Data-driven climate predictions can fail when relationships learned from climate models do not hold in observations. Here we show that a boreal-winter subsurface dipole controls the cross-domain performance of neural networks predicting the Indian Ocean Dipole. Across CMIP6 models, networks achieve higher model-domain skill when the December–February subsurface dipole is more strongly correlated with the subsequent September–November Dipole Mode Index. This relationship is strongest in models with strong ENSO–Indian Ocean Dipole coupling. Yet these networks perform worst when transferred to ocean reanalysis, in which the dipole has the opposite sign. Perturbing or removing this feature reduces prediction skill within CMIP6 but improves predictions in reanalysis, demonstrating that the networks exploit it as a domain-dependent shortcut. Networks trained on weakly coupled models are less skilful within CMIP6 but transfer more reliably because they depend less on this misleading precursor. Our results show that the strongest predictor in a training domain can become the principal source of error under physical domain shift, highlighting the need to diagnose and constrain learned precursors in data-driven climate prediction.