Research map / Label-efficient and robust learning / Recognition and long-tail
Dissenting Explanations: Leveraging Disagreement to Reduce Model Overreliance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces dissenting explanations (explanations for a disagreeing model's prediction) and shows in a human study on deceptive hotel review detection that they reduce overreliance without lowering accuracy, plus methods to induce global and local predictive disagreement.
Problem
Humans over-rely on AI predictions and explanations even when the AI is wrong; explanations that only argue for one prediction may increase trust in incorrect predictions in AI-assisted decision making.
Approach
Present an explanation for a disagreeing model's opposing prediction alongside the reference model's explanation (a "dissenting explanation"), treating explanations as arguments, and develop objective-based methods (regularization, reweighting, local retraining) to induce predictive and explanation multiplicity.
Key results
- Dissenting explanations (C2) significantly reduced overreliance versus single explanation (C1) (p<0.001) without a significant change in accuracy (p>0.05), with N=178 Prolific participants.
- Regularization objective with lambda=0.5 raised disagreement to 16.6% from 8.66% while accuracy fell from 0.889 to 0.807 on the reviews task.
- Local retraining of a neural network flipped the label of roughly 80% of test examples while keeping about 88% accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Provide a supporting explanation plus an opposing model's dissenting explanation instead of one explanation | Human decisions assisted by the reference model · Deceptive hotel reviews | Overreliance | Helped |
| Provide a supporting explanation plus an opposing model's dissenting explanation instead of one explanation | Human decisions assisted by the reference model · Deceptive hotel reviews | Accuracy | No clear effect |
| Show positive and negative features from one model instead of explanations from disagreeing models (ablation) | Human decisions assisted by the reference model · Deceptive hotel reviews | Overreliance | Hurt |
| Provide one supporting explanation instead of no explanation (ablation) | Human decisions assisted by the reference model · Deceptive hotel reviews | Agreement on incorrect AI predictions | No clear effect |
| Increase Reg λ from 0.0 to 0.5 (ablation) | Neural network trained against the linear SVM · Deceptive hotel reviews test set | Disagreement: 8.66 ± 0.6 % -> 16.6 ± 2.3 % | Helped |
| Increase Reg λ from 0.0 to 0.5 (ablation) | Neural network trained against the linear SVM · Deceptive hotel reviews test set | Accuracy: 0.889 ± .010 -> 0.807 ± .017 | Hurt |
| Increase Weights λ from 0 to 50 (ablation) | Neural network trained against the linear SVM · Deceptive hotel reviews test set | Disagreement: 8.68 ± 0.7 % -> 14.9 ± 0.7 % | Helped |
| Increase Weights λ from 0 to 50 (ablation) | Neural network trained against the linear SVM · Deceptive hotel reviews test set | Accuracy: 0.859 ± .019 -> 0.826 ± .018 | Hurt |
| Increase Reg λ from 0.0 to 0.50 (ablation) | Neural network trained against the linear SVM · Adult Income test set | Disagreement: 6.12 ± 0.6 % -> 27.3 ± 2.6 % | Helped |
| Increase Reg λ from 0.0 to 0.50 (ablation) | Neural network trained against the linear SVM · Adult Income test set | Accuracy: 0.754 ± .002 -> 0.706 ± .004 | Hurt |
| Increase Weights λ from 0 to 5.0 (ablation) | Neural network trained against the linear SVM · Adult Income test set | Disagreement: 9.10 ± 0.5 % -> 56.3 ± 6.8 % | Helped |
| Increase Weights λ from 0 to 5.0 (ablation) | Neural network trained against the linear SVM · Adult Income test set | Accuracy: 0.766 ± .002 -> 0.560 ± .002 | Hurt |
Surprising findings
- Showing both positive and negative features from a single model (C3) produced higher overreliance than dissenting explanations from two models (C2).
Limitations
- Human study covers only one task (deceptive reviews), so generality of dissenting explanations is unproven.
- Explanations are LIME feature highlights only; counterfactual or other explanation types not studied.
- Unclear whether gains come from feature differences or from framing of two models versus one.
Experimental setup
Datasets: Chicago Deceptive Reviews Folktables (UCI Adult, Hawaii 2018 Income)
Benchmarks: Chicago Deceptive Reviews Folktables Adult Income
Metrics: accuracy overreliance Cohen's kappa TopK explanation agreement predictive disagreement success rate
Baselines: single explanation condition (C1) prediction only (C0) positive and negative features from single model (C3) linear SVM reference model
Builds on: LIME Rashomon set boosting
Closest papers by meaning
- Towards Unifying Evaluation of Counterfactual Explanations: Leveraging Large Language Models for Human-Centric Assessments (AAAI 2025 · similarity 0.87)
- A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.87)
- Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations (ICLR 2025 · similarity 0.86)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (AAAI 2025 · similarity 0.86)
- Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations (ICML 2024 · similarity 0.85)
- The First Impression Problem: Internal Bias Triggers Overthinking in Reasoning Models (ICLR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.