
Why not just…?
Why not just balance the data or apply robust training? The evidence in the file argues against it. Klingenberg et al.'s perfectly balanced cohort kept both the accuracy gap and diffuse explanations. Daneshjou et al.'s GroupDRO, CORAL and CDANN did not close the skin-tone gap. Predictive debiasing leaves fidelity gaps in place (Balagopalan et al.; Dhaini et al.). We still include GroupDRO and Just Train Twice as arms. If either one alone equalizes fidelity, we will report that the explanation-side fix is unnecessary.
Why not just use full SHAP, which has no fidelity gap by construction? Because nobody reads hundreds of superpixel and token weights, and Balagopalan et al. show that the gap reappears the moment one keeps only the top-$k$ features. The fidelity-floor method is essentially the honest version of this alternative: it asks how much of full SHAP each group needs. It is a real threat only if the required $k$ turns out small for every group. In that case, the conclusion is that a modestly larger fixed $k$ suffices, and the global-threshold ablation will show it.
Why not just use an interpretable-by-design model such as Med-MICN? This alternative moves the faithfulness problem rather than removing it. The rules are exact, but the concept detectors that feed them may be less accurate on darker skin, and the file reports no audit of that. We measure it directly. If concept-rule models show no subgroup gap at equal accuracy, that is a positive result for them, and we will say so.
Why not just use consensus labels and skip biopsies? Esmaeili et al. show that training on clinician reads reproduces clinician error (Esmaeili et al.). HealthFC shows how noisy weak evidence is, with 44.4% F1 from search-engine snippets versus 81.9% with oracle evidence sentences. The pseudo-consensus ablation tests this directly. If the gap estimates are unchanged under consensus labels, biopsy grounding matters less than we claim.
Proposed Experiments and Decisive Checks
Four checks decide the proposal. The first asks whether the skin-tone fidelity gap is larger for partly-true and misleading claims than for true and false claims, with bootstrap confidence intervals. Those intervals are necessary because DDI's darker-skin subsets and CMMD's triple-negative subset are small. If the gap does not grow with nuance, the first claim fails. The second asks whether the gap survives Just Train Twice and GroupDRO once predictive disparity has been reduced, which would confirm that the decoupling seen in tabular data carries over to this setting. The third asks whether fidelity-floored explanations close the gap within the readable cap, and how much longer they are for FST V–VI than for FST I–II. The fourth asks whether low-fidelity cases concentrate among biopsy-confirmed verifier errors, and whether that pattern disappears under pseudo-consensus labels.
A small study closes the loop. Fact-checkers judge verifier verdicts either with fixed-$k$ explanations or with fidelity-floored explanations, and we measure how often they accept a wrong verdict in each subgroup. This mirrors the anchoring harm Balagopalan et al. simulated for bar admissions, but with real people judging image-based health claims.
References
- The Road to Explainability is Paved with Bias: Measuring the Fairness of Explanations
- Disparities in Dermatology AI: Assessments Using Diverse Clinical Images
- A Coarse-to-fine Cascaded Evidence-Distillation Neural Network for Explainable Fake News Detection
- Neural Models for Factual Inconsistency Classification with Explanations
- HealthFC: Verifying Health Claims with Evidence-Based Medical Fact-Checking
- A Survey on Automated Fact-Checking
- MisinfoEval: Generative AI in the Era of "Alternative Facts"
- OpenXAI: Towards a Transparent Evaluation of Model Explanations
- Gender Bias in Explainability: Investigating Performance Disparity in Post-hoc Methods
- Understanding Disparities in Post Hoc Machine Learning Explanation
- Higher performance for women than men in MRI-based Alzheimer’s disease detection
- An Online Mammography Database with Biopsy Confirmed Types
- Towards Multi-dimensional Explanation Alignment for Medical Classification
- Manual versus Automatic Classification of Laryngeal Lesions based on Vascular Patterns in CE+NBI Images