Disagreement is usually treated as a bug in artificial intelligence. When a machine learning system looks at the same patient, image, or event through several different lenses—different feature sets, different sensors, different expert opinions—those lenses rarely agree perfectly. The standard playbook for multi-view learning has been to smooth over these conflicts, suppress them, or discard the views responsible for them. A team of researchers in Shanghai argues that this instinct is exactly backwards, and they have built a new framework that treats conflict itself as the most valuable signal in the room.
The new method, called TECL, for trusted and explainable collective learning, is described in the journal Knowledge and Information Systems by Nengjun Zhu and Zhiyu Zhang of Shanghai University, Shenghui Lan of Shanghai Eighth People’s Hospital, Jian Cao of Shanghai Jiao Tong University, and Siji Zhu of Ruijin Hospital. Their work extends a prior conference paper presented at the IEEE International Conference on Data Mining, and it was tested on a real-world dataset drawn from multi-disciplinary medical consultations at Ruijin Hospital in Shanghai. The code and a portion of the anonymized data have been released on GitHub, allowing other researchers to scrutinize and build on the approach.
The core insight is deceptively simple. In a group decision-making setting—say, a tumor board where surgeons, radiologists, oncologists, and pathologists each weigh in—disagreement is not noise. It highlights genuinely divergent perspectives, forces each contributor to articulate the reasoning behind a position, and offers clues about how much trust to place in individual judgments and in the final consensus. The same logic applies to machine representations. Each view of a multi-view dataset captures a distinct aspect of the same underlying condition, and when those aspects collide, the collision carries information. TECL is designed to harvest that information rather than erase it.
Technically, the framework operates in several stages. First, it captures what the authors call dual concepts at two different levels: the feature level and the decision level, across all of the views. This dual capture is what gives the model interpretability at two distinct depths. At the feature-attention level, the system can show what it is looking at—which elements of the input in each view are driving its attention. At the decision-reasoning level, it can show what it decides and how that decision relates to the positions taken by the other views. This layered transparency matters enormously in high-stakes domains like medicine, where a bare prediction with no rationale is difficult for clinicians to act on or audit.
Next, TECL learns view-specific evidence that supports each individual opinion. This draws on a line of research known as evidential deep learning, which trains neural networks to output not just a classification but a measure of how much support—how much evidence—they have for that classification. Rather than forcing every view to produce a single confident answer, the model lets each view state its position together with the strength of the backing behind it. A view that has seen ambiguous or sparse input can express low evidential support, and a view with clear signals can express strong support. The distinction is preserved rather than flattened.
The genuinely novel machinery appears when the views conflict. TECL runs what the authors call a reliability-oriented collective learning phase, whose job is to determine which view deserves higher priority in a given situation. This is not a fixed global ranking of views; it is a learned, context-sensitive assessment of reliability. Once priorities are established, the model combines the priority information with the view-specific evidence to arrive at a final collective opinion, together with an associated reliability score. In other words, the output is not merely a decision but a decision bundled with a calibrated statement of how much the system trusts that decision—an explicit acknowledgment that group conclusions built on shaky or contradictory inputs should carry lower confidence.
This architecture addresses a longstanding tension in multi-view learning. Earlier approaches to trusted multi-view classification, including influential work by Han and colleagues on evidential fusion, have shown how to combine uncertainty across views, and more recent research on reliable conflictive multi-view learning has begun to grapple with disagreement directly. TECL pushes further by making the treatment of conflict proactive rather than defensive, and by tying the reliability estimation to an interpretable reasoning chain. The authors also situate the work within the broader concept-based interpretability literature, in which models are constrained to reason through human-understandable intermediate concepts—ideas explored in concept bottleneck models and their many probabilistic and post-hoc variants.
