Gregory et al. present in this preprint , a hybrid modeling approach that integrates machine learning (ML) into the GFDL SPEAR climate model to correct sea ice biases in real time. Two versions are tested: one that includes coupled feedbacks (HybridCPL) and one that does not (HybridIO). HybridCPL significantly improves Arctic and Antarctic sea ice forecasts on seasonal and subseasonal timescales. In contrast, HybridIO performs poorly due to unanticipated feedbacks. These results highlight the importance of training ML models within coupled climate systems for reliable predictions.