Anticipating biodiversity change is critical in rapidly warming regions, yet challenging because these areas often coincide with poor sampling. Data gaps are widely understood to interfere with species distribution models (SDMs), so bias-correction methods are generally trusted to account for this. However, when faced with real-world data gaps and climate change, it is unclear if these methods are reliable, as this is difficult to measure with biased data. To better understand SDMs in this context, we 1) measure performance with a new occurrence-checklist-range (OCR) validation, and 2) evaluate prediction discrepancy across space and time. We found: 1) bias-correction improves model performance against independent (checklist and range) data, but not against typical occurrence cross-validation, 2) predicted richness differed among methods (up to 2.7-fold), especially in the north, and 3) counterintuitively, future projections varied less (by 28%) because well-sampled climate space will shift north. Our findings suggest potential widespread overconfidence in SDM predictions for the unevenly sampled world, with implications for the growing reliance on biodiversity estimates for planning and policy. OCR validation and methodological discrepancy measurements are relatively easy ways to address this.

Keywords: SDMs, Sampling bias, Validation, Uncertainty