When Hazard Maps Are Not Predictions: A Critical Assessment of MCDA in Glacier Hazard Susceptibility

Background: Multi-criteria decision analysis (MCDA) has become a dominant approach for glacier hazard susceptibility mapping, widely used to support risk management and climate adaptation planning. However, despite its widespread adoption, the role of MCDA outputs remains conceptually ambiguous: hazard classifications are often interpreted as predictive representations of risk, even though they are derived from preference-dependent decision models. This raises a critical but underexamined question regarding the reliability of MCDA-based glacier hazard assessments. This issue becomes particularly relevant in the current transition toward data-driven and artificial intelligence (AI)-based approaches for hazard modelling, where similar challenges of interpretability, validation, and reliability arise. Methods: To address this issue, we conducted a systematic literature review following the PRISMA 2020 protocol, analysing peer-reviewed studies published between 2015 and 2025. After screening 571 records, 60 studies were included. Data were extracted using a structured framework and synthesised through quantitative descriptive analysis and qualitative assessment of modelling practices, including method selection, criteria weighting, uncertainty treatment, validation, and geographical distribution. This study conducts a structured methodological audit—not a catalogue—of multi-criteria decision analysis (MCDA) applications in glacier hazard susceptibility mapping. Results: The analysis reveals a consistent methodological pattern. The Analytic Hierarchy Process (AHP) dominates current practice (36/60 studies, 60%), typically implemented through GIS-based weighted overlay with expert-derived weights. Critically, 80% of studies (48/60) derive criteria weights exclusively from expert judgement, with no data-driven calibration or sensitivity testing of subjective inputs. This epistemic reliance on unstructured or semi-structured expert elicitation, presented without robustness analysis, forms a central concern of this review. Moreover, empirical validation is limited: only 21/60 studies (35.0%) report quantitative performance metrics. Uncertainty and robustness analyses are rarely conducted, and most studies rely on single-model configurations without comparative evaluation. Despite these limitations, the resulting hazard maps are frequently presented as objective spatial predictions. The evidence base is also geographically concentrated, with 48/60 studies (80.0%) located in High Mountain Asia. Conclusions: The findings indicate a systematic mismatch between how MCDA-based hazard maps are constructed and how they are interpreted. In most cases, MCDA functions as a decision-structuring framework rather than a validated predictive model, yet its outputs are commonly treated as predictive evidence. This gap has important implications for the use of such models in risk management and climate adaptation, particularly in the emerging context of AI-driven hazard modelling, where issues of model validation, interpretability, and reliability become even more critical. Advancing the field requires explicit validation against observed events, systematic robustness and sensitivity analysis, transparent uncertainty modelling, and comparative evaluation of alternative or hybrid decision frameworks.