The real-world testbed for TECL was multi-disciplinary consultation, one of the hardest collective decision-making problems in medicine. In such consultations, specialists examine different slices of the evidence—imaging, laboratory results, clinical history, physical examination—and frequently reach different conclusions about diagnosis or treatment. A decision support system built on this data must therefore fuse genuinely conflicting expert-style opinions, each grounded in a different view of the patient. According to the authors, experiments on their multi-disciplinary consultation dataset demonstrated the superiority of the method, though full quantitative details are available in the journal article itself. The dataset originated at Ruijin Hospital, was used with permission, was anonymized, and has been partially open-sourced.
The implications reach well beyond the clinic. Any setting in which multiple information sources must be reconciled—autonomous vehicles fusing camera, radar, and lidar; financial systems combining heterogeneous market indicators; sensor networks monitoring infrastructure—faces the same structural problem of conflictive multi-view data. A framework that can rank the reliability of conflicting sources on the fly, explain its reasoning at both the feature and decision levels, and attach a trust score to its own conclusions could change how engineers think about robustness. Instead of engineering conflicts away, developers might design systems that expect disagreement and exploit it.
There are also broader questions about trustworthiness in artificial intelligence that this line of work speaks to. Interpretability researchers, including Cynthia Rudin and collaborators, have argued that high-stakes decisions demand models whose reasoning is fundamentally inspectable, not bolted on afterward. By building interpretability into the learning process itself—through dual concept capture at feature and decision levels—and by grounding reliability estimates in the visible structure of disagreement, TECL offers a concrete template for what trusted AI might look like when it has to think the way a committee thinks: acknowledging dissent, weighing reputations, and stating its confidence out loud. The work was supported by funding bodies including the National Natural Science Foundation of China and the Shanghai Municipal Health Commission, and the authors report that all data supporting the findings are available within the paper and its associated repository.
Subject of Research: Trusted and explainable collective learning for conflictive multi-view decision-making
Article Title: Trusted and explainable collective learning for conflictive multi-view decision-making
Article References: Zhu, N., Zhang, Z., Lan, S., Cao, J., & Zhu, S. (2026). Trusted and explainable collective learning for conflictive multi-view decision-making. Knowledge and Information Systems, 68(1), Article 264. https://doi.org/10.1007/s10115-026-02884-1
Image Credits: AI Generated
DOI: 10.1007/s10115-026-02884-1
Keywords: multi-view learning, collective learning, conflictive opinions, explainable AI, evidential deep learning, reliability modeling, decision support, medical consultation, interpretability, uncertainty quantification, machine learning, trusted AI
Cite Scienmag News
APA MLA Chicago
Blake Davidson. (September 26, 2026). AI Turns Conflicting Opinions Into a Feature: New Model Weighs Disagreement to Make Group Decisions Trustworthy. Scienmag. https://scienmag.com/ai-turns-conflicting-opinions-into-a-feature-new-model-weighs-disagreement-to-make-group-decisions-trustworthy/
Blake Davidson. “AI Turns Conflicting Opinions Into a Feature: New Model Weighs Disagreement to Make Group Decisions Trustworthy.” Scienmag, 26 September 2026, https://scienmag.com/ai-turns-conflicting-opinions-into-a-feature-new-model-weighs-disagreement-to-make-group-decisions-trustworthy/. Accessed 26 September 2026.
Blake Davidson. “AI Turns Conflicting Opinions Into a Feature: New Model Weighs Disagreement to Make Group Decisions Trustworthy.” Scienmag. September 26, 2026. https://scienmag.com/ai-turns-conflicting-opinions-into-a-feature-new-model-weighs-disagreement-to-make-group-decisions-trustworthy/
Copy citation Download RIS
Tags: collaborative decision-making in AIcollective learningconflict as valuable signalconflicting opinions in AIconflictive opinionsdecision supportevidential deep learningexplainable AIExplainable Artificial Intelligencehandling disagreement in machine learninginterpretabilityMachine learningmedical consultationmedical decision support systemsmulti-disciplinary medical consultationsmulti-view learningopen-source AI researchreal-world healthcare datasetsreliability modelingTECL framework for collective learningtrusted AItrustworthiness in group decision-makinguncertainty quantification
